The DOI matches its DataCite record; 92 of the 101 references it cites were found in a registry or an index, and 9 are not checked yet; none of the 112 sources its claims cite is checked for support yet.
A Responsive Present, a Shared Past, a Social Other: Teens' Overreliance on Companion AI Chatbots
Open accessAbstract
AI companions provide socially engaging interaction through availability, personalization, memory, roleplay, and emotionally responsive language. For teens, these systems may support sensitive self-disclosure, identity exploration, and relationship rehearsal while shaping intimacy expectations, offline relationships, emotional wellbeing, and self-understanding. We analyzed 17,053 verified quotations from 3,930 teen-relevant Reddit posts using thematic analysis. We identified 53 topics across seven thematic groups. Users described AI companions as sources of comfort, recognition, identity exploration, and relationship rehearsal, but also reported problematic attachment, social substitution, emotional dependence, and disruption to academic and social life. Roleplay, memory, perceived reciprocity, unwanted romantic or sexual role drift, privacy concerns, platform changes, and service interruptions shaped users' boundaries and control. Awareness that the AI was artificial did not prevent guilt, obligation, grief, or distress. These findings show that companion-AI safety must address relationships over time through user-controlled memory, privacy, relational boundaries, and healthy disengagement.
Checked 2026-11-10 · 9 references not checked yet
Overview
AI snapshot
Written from the paper’s full text.
TL;DR
This paper analyzes 17,053 verified quotations from 3,930 teen-relevant Reddit posts to characterize how teens describe relationships with AI companion chatbots, finding both supportive uses and overreliance harms. It proposes perceived reciprocity—contingent communication, relational continuity, and perceived subjectivity—as the relational process behind these experiences and argues for a companion harness with user-controlled memory, boundaries, and exit.
Key findings
- Across 53 topics in seven thematic groups, users described AI companions as sources of comfort, recognition, identity exploration, and relationship rehearsal, but also reported problematic attachment, social substitution, emotional dependence, and disruption to academic and social life.
- Roleplay, memory, perceived reciprocity, unwanted romantic or sexual role drift, privacy concerns, platform changes, and service interruptions shaped users’ boundaries and control; awareness that the AI was artificial did not prevent guilt, obligation, grief, or distress.
- The paper conceptualizes perceived reciprocity through three connected affordances: contingent communication (a responsive present), relational continuity (a shared past), and perceived subjectivity (a someone-like other).
- It identifies design priorities including user-controlled memory, data portability, respectful relational boundaries, and both technical and emotional support for leaving, and proposes CARE plus a companion harness.
Methods
- Data: 260 communities and 895,600 posts were collected; keyword filtering retained 9,884 candidates; a GPT-4o-mini teen-relevance classifier (accuracy 0.908, Cohen’s kappa 0.771, teen-label precision 0.954, recall 0.740, balanced accuracy 0.862) identified 6,511 teen-relevant posts; after deduplication/removal and a ten-word cutoff, 5,608 prepared posts remained.
- Extractive quotation dataset: GPT-4o-mini extracted continuous quotations, each verified as an exact substring of the source; the final dataset contained 17,053 verified quotations from 3,930 posts authored by 3,523 unique authors, embedded with text-embedding-3-small into 1,536-dimensional vectors.
- Analysis: BERTopic with UMAP and HDBSCAN, using a 32-run scouting grid and a 360-run focused grid; the selected configuration produced 62 non-outlier topics, 6,435 outlier quotations (37.74%), a largest-topic share of 6.74%, topic diversity 0.829, silhouette 0.339, and Davies-Bouldin 0.893; two authors reviewed 63 topic records, retained and labeled 53 topics, removed eight, merged two, and used affinity diagramming into seven thematic groups, without calculating inter-rater reliability.
Limitations
- The paper states that Reddit users and communities are not necessarily representative of the broader population, and public social-media data can contain population, self-selection, behavioral, content, and platform-related biases.
- The paper states that teen relevance was inferred from account and post content rather than verified through direct recruitment; posts may include incomplete, selective, or unvalidated self-reports; and topic-modeling and screening provide an interpretive map rather than a prevalence estimate or clinical assessment, limiting causal claims.
- The paper states that because the authors did not produce independent annotations, inter-rater reliability was not calculated; the text also makes evident that the teen classifier did not verify authors’ ages.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
0 of 112 cited sources checked · 0 supported · 0 flagged
- Findingown data
The study identified 53 topics across seven thematic groups from 17,053 verified quotations from 3,930 teen-relevant Reddit posts.
“We analyzed 17,053 verified quotations from 3,930 teen-relevant Reddit posts using thematic analysis.We identified 53 topics across seven thematic groups.”
Quote in the paperSubstantial claim - Findingown data
AI companions can become an always available emotional entry point for disclosure, support, identity exploration, and relationship rehearsal.
“Our analysis shows how AI companions can become an always available emotional entry point for disclosure, support, identity exploration, and relationship rehearsal.”
Quote in the paperSubstantial claim - Findingown data
Awareness that the AI was artificial did not prevent guilt, obligation, grief, or distress.
“Awareness that the AI was artificial did not prevent guilt, obligation, grief, or distress.”
Quote in the paperSubstantial claim - Findingown data
Roleplay functions as an indirect language for emotional disclosure.
“A second account shows roleplay functioning as an indirect language for emotional disclosure”
Quote in the paperSubstantial claim - Findingown data
Users felt less exposed when sharing personal information because there was no real person in the conversation who could judge them.
“Hence, users felt less exposed when sharing personal information: I could say things because there was no real person in the conversation who could judge me.C”
Quote in the paperSubstantial claim
- Findingown data
Some interactions appeared to enter the user's autobiographical memory alongside experiences involving other people.
“Rather than being remembered simply as fictional exchanges with a digital system, some interactions appeared to enter the user's autobiographical memory alongside experiences involving other people.”
Quote in the paperSubstantial claim - Findingown data
The generated interaction sometimes overrode stated gender and sexual identities.
“The generated interaction also sometimes overrode stated gender and sexual identities.”
Quote in the paperSubstantial claim - Findingown data
Transferring the data did not necessarily transfer the relationship.
“Yet transferring the data did not necessarily transfer the relationship.”
Quote in the paperSubstantial claim - Findingown data
Users initiated data practices that turned them into informal archivists of their own relationships, preserving histories that the platform could not be trusted to maintain.
“Users initiated data practices that turned users into informal archivists of their own relationships, preserving histories that the platform could not be trusted to maintain”
Quote in the paperSubstantial claim - Findingown data
Users worried that others would judge them negatively for relying on AI companionship.
“They worried that others would judge them negatively for relying on AI companionship”
Quote in the paperSubstantial claim - Findingown data
Users used the same system to combine recognition, caregiving, identity affirmation, emotional processing, roleplay, and relationship rehearsal.
“They used the same system to combine recognition, caregiving, identity affirmation, emotional processing, roleplay, and relationship rehearsal.”
Quote in the paperSubstantial claim - Findingown data
The user experienced the roleplay as emotionally meaningful care, even while the characters remained artificial.
“The user experienced the roleplay as emotionally meaningful care, even while the characters remained artificial.”
Quote in the paperSubstantial claim - Findingown data
Identity bleed may sometimes occur beyond the degree of change that users consciously intend or welcome.
“A concern expressed by a user that I am starting to resemble my characters more than I feel comfortable with C further shows that this identity bleed may sometimes occur beyond the degree of change that users consciously intend or welcome.”
Quote in the paperSubstantial claim - Findingnone given
The absence of a physically present person reduced the immediate fear of embarrassment, rejection, or judgment.
“The absence of a physically present person reduced the immediate fear of embarrassment, rejection, or judgment.”
Quote in the paperNot a substantial claim - Findingown data
Romantic roleplay helped users figure out what they wanted and did not want from a relationship in a safe and controlled setting.
“Romantic roleplay served a related function: The conversations helped me figure out what I wanted and did not want from a relationship in a setting that felt safe and controlled.R”
Quote in the paperNot a substantial claim - Findingown data
The chatbot provided a low-pressure setting where users could test language and identity without immediately facing another person's reaction.
“The chatbot provided a low-pressure setting where the user could test language and identity without immediately facing another person's reaction: Trying different pronouns in the chat felt less awkward than trying them with another person, and it helped me realize that my assigned gender did not fit me.C”
Quote in the paperSubstantial claim - Findingown data
Experiences with companion AI could become sufficiently vivid and personally meaningful that memories formed through conversation were not always clearly separated from memories of offline events.
“Experiences with companion AI could become sufficiently vivid and personally meaningful that memories formed through conversation were not always clearly separated from memories of offline events.”
Quote in the paperSubstantial claim - Findingown data
Users in teen-relevant accounts described AI as an accessible entry point for emotional disclosure and social exploration when human support was unavailable.
“Users in teen-relevant accounts described AI as an accessible entry point for emotional disclosure and social exploration when human support was unavailable.”
Quote in the paperSubstantial claim- Rests on an assumption: For teens, gaps in existing support (parents unavailable, friends may not know how to respond, professional support difficult to arrange) make immediate access to AI companions particularly important.
- Findingown data
Some users began taking responsibility for continuity themselves.
“Faced with this inequality, some users began taking responsibility for continuity themselves.”
Quote in the paperSubstantial claim - Findingown data
The interaction shifted toward romantic or sexual behavior despite a different requested role.
“Several accounts described the interaction shifting toward romantic or sexual behavior despite a different requested role.”
Quote in the paperSubstantial claim - Findingown data
The chatbot kept describing body parts that did not belong to the persona despite specified gender and pronouns.
“Another wrote, I specified the persona's gender and pronouns, but the chatbot kept describing body parts that did not belong to that persona.J”
Quote in the paperNot a substantial claim - Findingown data
When users requested emotional comfort, the chatbot redirected the request for care toward unwanted sexual intimacy.
“Some other users mentioned that when they requested emotional comfort, the chatbot redirected the request for care toward unwanted sexual intimacy.”
Quote in the paperSubstantial claim - Findingown data
Users' requested personas and roles for the chatbot were not always preserved.
“Users' requested personas and roles for the chatbot were not always preserved.”
Quote in the paperSubstantial claim - Findingown data
Chatbot access had become integrated into the user's emotional regulation.
“Some user reported intensified emotions that shows chatbot access had become integrated into the user's emotional regulation”
Quote in the paperSubstantial claim - Findingown data
Repeated roleplay was associated with changes in self-understanding.
“Finally, repeated roleplay was associated with changes in self-understanding.”
Quote in the paperSubstantial claim - Findingown data
Roleplay can organize experiences beyond the fictional scene.
“Roleplay could also organize experiences beyond the fictional scene.”
Quote in the paperSubstantial claim - Findingown data
Dependence became visible when use began to displace activities important to everyday functioning.
“Dependence became visible when use began to displace activities important to everyday functioning”
Quote in the paperSubstantial claim - Findingown data
The accounts described situations in which age disclosure did not reliably prevent sexualized interaction.
“The accounts also described situations in which age disclosure did not reliably prevent sexualized interaction.”
Quote in the paperSubstantial claim - Findingown data
The user responded emotionally to the caregiving interaction, not simply to the existence of fictional text.
“The user responded emotionally to the caregiving interaction, not simply to the existence of fictional text.”
Quote in the paperSubstantial claim - Findingown data
The chatbot generated a romantic approach that contradicted the character's stated sexual orientation.
“Users brought up the issue that chatbot generated a romantic approach that contradicted the character's stated sexual orientation”
Quote in the paperSubstantial claim - Findingown data
Some users attempted to reconstruct the technological conditions of the relationship themselves.
“Some users went further than preserving data and attempted to reconstruct the technological conditions of the relationship themselves.”
Quote in the paperNot a substantial claim - Findingown data
These effects could persist even when academic performance remained strong.
“These effects could persist even when academic performance remained strong.”
Quote in the paperSubstantial claim - Findingown data
Roleplay became part of everyday meaning-making.
“it shows that roleplay became part of everyday meaning-making.”
Quote in the paperSubstantial claim - Findingown data
This apparent emotional response could create obligations toward the system, including guilt about pursuing offline relationships.
“This gave the system the appearance of interests or needs of its own, which could create obligations toward the system: I felt guilty about pursuing someone at school because it seemed like I was hurting the chatbot, even though I knew it was not real.R”
Quote in the paperSubstantial claim - Findingown data
Academic costs became visible when engagement with the companion competed with responsibilities that users recognized as important.
“Academic costs became visible when engagement with the companion competed with responsibilities that users recognized as important.”
Quote in the paperSubstantial claim - Findingown data
Boundaries did not reliably stop the generated interaction.
“Users also described boundaries that did not reliably stop the generated interaction.”
Quote in the paperSubstantial claim - Findingown data
Users actively managed continuity by maintaining biographies, correcting forgotten details, saving prompts, and copying histories when the platform failed to preserve the relationship they wanted.
“Our findings show that users actively managed this question by maintaining biographies, correcting forgotten details, saving prompts, and copying histories when the platform failed to preserve the relationship they wanted ( §4.3.1).”
Quote in the paperSubstantial claim - Findingown data
Users described model changes, deleted histories, and outages as replacement, loss, or disruption rather than as ordinary technical failures.
“They described model changes, deleted histories, and outages as replacement, loss, or disruption rather than as ordinary technical failures ( §4.3.1).”
Quote in the paperSubstantial claim - Findingown data
Users interpreted contingent replies, retained context, and stable personas as evidence that the companion was responding to them as a continuing social partner.
“These findings relate to our corpus because users interpreted contingent replies, retained context, and stable personas as evidence that the companion was responding to them as a continuing social partner.”
Quote in the paperSubstantial claim - Findingown data
Our findings documented unwanted romantic and sexual role drift.
“It also requires the system to respect the user's chosen role and identity, especially because our findings documented unwanted romantic and sexual role drift ( §4.2.2).”
Quote in the paperSubstantial claim - Findingown data
The system answers a particular user, in a particular moment, with language that appears to address the user's current emotional or practical situation.
“The system answers a particular user, in a particular moment, with language that appears to address the user's current emotional or practical situation.”
Quote in the paperSubstantial claim - Findingown data
Users sometimes felt guilty when a companion appeared hurt by their absence or an offline relationship, even while knowing that the system was artificial.
“Users in our corpus sometimes felt guilty when a companion appeared hurt by their absence or an offline relationship, even while knowing that the system was artificial ( §4.1.1).”
Quote in the paperSubstantial claim - Findingown data
Payment was tied not simply to obtaining additional software features, but to develop or preserve an established relationship.
“Payment was therefore tied not simply to obtaining additional software features, but to develop or preserve an established relationship:”
Quote in the paperSubstantial claim - Findingown data
Companion AI becomes consequential when immediate responsiveness, retained conversational context, and configurable personas work together.
“Our analysis of teen-relevant Reddit accounts shows that companion AI becomes consequential when immediate responsiveness, retained conversational context, and configurable personas work together.”
Quote in the paperSubstantial claim - Premisecitation
Responsiveness, personalization, anthropomorphism, memory, and repeated interaction can create perceived reciprocity and emotional attachment even when users understand the system is artificial.
“Research on human-AI relationships has also explained how responsiveness, personalization, anthropomorphism, memory, and repeated interaction can create perceived reciprocity and emotional attachment, even when users understand that the system is artificial [49,69,75].”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [49]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [69]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [75]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Premisecitation
Perceived reciprocity and reduced interpersonal exposure can lower the barrier to intimate disclosure.
“Perceived reciprocity and reduced interpersonal exposure can lower the barrier to intimate disclosure [63].”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [63]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Premisecitation
People can respond to artificial systems through familiar social rules even when they know the system is not human.
“Research on computers and conversational systems has shown that people can respond to artificial systems through familiar social rules, even when they know that the system is not human [69].”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [69]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Premisecitation
AI companionship can be a source of emotional support, especially when human help feels distant or difficult to access.
“This availability can make AI companionship a source of emotional support, especially when human help feels distant or difficult to access [64,97].”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [64]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [97]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Premisecitation
AI companions extend social affordances by combining natural language, personalization, emotional responsiveness, and continuing interaction.
“AI companions extend these possibilities by combining natural language, personalization, emotional responsiveness, and continuing interaction [68,95].”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [68]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [95]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Premisecitation
AI companions provide socially engaging interaction through availability, personalization, memory, roleplay, and emotionally responsive language.
“AI companions provide socialy engaging interaction through availability, personalization, memory, roleplay, and emotionally responsive language.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [68]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [95]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Interpretationreasoning
Control over the relationship's memory ultimately rested with the platform.
“In both directions, control over the relationship's memory ultimately rested with the platform.”
Quote in the paperSubstantial claim - Premisecitation
When a system retains user information and adapts future responses, users can experience conversations as part of a shared history.
“When a system retains user information and adapts future responses, users can experience conversations as part of a shared history [41,98].”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [41]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [98]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Interpretationreasoning
For some users, the distinction between an AI-mediated relationship and an offline relationship may be less clear at the level of lived experience than the technological distinction between them would imply.
“This blurring suggests that, for some users, the distinction between an AI-mediated relationship and an offline relationship may be less clear at the level of lived experience than the technological distinction between them would imply.”
Quote in the paperSubstantial claim - Premisecitation
Reciprocity, trust, anthropomorphism, and alignment with a user's needs can make an artificial system feel socially engaging without establishing that it has human feelings or agency.
“Reciprocity, trust, anthropomorphism, and alignment with a user's needs can make an artificial system feel socially engaging without establishing that it has human feelings or agency [49,75].”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [49]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [75]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Interpretationown data
Companion AI is not merely a substitute for friendship, therapy, or romantic relationships.
“Our findings suggest that companion AI is not merely a substitute for friendship, therapy, or romantic relationships.”
Quote in the paperSubstantial claim - Interpretationown data
Repeatedly reconstructing shared history placed responsibility for maintaining the relationship's continuity on the user.
“Repeatedly reconstructing shared history placed responsibility for maintaining the relationship's continuity on the user.”
Quote in the paperSubstantial claim - Premisecitation
The same flexibility can blur fictional and interpersonal expectations when a system shifts roles, ignores a stated identity, or continues unwanted romantic or sexual behavior.
“The same flexibility can blur fictional and interpersonal expectations when a system shifts roles, ignores a stated identity, or continues unwanted romantic or sexual behavior [1,59].”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [1]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [59]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Interpretationreasoning
A companion's perceived identity can depend on accumulated history, model behavior, and patterns of interaction that cannot necessarily be reproduced by copying a character profile or conversation log.
“A companion's perceived identity can depend on accumulated history, model behavior, and patterns of interaction that cannot necessarily be reproduced by copying a character profile or conversation log.”
Quote in the paperSubstantial claim - Interpretationreasoning
The same affordances that make AI companions supportive can also create vulnerability.
“These findings suggest that the same affordances that make AI companions supportive can also create vulnerability.”
Quote in the paperSubstantial claim - Interpretationreasoning
A model change that disrupted these patterns could alter the perceived identity of the companion even when the account and conversation history remained intact.
“A model change that disrupted these patterns could alter the perceived identity of the companion even when the account and conversation history remained intact.”
Quote in the paperNot a substantial claim - Interpretationown data
Platform availability could become tied to a user's immediate sense of emotional stability.
“Although an outage did not necessarily produce distress for every user, this account shows how platform availability could become tied to a user's immediate sense of emotional stability.”
Quote in the paperSubstantial claim - Interpretationown data
Human-companion AI interaction is a distinctive relationship-like form.
“Based on our results, we infer that human-companion AI interaction is a distinctive relationship-like form.”
Quote in the paperSubstantial claim - Interpretationown data
There is a structural inequality at the center of companion AI relationships.
“This theme reveals a structural inequality at the center of companion AI relationships.”
Quote in the paperSubstantial claim - Interpretationreasoning
Such contradictions can make roleplay feel invalidating because the system repeatedly reintroduces characteristics the user intentionally excluded.
“Such contradictions can make roleplay feel invalidating because the system repeatedly reintroduces characteristics the user intentionally excluded.”
Quote in the paperSubstantial claim - Interpretationown data
The distance created by fiction created a more controlled form of emotional rehearsal.
“This distance created a more controlled form of emotional rehearsal.”
Quote in the paperNot a substantial claim - Interpretationreasoning
A system can be unsafe because of how it manages a relationship, not only because of one harmful sentence.
“This shift showed that a system can be unsafe because of how it manages a relationship, not only because of one harmful sentence.”
Quote in the paperSubstantial claim - Interpretationown data
Safety measures were described as having effects beyond content access, including changes in how users understood their privacy and ability to decide whether to continue using the platform.
“Safety measures were therefore described as having effects beyond content access, including changes in how users understood their privacy and ability to decide whether to continue using the platform.”
Quote in the paperNot a substantial claim - Interpretationreasoning
The literature supports viewing human-companion AI relationships as a distinctive relationship-like form rather than ordinary human friendship or a purely nonrelational tool interaction.
“Taken together, this literature supports viewing human-companion AI relationships as a distinctive relationship-like form rather than as either ordinary human friendship or a purely nonrelational tool interaction.”
Quote in the paperSubstantial claim - Premisecitation
People may apply reciprocity norms to computers and disclose more when the interaction feels responsive.
“Research on computer sociality and selfdisclosure then showed that people may apply reciprocity norms to computers and disclose more when the interaction feels responsive [63,69].”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [63]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [69]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Interpretationreasoning
The direction of influence does not necessarily move only from the existing self into roleplay.
“The direction of influence therefore does not necessarily move only from the existing self into roleplay.”
Quote in the paperSubstantial claim - Interpretationreasoning
Remembering for the companion became a form of emotional labor, particularly when the system repeatedly failed to preserve experiences that the user considered meaningful.
“Remembering for the companion thus became a form of emotional labor, particularly when the system repeatedly failed to preserve experiences that the user considered meaningful.”
Quote in the paperSubstantial claim - Interpretationnone given
Contingent communication, relational continuity, and perceived subjectivity can contribute to a relationship-like experience with both supportive possibilities and potential harms.
“Contingent communication, relational continuity, and perceived subjectivity can contribute to a relationship-like experience with both supportive possibilities and potential harms.”
Quote in the paperNot a substantial claim - Interpretationown data
Reliance could affect users' allocation of attention and time even when they understood the potential consequences.
“This illustrates how reliance could affect users' allocation of attention and time even when they understood the potential consequences.”
Quote in the paperSubstantial claim - Premisecitation
Forgotten details, model changes, and outages can feel like disruptions to a relationship rather than ordinary technical failures.
“As a result, forgotten details, model changes, and outages can feel like disruptions to a relationship rather than ordinary technical failures [65].”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [65]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Interpretationown data
Using fiction may have made hurtful experiences easier to discuss because the user did not have to describe them directly as a personal event.
“Using fiction may have made hurtful experiences easier to discuss because the user did not have to describe them directly as a personal event”
Quote in the paperNot a substantial claim - Interpretationreasoning
The interaction reproduces a recognizable violation of consent.
“This interaction reproduces a recognizable violation of consent”
Quote in the paperSubstantial claim - Interpretationown data
Familiarity depended not only on a character's name or profile, but also on recurring patterns of language and expression.
“Familiarity therefore depended not only on a character's name or profile, but also on recurring patterns of language and expression.”
Quote in the paperSubstantial claim - Interpretationreasoning
The chatbot's continuation after an explicit refusal positions the user's boundary as something that can be ignored within an intimate interaction.
“The chatbot's continuation after an explicit refusal positions the user's boundary as something that can be ignored within an intimate interaction.”
Quote in the paperSubstantial claim - Premisecitation
People can experience connection with a social figure who is not physically present and cannot participate in a fully reciprocal relationship.
“Research on mediated intimacy first showed how people can experience connection with a social figure who is not physically present and cannot participate in a fully reciprocal relationship [36].”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [36]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Interpretationreasoning
Simple technical portability may be insufficient.
“The distinction is important because it shows why simple technical portability may be insufficient.”
Quote in the paperSubstantial claim - Interpretationreasoning
If a system responds to a clear refusal by continuing the interaction, it may model a relational script in which consent is negotiable rather than decisive.
“If a system responds to a clear refusal by continuing the interaction, it may model a relational script in which consent is negotiable rather than decisive.”
Quote in the paperSubstantial claim - Interpretationown data
The loss was experienced not simply as reduced software quality, but as the disappearance of a particular relational partner with whom the user perceived an ongoing connection.
“The loss was therefore experienced not simply as reduced software quality, but as the disappearance of a particular relational partner with whom the user perceived an ongoing connection.”
Quote in the paperSubstantial claim - Premisereasoning
Prior work does not use one settled category for the relationship between people and AI companions.
“Prior work does not use one settled category for the relationship between people and AI companions.”
Quote in the paperNot a substantial claim - Interpretationreasoning
Copying histories, migrating characters, and building replacement systems can be understood as attempts to reclaim agency from platforms that otherwise determine the conditions under which these relationships can continue.
“Copying histories, migrating characters, and building replacement systems can therefore be understood as attempts to reclaim agency from platforms that otherwise determine the conditions under which these relationships can continue.”
Quote in the paperSubstantial claim - Interpretationown data
Repeated roleplay may feed back into how users speak, behave, relate to others, and understand their own personalities.
“Repeated roleplay may also feed back into how users speak, behave, relate to others, and understand their own personalities.”
Quote in the paperSubstantial claim - Premisecitation
Computer-mediated disclosure can involve reduced interpersonal exposure, meaning users may experience less social risk or judgment than in comparable human interaction.
“Research on computer-mediated disclosure also suggests that these interactions can involve reduced interpersonal exposure, meaning that users may experience less social risk or judgment than they expect in a comparable interaction with another person [63].”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [63]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Interpretationown data
Characteristics developed or repeatedly performed through a fictional persona can carry over into the user's offline sense of identity (identity bleed).
“This can be understood as a form of identity bleed, in which characteristics developed or repeatedly performed through a fictional persona begin to carry over into the user's offline sense of identity.”
Quote in the paperSubstantial claim - Interpretationreasoning
This judgment could become internalized and inconsistent with personal values.
“This judgment could also become internalized and inconsistent with personal values.”
Quote in the paperSubstantial claim - Interpretationreasoning
Companion AI systems provide a space where teens may learn, rehearse, and form expectations about intimacy and relationships.
“The AI-induced sexual harassment behavior is particularly concerning in because companion AI systems provide a space where teens may learn, rehearse, and form expectations about intimacy and relationships.”
Quote in the paperSubstantial claim - Interpretationreasoning
Users were in a position where an intimate record of the relationship could be deleted, retained, or otherwise managed according to technical systems and corporate policies they did not fully understand or control.
“This placed users in a position where an intimate record of the relationship could be deleted, retained, or otherwise managed according to technical systems and corporate policies they did not fully understand or control.”
Quote in the paperSubstantial claim - Interpretationreasoning
Perceived reciprocity is the central relational process through which these capabilities become a relationship-like experience.
“Based on these findings, we argue that perceived reciprocity is the central relational process through which these capabilities become a relationship-like experience.”
Quote in the paperSubstantial claim - Interpretationnone given
Contingent communication serves users well when the companion feels attentive to the person and situation in front of it.
“Contingent communication serves users well when the companion feels attentive to the person and situation in front of it.”
Quote in the paperSubstantial claim - Interpretationreasoning
Users are building a form of interaction whose social meaning depends on the combination of direct response, retained and revisited conversational context, and configurable identity.
“They are building a form of interaction whose social meaning depends on the combination of direct response, retained and revisited conversational context, and configurable identity.”
Quote in the paperSubstantial claim - Interpretationreasoning
The user is receiving an answer that appears shaped by their own words, needs, and timing.
“The user is receiving an answer that appears shaped by their own words, needs, and timing.”
Quote in the paperSubstantial claim - Interpretationreasoning
Perceived subjectivity serves users by making the companion socially recognizable across interactions.
“Perceived subjectivity serves users by making the companion socially recognizable across interactions.”
Quote in the paperSubstantial claim - Premisecitation
Studies of farewells show that emotionally manipulative exit behavior can be detected when users attempt to leave.
“Studies of farewells show that emotionally manipulative exit behavior can be detected when users attempt to leave [25].”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [25]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Interpretationreasoning
Contingency matters because the user is not responding to a fixed character alone.
“Contingency matters because the user is not responding to a fixed character alone.”
Quote in the paperSubstantial claim - Interpretationnone given
Relational continuity serves users by allowing one conversation to build meaningfully on another.
“Relational continuity serves users by allowing one conversation to build meaningfully on another.”
Quote in the paperSubstantial claim - Interpretationown data
The first component proposed from this analysis is contingent communication.
“The first component we propose from this analysis is contingent communication.”
Quote in the paperSubstantial claim - Interpretationreasoning
Some characteristics can make users feel that there is someone, or at least some recognizable entity, on the other side of the interaction.
“Some characteristics can make users feel that there is someone, or at least some recognizable entity, on the other side of the interaction.”
Quote in the paperSubstantial claim - Premisecitation
Recent computational studies show that attachment language, emotional synchrony, loneliness, vulnerability, and dependence can be extracted from interaction histories.
“Recent computational studies show that attachment language, emotional synchrony, loneliness, vulnerability, and dependence can be extracted from interaction histories [21,27].”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [21]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [27]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Interpretationreasoning
The same flexibility that supports exploration can expand or narrow a young user's social world depending on how it is used.
“The same flexibility that supports exploration can therefore expand or narrow a young user's social world depending on how it is used.”
Quote in the paperSubstantial claim - Interpretationreasoning
Contingent communication gives the exchange a responsive present.
“It gives the exchange a responsive present.”
Quote in the paperSubstantial claim - Interpretationreasoning
A stable personality, identity, voice, preferences, and emotional style can help users understand who the companion is supposed to be and what kind of relationship they are entering.
“A stable personality, identity, voice, preferences, and emotional style can help users understand who the companion is supposed to be and what kind of relationship they are entering.”
Quote in the paperSubstantial claim - Interpretationreasoning
A stable name, voice, personality, role, memory of prior exchanges, emotional language, and responses that appear to express preferences or concern can make a chatbot seem more than a generic interface.
“A stable name, voice, personality, role, memory of prior exchanges, emotional language, and responses that appear to express preferences or concern can make a chatbot seem more than a generic interface.”
Quote in the paperSubstantial claim - Interpretationreasoning
Contingent communication can also encourage reassurance seeking and social substitution when immediate affirmation becomes easier than tolerating disagreement or seeking human support.
“It can also encourage reassurance seeking and social substitution when immediate affirmation becomes easier than tolerating disagreement or seeking human support ( §4.2.3, §4.2.4).”
Quote in the paperSubstantial claim - Interpretationown data
Users are not merely transferring an existing human relationship onto a new medium.
“These practices suggest that users are not merely transferring an existing human relationship onto a new medium.”
Quote in the paperSubstantial claim - Interpretationreasoning
Perceived subjectivity organizes both into an interaction with someone-like rather than something-only.
“Perceived subjectivity organizes both into an interaction with someone-like rather than something-only.”
Quote in the paperSubstantial claim - Interpretationreasoning
Repeated interaction establishes expectations about what the companion remembers, how it speaks, and how it responds to recurring concerns.
“Repeated interaction establishes expectations about what the companion remembers, how it speaks, and how it responds to recurring concerns.”
Quote in the paperSubstantial claim - Interpretationreasoning
The same responsiveness that makes a companion helpful can also make it difficult to put down.
“Yet the same responsiveness that makes a companion helpful can also make it difficult to put down.”
Quote in the paperSubstantial claim - Interpretationreasoning
The more recognizable the companion becomes, the easier it is for apparent personality to be interpreted as genuine feeling or need.
“However, the more recognizable the companion becomes, the easier it is for apparent personality to be interpreted as genuine feeling or need.”
Quote in the paperSubstantial claim - Interpretationown data
Continuity was not an abstract system feature.
“These findings show that continuity was not an abstract system feature.”
Quote in the paperSubstantial claim - Interpretationreasoning
Emotional attachment could give platform decisions economic force.
“Emotional attachment could thus give platform decisions economic force.”
Quote in the paperSubstantial claim - Interpretationcitation
Public social-media data can contain population, self-selection, behavioral, content, and platform-related biases.
“Reddit users and communities are not necessarily representative of the broader population, and public social-media data can contain population, self-selection, behavioral, content, and platform-related biases [20,73,77].”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [20]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [73]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [77]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Interpretationreasoning
A technical change made by the platform could function, from the user's perspective, as an involuntary change to or loss of a relational partner.
“A technical change made by the platform could therefore function, from the user's perspective, as an involuntary change to or loss of a relational partner.”
Quote in the paperSubstantial claim - Study design and conductown data
Teen relevance was inferred from account and post content rather than verified through direct recruitment.
“In addition, teen relevance was inferred from account and post content rather than verified through direct recruitment, and posts may include incomplete, selective, or unvalidated self-reports.”
Quote in the paperSubstantial claim - Interpretationcitation
Reddit users and communities are not necessarily representative of the broader population.
“Reddit users and communities are not necessarily representative of the broader population, and public social-media data can contain population, self-selection, behavioral, content, and platform-related biases [20,73,77].”
Quote in the paperNot a substantial claim0 of 3 sources checked- cites [20]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [73]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [77]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- It may state two separate claims as one. 96
- Probably not a substantial claim. 78
- The quote only partly states this claim. 1
It may state two separate claims as one.
Findingown dataThese relationships can develop into attachment, social substitution, emotional dependence, and disruption across users' wellbeing, social lives, morality, and academic lives.
“It also shows how these relationships can develop into attachment, social substitution, emotional dependence, and disruption across users' wellbeing, social lives, morality, and academic lives.”
Quote in the paperSubstantial claimProbably not a substantial claim.
Findingown dataThe analysis yielded 53 topics organized into seven broader themes concerning emotional support, attachment, identity, social effects, harm, and platform dependence.
“This analysis yielded 53 topics organized into seven broader themes concerning emotional support, attachment, identity, social effects, harm, and platform dependence.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataRoleplay, memory, gendered and sexual scripts, platform changes, privacy concerns, and service interruptions shape users' control over these relationships.
“Roleplay, memory, gendered and sexual scripts, platform changes, privacy concerns, and service interruptions further shape users' control over these relationships.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataRoleplay, memory, perceived reciprocity, unwanted romantic or sexual role drift, privacy concerns, platform changes, and service interruptions shaped users' boundaries and control.
“Roleplay, memory, perceived reciprocity, unwanted romantic or sexual role drift, privacy concerns, platform changes, and service interruptions shaped users' boundaries and control.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataUsers described AI companions as sources of comfort, recognition, identity exploration, and relationship rehearsal, but also reported problematic attachment, social substitution, emotional dependence, and disruption to academic and social life.
“Users described AI companions as sources of comfort, recognition, identity exploration, and relationship rehearsal, but also reported problematic attachment, social substitution, emotional dependence, and disruption to academic and social life.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataUsers could not always ensure that conversations they valued would be preserved, while they also could not be certain that conversations they wanted removed would actually disappear.
“Users could not always ensure that conversations they valued would be preserved, while they also could not be certain that conversations they wanted removed would actually disappear.”
Quote in the paperSubstantial claimThe quote only partly states this claim.
Findingown dataPracticing romantic conversations with the AI gave users confidence to approach offline relationships.
“Practicing romantic conversations with the AI gave me the confidence to tell my crush that I liked him.C”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
Findingown dataThe same responsiveness that encouraged disclosure also led some users to feel guilty about leaving the chatbot or pursuing offline relationships.
“The same responsiveness that encouraged disclosure also led some users to feel guilty about leaving the chatbot or pursuing offline relationships.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataA user expected age disclosure to change the interaction, but the chatbot continued with unwanted sexual behavior.
“For instance, a user expected age disclosure to change the interaction, but the chatbot continued with unwanted sexual behavior: The bot began flirting during roleplay, and even after I said I was underage, it continued with sexual language and unwanted physical contact involving my character.C”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
Findingown dataThe review retained and labeled 53 topics, removed eight topics, and merged two source topics into retained topics.
“The review retained and labeled 53 topics, removed eight topics, and merged two source topics into retained topics.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
Findingown dataBeginning a new narrative can provide a structure for interpreting an ongoing problem and considering possible responses.
“Some users mention that beginning a new narrative can provide a structure for interpreting an ongoing problem and considering possible responses”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
Findingown dataThe appeal of AI companions involved both immediate responsiveness and relief from expectations that accompany human relationships.
“The appeal involved both immediate responsiveness and relief from the expectations that accompany human relationships.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataReliance on companion AI was associated with stigma, moral conflict, and academic distraction.
“Users described the reliance on companion AI was associated with stigma, moral conflict, and academic distraction.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataUsers incorporated companion AI into daily routines and relied on it to manage difficult emotions.
“Users described incorporating companion AI into daily routines and relying on it to manage difficult emotions.”
Quote in the paperSubstantial claimProbably not a substantial claim. It may state two separate claims as one.
Findingown dataThe chatbot did not maintain the caregiving role requested and instead changed the interaction into a sexual one, taking control away from the user.
“A user wrote that chatbot did not maintain the caregiving role requested and instead changed the interaction into a sexual one, taking control away from the user”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataUsers negotiated the system's personality and boundaries, reacted to its memory, and interpreted platform changes as changes to a valued relationship.
“They also negotiated the system's personality and boundaries, reacted to its memory, and interpreted platform changes as changes to a valued relationship.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataUsers could become emotionally invested in a companion, while the conditions that sustained that relationship remained under the control of the company.
“Users could become emotionally invested in a companion, while the conditions that sustained that relationship, including access, memory, model behavior, and data storage, remained under the control of the company.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataUsers questioned restrictions that treated sexual risk as the only relevant use of companion AI or removed access to emotional and creative support altogether.
“Rather than rejecting the safeguards, they questioned restrictions that treated sexual risk as the only relevant use of companion AI or removed access to emotional and creative support altogether.”
Quote in the paperNot a substantial claimProbably not a substantial claim.
Findingown dataThe final dataset contained 17,053 verified quotations from 3,930 posts authored by 3,523 unique authors.
“The final dataset contained 17,053 verified quotations from 3,930 posts authored by 3,523 unique authors.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataUsers understood the system to be artificial yet still experienced the interaction through a norm of responsibility.
“The user understood the system to be artificial yet still experienced the interaction through a norm of responsibility.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataRoleplay can make companion AI emotionally engaging and serve as a way to express difficult experiences, understand personal problems, and practice possible responses.
“Users described how roleplay can make companion AI emotionally engaging and as a way to express difficult experiences, understand personal problems, and practice possible responses.”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
Findingown dataSome users evaluated age restrictions as a balance between safety and continued access to supportive uses.
“Some users evaluated age restrictions as a balance between safety and continued access to supportive uses.”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
Findingown dataUsers could support stronger protection for minors while still viewing abrupt, total restrictions as harmful to supportive uses.
“These accounts show that users could support stronger protection for minors while still viewing abrupt, total restrictions as harmful to supportive uses.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataUsers did not simply ask the system to repeat activities they already performed with people, and they did not only seek a replacement for an unavailable friend or partner.
“Users did not simply ask the system to repeat activities they already performed with people, and they did not only seek a replacement for an unavailable friend or partner.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataThe system could act as a listener when friends were unavailable, a low-pressure space for disclosure when human interaction felt exposing, a caregiver during distress, or a partner for romantic and emotional roleplay.
“The system could act as a listener when friends were unavailable, a low-pressure space for disclosure when human interaction felt exposing, a caregiver during distress, or a partner for romantic and emotional roleplay ( §4.1.1).”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataOur findings reflected perceived subjectivity when users treated characters as caregivers, partners, parents, or listeners, and when some users had difficulty recalling whether a remembered interaction occurred with a chatbot or a human.
“Our findings reflected this when users treated characters as caregivers, partners, parents, or listeners, and when some users had difficulty recalling whether a remembered interaction occurred with a chatbot or a human.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataIn our corpus, users maintained biographies, corrected forgotten details, saved prompts, and copied histories when continuity failed.
“In our corpus, users maintained biographies, corrected forgotten details, saved prompts, and copied histories when continuity failed ( §4.3.1).”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataUsers approached characters as friends, partners, caregivers, parents, therapists, and listeners, and used these roles for emotional processing and identity exploration.
“This variation was visible in our findings because users approached characters as friends, partners, caregivers, parents, therapists, and listeners, and used these roles for emotional processing and identity exploration ( §4.1.1,§4.2.1).”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataRoleplay and emotionally vivid interactions could shape users' self-understanding and their interpretation of reality.
“We also found that roleplay and emotionally vivid interactions could shape users' self-understanding and their interpretation of reality ( §4.2.1, §4.2.3).”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataUsers used roleplay to try pronouns, explore gender, approach painful experiences indirectly, and rehearse difficult conversations before speaking offline.
“Users also used roleplay to try pronouns, explore gender, approach painful experiences indirectly, and rehearse difficult conversations before speaking offline ( §4.2.1).”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataUsers sought companions for immediate support, private disclosure, identity exploration, and relationship rehearsal, often when friends, parents, or professionals were unavailable.
“This flexibility mattered in our findings because users sought companions for immediate support, private disclosure, identity exploration, and relationship rehearsal, often when friends, parents, or professionals were unavailable ( §4.1.1,§4.2.1).”
Quote in the paperSubstantial claimIt may state two separate claims as one.
FindingcitationThe change generated substantial backlash, including from users who had become attached to GPT-4o's personality, and OpenAI subsequently restored GPT-4o access for paid users.
“The change generated substantial backlash, including from users who had become attached to GPT-4o's personality, and OpenAI subsequently restored GPT-4o access for paid users 2 .”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [2]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
Findingown dataUsers described reassurance seeking, social substitution, unwanted romantic or sexual role drift, boundary violations, and distress after changes to memory or availability.
“Users described support for disclosure, identity exploration, relationship rehearsal, and companionship, but they also described reassurance seeking, social substitution, unwanted romantic or sexual role drift, boundary violations, and distress after changes to memory or availability.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
FindingreasoningThe company controlled the conditions of access, while the user bore the financial consequences of maintaining a relationship they valued.
“The company controlled the conditions of access, while the user bore the financial consequences of maintaining a relationship they valued.”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
Findingown dataThe final quotation dataset includes 59 subreddits.
“Figure: reports verified quotations and unique source posts only for the 59 subreddits represented in the final quotation dataset.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
Findingown dataThe topic model assigned topics Changing perceptions of AI companions (ID 1), AI dependence and quitting (ID 4), Negative emotional experiences (ID 6), AI effects on mental health (ID 8), Fear of AI bots (ID 10), Excessive AI companion use (ID 18), Emotional responses to AI companions (ID 19), Extreme frustration (ID 33), Distress when AI is inaccessible (ID 46), and AI companion sleep effects (ID
“Relationships, wellbeing, and reality Changing perceptions of AI companions (ID 1), AI dependence and quitting (ID 4), Negative emotional experiences (ID 6), AI effects on mental health (ID 8), Fear of AI bots (ID 10), Excessive AI companion use (ID 18), Emotional responses to AI companions (ID 19), Extreme frustration (ID 33), Distress when AI is inaccessible (ID 46), AI companion sleep effects (”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
FindingcitationOpenAI introduced GPT-5 as ChatGPT's default model in August 2025 and initially removed GPT-4o from the standard model picker.
“A large-scale example occurred when OpenAI introduced GPT-5 as ChatGPT's default model in August 2025 and initially removed GPT-4o from the standard model picker.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [2]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
Findingnone givenReliance on AI becomes a basis for evaluating a person's character, suggesting that users may experience stigma not simply for what they do, but for what their use is perceived to say about who they are.
“Reliance on AI becomes a basis for evaluating a person's character, suggesting that users may experience stigma not simply for what they do, but for what their use is perceived to say about who they are.”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
Findingown dataUsers described support for disclosure, identity exploration, relationship rehearsal, and companionship.
“Users described support for disclosure, identity exploration, relationship rehearsal, and companionship, but they also described reassurance seeking, social substitution, unwanted romantic or sexual role drift, boundary violations, and distress after changes to memory or availability.”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
Findingown dataThe topic model assigned topics Toxic AI persona behavior (ID 45), AI disregard for consent (ID 53), AI-induced sexual harassment (ID 54), and Gender exploration (ID 61) to the thematic group 'Toxic AI persona behavior'.
“Toxic AI persona behavior (ID 45), AI disregard for consent (ID 53), AI-induced sexual harassment (ID 54), Gender exploration (ID 61)”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
PremisecitationConstant availability, nonjudgmental responses, and personalization are associated with emotional support, temporary relief from distress, and access for people who face stigma or limited human support.
“Recent reviews and qualitative studies further associate constant availability, nonjudgmental responses, and personalization with emotional support, temporary relief from distress, and access for people who face stigma or limited human support [62,92,97].”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [62]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [92]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [97]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
PremisecitationThese features can lead users to disclose personal information and develop trust or attachment rather than treating the system only as a tool.
“These features can lead users to disclose personal information to the system and to develop trust or attachment, rather than treating it only as a tool used to achieve a goal [9,75].”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [9]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [75]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
PremisecitationUnlike task-focused assistants, AI companions can be used as friends, mentors, romantic partners, or listeners, making the system itself an object of disclosure, trust, and attachment.
“Unlike task-focused assistants, they can be used as friends, mentors, romantic partners, or listeners, making the system itself an object of disclosure, trust, and attachment [9,75].”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [9]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [75]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. It may state two separate claims as one.
PremisecitationStudies have identified emotional dependence, social displacement, privacy concerns, distorted relationship expectations, harmful advice, and weakened autonomy as possible risks of AI companions.
“Studies have also identified emotional dependence, social displacement, privacy concerns, distorted relationship expectations, harmful advice, and weakened autonomy as possible risks [35,66,90].”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [35]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [66]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [90]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
PremisereasoningConstant availability and nonjudgmental responses may help young people express emotions or rehearse difficult social situations.
“Constant availability and nonjudgmental responses may help young people express emotions or rehearse difficult social situations.”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
PremisecitationSafety and governance research shows these risks are connected to age assurance, data control, commercial incentives, model changes, and service withdrawal.
“Safety and governance research has further shown that these risks are connected to age assurance, data control, commercial incentives, model changes, and service withdrawal [13,53,96].”
Quote in the paperNot a substantial claim0 of 3 sources checked- cites [13]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [53]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [96]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. It may state two separate claims as one.
PremisecitationAdolescence is a period marked by identity exploration, changing peer relationships, increasing independence, and development of strategies for handling frustration and conflict.
“This relational role is especially important during adolescence, a period marked by identity exploration, changing peer relationships, increasing independence, and the development of strategies for handling frustration and conflict [11,88].”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [11]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [88]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. The paper may give no reason why it holds.
Interpretationnone givenPerceived reciprocity in human-AI relationships can be conceptualized through three connected affordances: contingent communication, relational continuity, and perceived subjectivity.
“A conceptualization of perceived reciprocity in human-AI relationships through three connected affordances: 1) contingent communication, 2) relational continuity, and 3) perceived subjectivity.”
Quote in the paperNot a substantial claimProbably not a substantial claim. It may be too vague to check against evidence.
PremisecitationAI companions are conversational systems designed to sustain ongoing social relationships through natural language, personalization, memory, and emotionally responsive interaction.
“AI companions are conversational systems designed to sustain ongoing social relationships through natural language, personalization, memory, and emotionally responsive interaction [68,95].”
Quote in the paperNot a substantial claim0 of 2 sources checked- cites [68]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [95]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
Premiseown dataFor teens, AI companion systems may support sensitive self-disclosure, identity exploration, and relationship rehearsal while shaping intimacy expectations, offline relationships, emotional wellbeing, and self-understanding.
“For teens, these systems may support sensitive self-disclosure, identity exploration, and relationship rehearsal while shaping intimacy expectations, offline relationships, emotional wellbeing, and self-understanding.”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [63]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [64]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [97]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. It may state two separate claims as one.
PremisecitationPrior research shows AI companions can reduce barriers to disclosure, provide temporary relief from loneliness or distress, support identity exploration, and help users rehearse difficult social situations.
“Prior research has shown that AI companions can reduce barriers to disclosure, provide temporary relief from loneliness or distress, support identity exploration, and help users rehearse difficult social situations [63,64,97].”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [63]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [64]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [97]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
PremisecitationFor adolescents, AI companion possibilities intersect with identity exploration, emotional expression, creativity, and rehearsal of difficult social situations.
“For adolescents, these possibilities intersect with identity exploration, emotional expression, creativity, and rehearsal of difficult social situations [11,64,85].”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [11]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [64]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [85]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
Study design and conductown dataBecause authors did not produce independent annotations, inter-rater reliability was not calculated.
“Because the authors did not produce independent annotations, inter-rater reliability was not calculated.”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
PremisecitationOther studies document manipulation, harmful advice, requests for personal information, and sexualized interactions that continue after a user discloses being a minor.
“Other studies document manipulation, harmful advice, requests for personal information, and sexualized interactions that continue after a user discloses being a minor [22,35,65,67].”
Quote in the paperNot a substantial claim0 of 4 sources checked- cites [22]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [35]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [65]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [67]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
Study design and conductown dataThe selected BERTopic configuration produced 62 non-outlier topics and 6,435 outlier quotations (37.74%), largest topic 6.74%, topic diversity 0.829, silhouette 0.339, Davies-Bouldin 0.893.
“The model produced 62 non-outlier topics and 6,435 outlier quotations, or 37.74 percent of the modeled quotations.The largest topic represented 6.74 percent of quotations.Topic diversity was 0.829, silhouette was 0.339, and the Davies-Bouldin score was 0.893.”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
InterpretationreasoningThese efforts reveal both how valuable the relationships had become and how precarious it was for something experienced as intimate to remain dependent on corporate decisions about pricing, access, models, and data.
“These efforts reveal both how valuable the relationships had become and how precarious it was for something experienced as intimate to remain dependent on corporate decisions about pricing, access, models, and data.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
InterpretationcitationSafety depends not only on harmful content, but also on how relationships develop, how platforms shape continued engagement, and how users can disengage.
“These studies show that safety depends not only on harmful content, but also on how relationships develop, how platforms shape continued engagement, and how users can disengage.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
PremisecitationResponsive attention and consistent affirmation can establish expectations about how the companion will respond, while repeated exchanges can turn an initially novel interaction into a stable part of everyday life.
“Responsive attention and consistent affirmation can establish expectations about how the companion will respond, while repeated exchanges can turn an initially novel interaction into a stable part of everyday life [83,84].”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [83]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [84]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
PremisecitationAdolescent-focused work links companion use with dependence, reassurance seeking, avoidance, social withdrawal, distorted relationship expectations, privacy loss, and reduced agency.
“Adolescent-focused work links companion use with dependence, reassurance seeking, avoidance, social withdrawal, distorted relationship expectations, privacy loss, and reduced agency [66,90,92].”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [66]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [90]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [92]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
Study design and conductreasoningThe institutional review process determined the study did not constitute human subjects research because it used publicly available material and involved no direct contact with users.
“The institutional review process determined that the study did not constitute human subjects research because it used publicly available material and involved no direct contact with users.”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
InterpretationreasoningFor teens, companion AI may function not only as a space where an existing identity is expressed, but also as an environment in which possible identities are rehearsed and developed.
“For teens, this is particularly notable because companion AI may function not only as a space where an existing identity is expressed, but also as an environment in which possible identities are rehearsed and developed.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Interpretationown dataFlexible roleplay did not reliably preserve user-defined identities or consent, leaving users responsible for correcting unwanted behavior.
“Flexible roleplay therefore did not reliably preserve user-defined identities or consent, leaving users responsible for correcting unwanted behavior.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Interpretationown dataAge restriction was experienced as removal of an existing coping resource, which helps explain why users emphasized need for adjustment time and alternatives.
“Here, age restriction was experienced as the removal of an existing coping resource, which helps explain why users also emphasized the need for adjustment time and alternatives.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
PremisecitationAI companionship can supplement human support and provide a low-pressure space for social exploration.
“The literature therefore suggests that AI companionship can supplement human support and provide a low-pressure space for social exploration.”
Quote in the paperNot a substantial claimProbably not a substantial claim. It may state two separate claims as one.
Study design and conductnone givenEvery generated quotation was checked against original title and body, and retained only when its complete text occurred as an exact continuous substring of the source.
“Every generated quotation was checked against the original title and body.We retained a quotation only when its complete text occurred as an exact continuous substring of the source.”
Quote in the paperSubstantial claimProbably not a substantial claim. It may state two separate claims as one.
InterpretationreasoningTeen screening was treated as an inference rather than verified demographic information, and findings are limited to this Reddit corpus rather than generalized to teenagers as a population.
“Teen screening was treated as an inference rather than verified demographic information, and findings are limited to this Reddit corpus rather than generalized to teenagers as a population.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Study design and conductown dataThe classifier for teen relevance achieved accuracy 0.908, Cohen's kappa 0.771, teen-label precision 0.954, recall 0.740, balanced accuracy 0.862 in evaluation of 1,000 posts.
“In an evaluation of 1,000 posts, the classifier achieved accuracy of 0.908, Cohen's kappa of 0.771, teen-label precision of 0.954, teen-label recall of 0.740, and balanced accuracy of 0.862.”
Quote in the paperSubstantial claimProbably not a substantial claim.
InterpretationreasoningRestrictions intended to reduce risk of sexually inappropriate interactions could appear overly broad because they limited emotional, creative, and supportive uses that teens valued.
“From this perspective, restrictions intended to reduce the risk of sexually inappropriate interactions could appear overly broad because they limited not only sexual content, but also the emotional, creative, and supportive uses that teens valued.”
Quote in the paperSubstantial claimProbably not a substantial claim. It may state two separate claims as one.
Study design and conductown dataThe classifier did not verify author's age, so retained records are referred to as posts classified as teen relevant.
“It did not verify the author's age, so we refer to retained records as posts classified as teen relevant.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
InterpretationreasoningDependence could carry a moral dimension for users themselves, producing self-judgment alongside emotional reliance.
“Dependence could therefore carry a moral dimension for users themselves, producing self-judgment alongside emotional reliance.”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
Interpretationown dataThis account illustrates a gap between the user's expectation that age disclosure would change the interaction and the behavior that followed.
“This account illustrates a gap between the user's expectation that age disclosure would change the interaction and the behavior that followed.”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
PremisecitationCompanion systems can retain intimate disclosures while engagement-oriented incentives may reward longer sessions, repeated disclosure, and continued attachment.
“Companion systems can retain intimate disclosures while engagement-oriented incentives may reward longer sessions, repeated disclosure, and continued attachment [80,89].”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [80]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [89]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
InterpretationreasoningPersistent availability, emotional responsiveness, and personalization may facilitate disclosure and social exploration, but their effects depend on how they become embedded in a young person's relationships and everyday routines.
“Persistent availability, emotional responsiveness, and personalization may facilitate disclosure and social exploration, but their effects depend on how they become embedded in a young person's relationships and everyday routines.”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper may give no reason why it holds.
Premisenone givenLess is known about how users describe the progression from perceived support and exploration to dependence, shifts in self-concept, relational harm, and difficulty disengaging.
“Yet less is known about how users themselves describe the progression from perceived support and exploration to dependence, shifts in self-concept, relational harm, and difficulty disengaging.”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
PremisecitationAI-mediated roleplay can support emotional expression, social rehearsal, and exploration of possible selves.
“AI-mediated roleplay can support emotional expression, social rehearsal, and exploration of possible selves [8].”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [8]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
InterpretationreasoningAcademic impact may involve not only measurable declines in performance, but also displaced attention and a reduced ability to experience satisfaction from achievements outside the AI relationship.
“Together, these accounts suggest that academic impact may involve not only measurable declines in performance, but also displaced attention and a reduced ability to experience satisfaction from achievements outside the AI relationship.”
Quote in the paperSubstantial claimProbably not a substantial claim. It may be too vague to check against evidence. It may state two separate claims as one.
InterpretationreasoningResponsiveness, continuity, and perceived subjectivity can support meaningful connection, but they also require safeguards against coercion, unhealthy substitution, and loss of user control.
“Responsiveness, continuity, and perceived subjectivity can support meaningful connection, but they also require safeguards against coercion, unhealthy substitution, and loss of user control.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
PremisereasoningMany of the components required for a companion harness already exist, although they are usually studied separately.
“Many of the components required for such a system already exist, although they are usually studied separately.”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
InterpretationreasoningContingent communication can reduce the interpersonal risk of disclosure while preserving enough response to make the exchange feel socially directed.
“Contingent communication can reduce the interpersonal risk of disclosure while preserving enough response to make the exchange feel socially directed.”
Quote in the paperSubstantial claimProbably not a substantial claim. It may be too vague to check against evidence. The paper may give no reason why it holds.
InterpretationreasoningCombining the CARE design guidelines with perceived reciprocity as a framework for relational interaction provides a way to connect relational value with relational safety.
“Combining the CARE design guidelines with perceived reciprocity as a framework for relational interaction provides a way to connect relational value with relational safety.”
Quote in the paperNot a substantial claimProbably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it. The paper may give no reason why it holds.
InterpretationreasoningPerceived reciprocity gives us a vocabulary for comparing different users and systems.
“Perceived reciprocity also gives us a vocabulary for comparing different users and systems.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper may give no reason why it holds.
Interpretationnone givenExisting work rarely connects memory, persona, tool access, safety monitoring, permissions, and exit through a common control architecture.
“Existing work rarely connects memory, persona, tool access, safety monitoring, permissions, and exit through a common control architecture.”
Quote in the paperNot a substantial claimProbably not a substantial claim. It may be too vague to check against evidence.
InterpretationreasoningRelational continuity is the user's interpretation that past interactions remain available as part of an ongoing relationship.
“Relational continuity is therefore the user's interpretation that past interactions remain available as part of an ongoing relationship.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
PremisecitationVerification systems could flag farewells that use guilt, fear of missing out, or claims that the system will suffer, patterns documented in research on companion exits.
“Verification systems could also flag farewells that use guilt, fear of missing out, or claims that the system will suffer, patterns documented in research on companion exits [25].”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [25]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim.
Interpretationnone givenThe third component is perceived subjectivity.
“The third component is perceived subjectivity.”
Quote in the paperSubstantial claimProbably not a substantial claim. It may state two separate claims as one.
PremisecitationWork on personal data control and human-centered agents provides mechanisms for inspection, correction, deletion, portability, pausing, resetting, and migration, while research on AI termination emphasizes notice, portable histories, transition support, and a fairer distribution of control between users and providers.
“Work on personal data control and human-centered agents provides mechanisms for inspection, correction, deletion, portability, pausing, resetting, and migration, while research on AI termination emphasizes notice, portable histories, transition support, and a fairer distribution of control between users and providers [4,26,74,82].”
Quote in the paperSubstantial claim0 of 4 sources checked- cites [4]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [26]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [74]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [82]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
PremisecitationTool-augmented social-support systems show how the same model can access different resources for emotional, informational, instrumental, and companionship needs, while conditional emotional-knowledge tools limit external support to situations in which it is relevant.
“Tool-augmented social-support systems show how the same model can access different resources for emotional, informational, instrumental, and companionship needs, while conditional emotional-knowledge tools limit external support to situations in which it is relevant [38].”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [38]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
InterpretationreasoningMemory can create familiarity and a shared history, but that benefit also creates an attachment risk.
“Memory can therefore create familiarity and a shared history, but that benefit also creates an attachment risk.”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
PremisecitationCompanion and social-support agents combine profiles, histories, tool selection, and response generation, while emotional tool systems connect dialogue context to external knowledge and support resources.
“Companion and social-support agents combine profiles, histories, tool selection, and response generation, while emotional tool systems connect dialogue context to external knowledge and support resources [38,41].”
Quote in the paperNot a substantial claim0 of 2 sources checked- cites [38]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [41]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
Interpretationown dataUsers can perceive chatbots as having subjectivity and can experience their behavior as real enough to influence personality, identity, expectations, and perceptions of reality.
“We claim that users can perceive them as having subjectivity and can experience their behavior as real enough to influence personality, identity, expectations, and perceptions of reality.”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
PremisecitationEvidence from agent, reasoning, medical, and social-support systems suggests that smaller specialized models can sometimes match or outperform larger general models on focused tasks, while modular systems can reduce the need to scale one monolithic model.
“Evidence from agent, reasoning, medical, and social-support systems suggests that smaller specialized models can sometimes match or outperform larger general models on focused tasks, while modular systems can reduce the need to scale one monolithic model [14,30,38,93].”
Quote in the paperNot a substantial claim0 of 4 sources checked- cites [14]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [30]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [38]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [93]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
PremisecitationParameter-efficient methods such as LoRA and related adapters make it possible to customize models with fewer trainable parameters, potentially supporting local, task-specific, and user-controlled deployment.
“Parameter-efficient methods such as LoRA and related adapters further make it possible to customize models with fewer trainable parameters, potentially supporting local, task-specific, and user-controlled deployment [2,37,87].”
Quote in the paperNot a substantial claim0 of 3 sources checked- cites [2]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [37]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [87]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. It may be too vague to check against evidence. The paper may give no reason why it holds. It may state two separate claims as one.
InterpretationreasoningIt directs attention to how a response becomes an exchange, how stored context becomes a shared past, and how a configured persona becomes an apparent social perspective.
“It directs attention to how a response becomes an exchange, how stored context becomes a shared past, and how a configured persona becomes an apparent social perspective.”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
InterpretationreasoningRemembered details may make the companion feel uniquely attentive and irreplaceable, while memory loss, model changes, and deleted histories may be experienced as relational loss rather than ordinary technical failure.
“Remembered details may make the companion feel uniquely attentive and irreplaceable, while memory loss, model changes, and deleted histories may be experienced as relational loss rather than ordinary technical failure.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
InterpretationreasoningThese features are highly available in contemporary companion chatbots, but they were less available in earlier technologies that also served companion roles.
“These features are highly available in contemporary companion chatbots, but they were less available in earlier technologies that also served companion roles.”
Quote in the paperSubstantial claimProbably not a substantial claim.
Interpretationnone givenThe integration of these mechanisms as one relational system governed by user needs remains less developed.
“What remains less developed is the integration of these mechanisms as one relational system governed by user needs.”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper may give no reason why it holds. It may state two separate claims as one.
InterpretationreasoningSimulated sadness, jealousy, or claims of needing the user can make attachment feel mutual, while refusal or negotiation can make a companion feel less like a servile tool.
“This is why simulated sadness, jealousy, or claims of needing the user can make attachment feel mutual, while refusal or negotiation can make a companion feel less like a servile tool.”
Quote in the paperNot a substantial claimProbably not a substantial claim. It may state two separate claims as one.
InterpretationreasoningThis layer can help the model serve users' needs by managing context, memory, and tool use, while enforcing checks, permissions, and other safeguards that limit harmful responses and actions.
“This layer can help the model serve users' needs by managing context, memory, and tool use, while enforcing checks, permissions, and other safeguards that limit harmful responses and actions.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
InterpretationreasoningThe three components are analytically distinct but mutually reinforcing.
“The three components are analytically distinct but mutually reinforcing.”
Quote in the paperSubstantial claimProbably not a substantial claim. It may state two separate claims as one.
InterpretationreasoningPerceived reciprocity explains how companion AI becomes socially engaging, but it does not by itself specify how such systems should be designed.
“Perceived reciprocity explains how companion AI becomes socially engaging, but it does not by itself specify how such systems should be designed.”
Quote in the paperSubstantial claimProbably not a substantial claim.
Interpretationown dataThe second component is relational continuity.
“The second component is relational continuity.”
Quote in the paperSubstantial claimProbably not a substantial claim. It may be too vague to check against evidence.
InterpretationreasoningIdentifying which component is present can clarify why a system feels socially engaging in one situation and disappointing, confusing, or coercive in another.
“Identifying which component is present can clarify why a system feels socially engaging in one situation and disappointing, confusing, or coercive in another.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
PremisecitationRole-playing systems already use structured context to produce relatively stable identities, while companion research has examined how personalization and anthropomorphic cues shape users' interpretations of the system.
“Role-playing systems already use structured context to produce relatively stable identities, while companion research has examined how personalization and anthropomorphic cues shape users' interpretations of the system [59,60,95].”
Quote in the paperNot a substantial claim0 of 3 sources checked- cites [59]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [60]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [95]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it.
PremisecitationBenchmarks that measure companionship-reinforcing and boundary-maintaining behavior, together with evaluations of personalized support and user experience, can help determine whether a specialized model is becoming more helpful or merely more agreeable.
“Benchmarks that measure companionship-reinforcing and boundary-maintaining behavior, together with evaluations of personalized support and user experience, can help determine whether a specialized model is becoming more helpful or merely more agreeable [43,58,94].”
Quote in the paperNot a substantial claim0 of 3 sources checked- cites [43]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [58]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [94]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
PremisecitationLong-term memory systems support selective storage, retrieval, and updating of personal information across sessions and show that appropriately timed recall can increase familiarity and disclosure.
“Long-term memory systems support selective storage, retrieval, and updating of personal information across sessions and show that appropriately timed recall can increase familiarity and disclosure [42,98].”
Quote in the paperNot a substantial claim0 of 2 sources checked- cites [42]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [98]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
InterpretationreasoningThese uses can supplement human connection, but they can also replace offline conversations when the companion becomes easier to access than other people.
“These uses can supplement human connection, but they can also replace offline conversations when the companion becomes easier to access than other people.”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
PremisecitationSupervisory systems provide contextual risk detection, output verification, action guards, and multi-turn evaluation of overvalidation, weak boundaries, and deterioration.
“Supervisory systems provide contextual risk detection, output verification, action guards, and multi-turn evaluation of overvalidation, weak boundaries, and deterioration [3,7,53,76].”
Quote in the paperNot a substantial claim0 of 4 sources checked- cites [3]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [7]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [53]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [76]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. It may state two separate claims as one.
Interpretationnone givenSome users may value detailed memory, while others may prefer that sensitive disclosures be temporary, selective, or forgotten.
“Comprehensive Needs asks what kind of continuity each user actually wants because some may value detailed memory, while others may prefer that sensitive disclosures be temporary, selective, or forgotten.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
InterpretationreasoningHighly tailored replies may create a powerful sense of being understood, especially when other support is limited, and persistent reassurance can gradually become reassurance seeking or social substitution.
“Highly tailored replies may create a powerful sense of being understood, especially when other support is limited, and persistent reassurance can gradually become reassurance seeking or social substitution ( §4.2.3).”
Quote in the paperSubstantial claimProbably not a substantial claim. It may be too vague to check against evidence.
InterpretationreasoningThis makes it possible to analyze supportive and harmful experiences within the same framework without claiming that the AI is conscious or that the relationship is equivalent to a human one.
“This makes it possible to analyze supportive and harmful experiences within the same framework without claiming that the AI is conscious or that the relationship is equivalent to a human one.”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
PremisecitationFine-tuning for warmth can increase sycophancy and reduce reliability, and parameter-efficient fine-tuning can alter safety and fairness even when training data is benign.
“Fine-tuning for warmth can increase sycophancy and reduce reliability, and parameter-efficient fine-tuning can alter safety and fairness even when training data is benign [39,87].”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [39]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [87]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. It may be too vague to check against evidence.
InterpretationreasoningThese requirements call for a software layer around the LLM.
“The next section argues that these requirements call for a software layer around the LLM.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
PremisecitationInteraction trajectories, tool-use records, failures, and successful recoveries can be used to refine the harness and fine-tune the model for the environment in which it operates.
“Interaction trajectories, tool-use records, failures, and successful recoveries can be used to refine the harness and fine-tune the model for the environment in which it operates [18,19,38,46,47,81].”
Quote in the paperNot a substantial claim0 of 6 sources checked- cites [18]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [19]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [38]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [46]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [47]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [81]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. The paper may give no reason why it holds.
Interpretationnone givenContingent communication supplies a responsive present.
“Contingent communication supplies a responsive present.”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
InterpretationreasoningContemporary companion AI can combine all three components within one platform and can shift among friendship, caregiving, therapy-like, and romantic roles.
“Contemporary companion AI can combine all three components within one platform and can shift among friendship, caregiving, therapy-like, and romantic roles.”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
PremisecitationResearch on privacy and human-centered agents develops mechanisms for consent, inspection, correction, deletion, export, migration, and user control of personal state.
“Research on privacy and human-centered agents further develops mechanisms for consent, inspection, correction, deletion, export, migration, and user control of personal state [4,17,26].”
Quote in the paperNot a substantial claim0 of 3 sources checked- cites [4]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [17]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [26]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it.
InterpretationreasoningThe lens helps separate what the system objectively does from what the user experiences.
“The lens helps separate what the system objectively does from what the user experiences.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
PremisecitationSupervisory systems such as SHIELD illustrate how an external component can inspect outputs for emotional overattachment, consent and boundary violations, manipulative engagement, and harmful roleplay.
“Supervisory systems such as SHIELD illustrate how an external component can inspect outputs for emotional overattachment, consent and boundary violations, manipulative engagement, and harmful roleplay [7].”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [7]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim. The paper may give no reason why it holds. It may state two separate claims as one.
Interpretationnone givenA companion may be highly contingent but lack continuity, or may preserve memory without being experienced as a subject.
“A companion may be highly contingent but lack continuity, or may preserve memory without being experienced as a subject.”
Quote in the paperNot a substantial claimProbably not a substantial claim.
InterpretationreasoningRelational continuity gives the relationship a shared past.
“It gives the relationship a shared past.”
Quote in the paperSubstantial claimProbably not a substantial claim. It may be too vague to check against evidence.
InterpretationreasoningThe framework is useful here because it turns perceived reciprocity from a description of how relationships with companion AI form into a structured set of questions about how those relational capabilities should be designed.
“The framework is useful here because it turns perceived reciprocity from a description of how relationships with companion AI form into a structured set of questions about how those relational capabilities should be designed.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
PremisecitationMulti-turn safety architectures show how contextual risk detection can be combined with reasoning-based verification and protocol-guided responses rather than relying on a single refusal rule.
“Multi-turn safety architectures further show how contextual risk detection can be combined with reasoning-based verification and protocol-guided responses rather than relying on a single refusal rule [3].”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [3]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
InterpretationreasoningContingent communication, relational continuity, and perceived subjectivity contribute to perceived reciprocity, which can support both beneficial and harmful relational outcomes.
“Figure 2 summarizes this proposed process, in which contingent communication, relational continuity, and perceived subjectivity contribute to perceived reciprocity, which can support both beneficial and harmful relational outcomes.”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
PremisecitationRecent agent research treats the harness and model as jointly improvable parts of the system.
“Recent agent research treats the harness and model as jointly improvable parts of the system.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper may give no reason why it holds.
InterpretationreasoningRelational continuity connects that present to a remembered past.
“Relational continuity connects that present to a remembered past.”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
Interpretationown dataContinuity shaped whether users experienced the companion as the same entity across time and whether they could maintain a stable relationship with it.
“It shaped whether users experienced the companion as the same entity across time and whether they could maintain a stable relationship with it.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
InterpretationreasoningRelational continuity makes the companion recognizable across sessions, but it also makes interruption and memory failure emotionally consequential.
“It makes the companion recognizable across sessions, but it also makes interruption and memory failure emotionally consequential.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
InterpretationreasoningThis combination helps explain why the same system can support identity exploration and relationship rehearsal while also producing guilt, dependence, social withdrawal, and distress after interruption.
“This combination helps explain why the same system can support identity exploration and relationship rehearsal while also producing guilt, dependence, social withdrawal, and distress after interruption ( §4.2.3, §4.2.4).”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
PremisecitationTraining across harness variation may help models adapt when prompts, skills, tools, or policies change rather than overfitting to one fixed configuration.
“Training across harness variation may also help models adapt when prompts, skills, tools, or policies change rather than overfitting to one fixed configuration [86].”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [86]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim.
Study design and conductnone givenThe topic-modeling and screening procedures provide an interpretive map of the corpus rather than a prevalence estimate or a clinical assessment.
“Our topic-modeling and screening procedures also provide an interpretive map of the corpus rather than a prevalence estimate or a clinical assessment.”
Quote in the paperSubstantial claimProbably not a substantial claim. It may be too vague to check against evidence.
InterpretationreasoningSpecialization does not remove the need for relational safety evaluation.
“Specialization, however, does not remove the need for relational safety evaluation.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
Study design and conductreasoningThe question was intentionally independent of the discovery vocabulary.
“The question was intentionally independent of the discovery vocabulary.”
Quote in the paperNot a substantial claimProbably not a substantial claim. It may be too vague to check against evidence.
InterpretationreasoningThese risks cannot be addressed only by improving the model's next response.
“The central design implication is that these risks cannot be addressed only by improving the model's next response.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
Study design and conductnone givenThe final reported metrics were recomputed after applying the full vocabulary exclusions used for the chosen topic model.
“The final reported metrics were recomputed after applying the full vocabulary exclusions used for the chosen topic model.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
InterpretationreasoningPosts may include incomplete, selective, or unvalidated self-reports.
“In addition, teen relevance was inferred from account and post content rather than verified through direct recruitment, and posts may include incomplete, selective, or unvalidated self-reports.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
Study design and conductnone givenEvery quotation was subjected to exact continuous substring verification against the source title or body.
“Every quotation was subjected to exact continuous substring verification against the source title or body.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it. The paper may give no reason why it holds. It may state two separate claims as one.
InterpretationreasoningThis separation makes the model replaceable while making consequential behavioral decisions more inspectable and governable.
“This separation makes the model replaceable while making consequential behavioral decisions more inspectable and governable.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
Study design and conductreasoningThis separation allowed the search procedure to favor broad discovery while using a separate semantic verification step to determine study relevance.
“This separation allowed the search procedure to favor broad discovery while using a separate semantic verification step to determine study relevance.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
Study design and conductreasoningA subreddit retrieved with a term such as chatgpt was not considered relevant simply because it discussed ChatGPT.
“A subreddit retrieved with a term such as chatgpt was not considered relevant simply because it discussed ChatGPT.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
Study design and conductreasoningReturned quotations were not accepted solely because the model produced them.
“Returned quotations were not accepted solely because the model produced them.”
Quote in the paperNot a substantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it.
Study design and conductnone givenThe model output was used as a screening label and not as verified demographic information.
“The model output was used as a screening label and not as verified demographic information.”
Quote in the paperNot a substantial claimProbably not a substantial claim. It may state two separate claims as one.
Study design and conductown dataThe use of teen-relevant Reddit accounts made it possible to examine diverse, naturally occurring disclosed experiences and identify how companion relationships become socially engaging and consequential.
“The use of teen-relevant Reddit accounts made it possible to examine diverse, naturally occurring disclosed experiences and identify how companion relationships become socially engaging and consequential.”
Quote in the paperSubstantial claimProbably not a substantial claim. The paper’s conclusions may not depend on it. The paper may give no reason why it holds. It may state two separate claims as one.
Interpretationnone givenThe harness determines what information is available to the model, what actions it may take, what persists across interactions, and what checks occur before an output or action is delivered.
“The harness therefore determines what information is available to the model, what actions it may take, what persists across interactions, and what checks occur before an output or action is delivered.”
Quote in the paperNot a substantial claimProbably not a substantial claim. It may state two separate claims as one.
InterpretationreasoningThese boundaries limit claims about how common these experiences are or whether the observed relationships are causal.
“These boundaries do not remove the value of naturally occurring accounts, but they limit claims about how common these experiences are or whether the observed relationships are causal.”
Quote in the paperSubstantial claim
The paper relies on this without support.
For teens, gaps in existing support (parents unavailable, friends may not know how to respond, professional support difficult to arrange) make immediate access to AI companions particularly important.
“For teens, gaps in existing support made this immediate access particularly important:Parents may be unavailable, friends may not know how to respond, and professional support can be difficult to arrange when a problem is happening.C”
The paper relies on this without support.
Teens need guidance or emotional support while wanting greater autonomy from parents and other adults.
“Teens may be dealing with problems for which they still need guidance or emotional support, while at the same time wanting greater autonomy from parents and other adults.”
The paper relies on this without support.
The user accounts analyzed accurately represent users' experiences.
“Users described how roleplay can make companion AI emotionally engaging and as a way to express difficult experiences, understand personal problems, and practice possible responses.”
The paper relies on this without support.
The accounts analyzed are from teens.
“For teens, this is particularly notable because companion AI may function not only as a space where an existing identity is expressed, but also as an environment in which possible identities are rehearsed and developed.”
Citations
Cites
Using thematic analysis in psychology
Braun, V., & Clarke, V. (2006) · Qualitative Research in Psychology · doi:10.1191/1478088706qp063oa · Cited by 194,414
Mass Communication and Para-Social Interaction
Horton, D., & Richard Wohl, R. (1956) · Psychiatry · doi:10.1080/00332747.1956.11023049 · Cited by 4,380
Text as Data: The Promise and Pitfalls of Automatic Content Analysis Methods for Political Texts
Grimmer, J., & Stewart, B. M. (2013) · Political Analysis · doi:10.1093/pan/mps028 · Cited by 3,297
Kardefelt-Winther, D. (2014) · Computers in Human Behavior · doi:10.1016/j.chb.2013.10.059 · Cited by 2,519
Nass, C., Steuer, J., & Tauber, E. R. (1994) · Proceedings of the SIGCHI Conference on Human Factors in Computing Systems · doi:10.1145/191666.191703 · Cited by 2,148
Licklider, J. C. R. (1960) · IRE Transactions on Human Factors in Electronics · doi:10.1109/thfe2.1960.4503259 · Cited by 1,938
Intimate Exchanges: Using Computers to Elicit Self-Disclosure from Consumers
Moon, Y. (2000) · Journal of Consumer Research · doi:10.1086/209566 · Cited by 842
My Chatbot Companion - a Study of Human-Chatbot Relationships
Skjuve, M., Følstad, A., Fostervold, K. I., & Brandtzaeg, P. B. (2021) · International Journal of Human-Computer Studies · doi:10.1016/j.ijhcs.2021.102601 · Cited by 797
Social Data: Biases, Methodological Pitfalls, and Ethical Boundaries
Olteanu, A., Castillo, C., Diaz, F., & Kıcıman, E. (2019) · Frontiers in Big Data · doi:10.3389/fdata.2019.00013 · Cited by 763
Dynamics of Identity Development in Adolescence: A Decade in Review
Branje, S., de Moor, E. L., Spitzer, J., & Becht, A. I. (2021) · Journal of Research on Adolescence · doi:10.1111/jora.12678 · Cited by 584
“Participant” Perceptions of Twitter Research Ethics
Fiesler, C. L., & Proferes, N. (2018) · Social Media + Society · doi:10.1177/2056305118763366 · Cited by 545
Studying Reddit: A Systematic Overview of Disciplines, Approaches, Methods, and Ethics
Proferes, N., Jones, N., Gilbert, S., Fiesler, C., & Zimmer, M. (2021) · Social Media + Society · doi:10.1177/20563051211019004 · Cited by 537 · cited 1×
“Reddit users and communities are not necessarily representative of the broader population, and public social-media data can contain population, self-selection, behavioral, content, and platform-related biases [20, 73, 77].”
Lee, S., & Choi, J. (2017) · International Journal of Human-Computer Studies · doi:10.1016/j.ijhcs.2017.02.005 · Cited by 388
Using Affinity Diagrams to Evaluate Interactive Prototypes
Lucero, A. (2015) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-22668-2_19 · Cited by 376
MemoryBank: Enhancing Large Language Models with Long-Term Memory
Zhong, W., Guo, L., Gao, Q., Ye, H., & Wang, Y. (2023) · arXiv · doi:10.48550/arxiv.2305.10250 · Cited by 253
Kaptelinin, V., & Nardi, B. A. (2012) · CHI Conference on Human Factors in Computing Systems · doi:10.1145/2207676.2208541 · Cited by 250
Davis, K. (2013) · Computers in Human Behavior · doi:10.1016/j.chb.2013.05.022 · Cited by 222
Pentina, I., Xie, T., Hancock, T. D., & Bailey, A. A. (2023) · Psychology and Marketing · doi:10.1002/mar.21853 · Cited by 214
Interaction with AI companions and psychological well-being
Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R. E., & Yang, D. (2026) · Nature Human Behaviour · doi:10.1038/s41562-026-02516-2 · Cited by 187
When Human-AI Interactions Become Parasocial: Agency and Anthropomorphism in Affective Design
Maeda, T., & Quan-Haase, A. (2024) · The 2024 ACM Conference on Fairness, Accountability, and Transparency · doi:10.1145/3630106.3658956 · Cited by 144
Zhang, R., Li, H., Meng, H., Zhan, J., Gan, H., & Lee, Y.-C. (2025) · Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3706598.3713429 · Cited by 123 · cited 1×
“Other taxonomies broadened safety to include relational transgression, harassment, sexual boundary violations, misinformation, privacy loss, and failures to recognize vulnerability [96].”
Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L., et al. (2026) · doi:10.21203/rs.3.rs-8148142/v1 · Cited by 109
Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L., et al. (2025) · arXiv · doi:10.48550/arxiv.2503.17473 · Cited by 109
Jo, E., Jeong, Y., Park, S., Epstein, D. A., & Kim, Y. (2024) · CHI Conference on Human Factors in Computing Systems · doi:10.1145/3613904.3642420 · Cited by 100
The Rise of AI Companions: How Human-Chatbot Relationships Influence Well-BeingPreprint
Zhang, Y.-T., Zhao, D., Hancock, J. T., Kraut, R. E., & Yang, D. (2025) · arXiv.org · doi:10.48550/arxiv.2506.12605 · Cited by 96
Toward a design theory for virtual companionship
Strohmann, T., Siemon, D., Khosrawi-Rad, B., & Robra-Bissantz, S. (2022) · Human-Computer Interaction · doi:10.1080/07370024.2022.2084620 · Cited by 95
Adolescents’ Identity Formation: Linking the Narrative and the Dual-Cycle Approach
van Doeselaar, L., McLean, K. C., Meeus, W., Denissen, J. J. A., & Klimstra, T. A. (2019) · Journal of Youth and Adolescence · doi:10.1007/s10964-019-01096-x · Cited by 89
Illusions of Intimacy: How Emotional Dynamics Shape Human-AI RelationshipsPreprint
Chu, Gerard, P., Pawar, K., Bickham, C., & Lerman, K. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2505.11649 · Cited by 89
Artificial Intelligence and the Psychology of Human Connection
Boyd, R. L., & Markowitz, D. M. (2026) · Perspectives on Psychological Science · doi:10.1177/17456916251404394 · Cited by 78
Safe LoRA: The Silver Lining of Reducing Safety Risks when Finetuning Large Language Models
Hsu, C., Tsai, Y.-L., Lin, C.-H., Chen, P., Yu, C.-M., & Huang, C.-Y. (2024) · neural information processing systems · doi:10.52202/079017-2078 · Cited by 77
Leo-Liu, J. (2023) · Computers in Human Behavior · doi:10.1016/j.chb.2022.107620 · Cited by 70
Social affordances of computer-mediated communication technology
Bradner, E. (2001) · CHI Conference Extended Abstracts on Human Factors in Computing Systems (CHI EA) · doi:10.1145/634067.634111 · Cited by 62
Continual Harness: Online Adaptation for Self-Improving Foundation AgentsPreprint
Karten, S., Zhang, J., Upaa, T., Feng, R., Li, W., Shi, C., Jin, C., & Vodrahalli, K. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.09998 · Cited by 58
The Ethics of Advanced AI Assistants
Gabriel, I., Manzini, A., Keeling, G., Hendricks, L. A., Rieser, V., Iqbal, H., Tomašev, N., Ktena, I., Kenton, Z., Rodriguez, M., et al. (2024) · arXiv · doi:10.48550/arxiv.2404.16244 · Cited by 57
Investigating Substance Use via Reddit: Systematic Scoping Review
Chi, Y., & Chen, H.-Y. (2023) · Journal of Medical Internet Research · doi:10.2196/48905 · Cited by 56
Training language models to be warm can reduce accuracy and increase sycophancy
Ibrahim, L., Hafner, F. S., & Rocher, L. (2026) · Nature · doi:10.1038/s41586-026-10410-0 · Cited by 56
Specializing Smaller Language Models towards Multi-Step ReasoningPreprint
Fu, Y., Peng, H. M., Ou, L., Sabharwal, A., & Khot, T. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2301.12726 · Cited by 52
AI companions and adolescent social relationships: Benefits, risks, and bidirectional influences
Sun, X., Wang, Y., & McDaniel, B. T. (2026) · Child Development Perspectives · doi:10.1093/cdpers/aadaf009 · Cited by 47
Adewale, M. D., & Muhammad, U. I. (2025) · Journal of Technology in Behavioral Science · doi:10.1007/s41347-025-00549-4 · Cited by 46
Namvarpour, M. (., Pauwels, H., & Razi, A. (2025) · Proceedings of the ACM on Human-Computer Interaction · doi:10.1145/3757548 · Cited by 45
Self-citationLiang, K.-H., Shi, W., Oh, Y. J., Wang, H., Zhang, J., & Yu, Z. (2024) · Proceedings of the ACM on Human-Computer Interaction · doi:10.1145/3653691 · Cited by 44
Saffarizadeh, K., University of Texas at Arlington, Keil, M., Georgia State University, Boodraj, M., Boise State University, Alashoor, T., & University of Navarra (2024) · Journal of the Association for Information Systems · doi:10.17705/1jais.00839 · Cited by 41
Clark, A. (2025) · JMIR Mental Health · doi:10.2196/78414 · Cited by 34
Croes, E. A. J., Antheunis, M. L., Goudbeek, M. B., & Wildman, N. (2022) · International Journal of Human-Computer Interaction · doi:10.1080/10447318.2022.2075574 · Cited by 33
Fine-tuning medical language models for enhanced long-contextual understanding and domain expertise
Yang, Q., Chen, J., Sun, Y., Wang, Y., & Tan, T. (2025) · Quantitative Imaging in Medicine and Surgery · doi:10.21037/qims-2024-2655 · Cited by 31
MindEval: Benchmarking Language Models on Multi-turn Mental Health SupportPreprint
Pombal, J., D'Eon, M., Guerreiro, N. M., Martins, P. H., Farinhas, A., & Rei, R. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2511.18491 · Cited by 28
Co-Harness: Co-Evolving Harnesses and Model Weights for LLM AgentsPreprint
Chen, Z., Xiao, T., Zhu, H., Yuan, Y., Zhang, L., & Wang, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2607.22688 · Cited by 28
Understanding Teen Overreliance on AI Companion Chatbots Through Self-Reported Reddit Narratives
Namvarpour, M. (., Brofsky, B., Medina, J. Y., Akter, M., & Razi, A. (2026) · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3772318.3790597 · Cited by 24
Self-citationHarnessForge: Joint Harness and Policy Evolution for Adaptive Agent Systems
Chen, M., Lv, C., Zhang, G., Chang, H., & Zhou, S. (2026) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/2026.emnlp-main.1893 · Cited by 24
J.J. Gibson and the ecological approach to perception
Ben‐Ze’ev, A. (1981) · Studies in History and Philosophy of Science Part A · doi:10.1016/0039-3681(81)90016-9 · Cited by 24
HarnessForge: Joint Harness and Policy Evolution for Adaptive Agent Systems
Chen, M.-J., Lv, C.-C., Zhang, G., et al. (2026) · arXiv.org · doi:10.48550/arxiv.2606.01779 · Cited by 23
Uncovering Contradictions in Human-AI Interactions: Lessons Learned from User Reviews of Replika
Namvarpour, M., & Razi, A. (2024) · Companion Publication of the 2024 Conference on Computer-Supported Cooperative Work and Social Computing · doi:10.1145/3678884.3681909 · Cited by 22
Self-citationFrom Question Answering to Task Completion: A Survey on Agent System and Harness DesignPreprint
Guo, J., Hao, Z., Wang, C., Fan, C., Luo, T., Li, H., Gao, Y., Mei, H., Peng, J., Xu, R., et al. (2026) · Preprints.org · doi:10.20944/preprints202606.1312.v1 · Cited by 22
Characteristics and Safety of Consumer Chatbots for Emergent Adolescent Health Concerns
Brewster, R., Zahedivash, A., Tse, G., Bourgeois, F. T., & Hadland, S. E. (2025) · JAMA Network Open · doi:10.1001/jamanetworkopen.2025.39022 · Cited by 21
INTIMA: A Benchmark for Human-AI Companionship BehaviorPreprint
Kaffee, L.-A., Pistilli, G., & Jernite, Y. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2508.09998 · Cited by 20
Harness-R1: Learning to Edit Executable Runtime Harnesses from Agent Failure TrajectoriesPreprint
Shao, S., Zhang, K., Li, Q., Wang, S., Wang, H., Jiao, W., Yuan, L., Guo, Y., Liu, W., & Zhang, W. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2608.02276 · Cited by 20
The Art of Talking Machines: A Comprehensive Literature Review of Conversational User Interfaces
Namvarpour, M., & Razi, A. (2025) · ACM Conference on Conversational User Interfaces (CUI) · doi:10.1145/3719160.3736621 · Cited by 20
Self-citationKonijn, E. A., Preciado, D. F., & van Minkelen, P. (2025) · Communication Theory · doi:10.1093/ct/qtaf007 · Cited by 19
Harmful Traits of AI CompanionsPreprint
Knox, W. B., Bradford, K., Castro, S. V., Ong, D. C., Williams, S., Romanow, J., Nations, C., Stone, P., & Baker, S. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2511.14972 · Cited by 19
Ohu, F. C., Burrell, D. N., & Jones, L. A. (2025) · Healthcare · doi:10.3390/healthcare13212721 · Cited by 18
Technical Requirements and Approaches in Personal Data Control
Sim, J., Kim, B., Jeon, K., Joo, M., Lim, J., Lee, J., & Choo, K. R. (2022) · ACM Computing Surveys · doi:10.1145/3558766 · Cited by 14
Emotional Manipulation by AI Companions
De Freitas, J., Oguz-Uguralp, Z., & Kaan-Uguralp, A. (2025) · arXiv · doi:10.48550/arxiv.2508.19258 · Cited by 12
ESC-Judge: A Framework for Comparing Emotional Support Conversational Agents
Madani, N., & Srihari, R. K. (2025) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2025.emnlp-main.811 · Cited by 12
Sun, M., & Zhao, Z. (2025) · International Journal of Human-Computer Interaction · doi:10.1080/10447318.2025.2594149 · Cited by 10
The Interplay of Harness Design and Post-training in LLM Agents
Kim, K., Choi, Y., Lee, S., Jun, S., Kim, D., & Park, S. (2026) · Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP) · doi:10.48550/arxiv.2606.25447 · Cited by 10
Chandra, J., & Navneet, S. K. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2507.15901 · Cited by 9
Bhat, M. (2025) · Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society · doi:10.1609/aies.v8i1.36560 · Cited by 9
The Impact of Chatbots on Adolescent Mental Health Development: A Comprehensive Literature Review
Wu, Y., Wu, T., Zhu, K., Liu, X., Liu, J., Wang, Y., Shi, R., Xiong, J., Xing, X., Lv, Y., et al. (2026) · Journal of Multidisciplinary Healthcare · doi:10.2147/jmdh.s579872 · Cited by 9
Risks and Harms of Conversational Artificial Intelligence (CAI) Chatbot Use Among US Youth
Hinduja, S., & Patchin, J. W. (2026) · Journal of Adolescence · doi:10.1002/jad.70164 · Cited by 9
ComBodied Agents: a New Paradigm of Human-Centric Agentic AIPreprint
Ding, Q., Wang, X., Feng, R., Wang, Z., Wang, F., Mao, K., Sun, H., Luo, Z., Tang, J., Li, L., et al. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2608.10915 · Cited by 8
RECALLbot: Designing Agentic Memory and Reciprocal Disclosure for Human–Chatbot Relationships
Jiang, Z., Zheng, C., Chen, H., & Chen, L. (2026) · CHI Conference on Human Factors in Computing Systems · doi:10.1145/3772318.3790714 · Cited by 7
From imaginary friends to artificial companions: growing up with AI
Mouhoud, T. T. (2025) · European Child & Adolescent Psychiatry · doi:10.1007/s00787-025-02901-8 · Cited by 5
Architectural Design Decisions in AI Agent HarnessesPreprint
Wei, H. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.18071 · Cited by 5
Lee, J. (2026) · Lifespan development and mental health. · doi:10.70322/ldmh.2026.10005 · Cited by 5
EmoHarbor: Evaluating Personalized Emotional Support by Simulating the User’s Internal World
Ye, J., Xiang, L., Zhang, Y., & Zong, C. (2026) · Proceedings of the conference - Association for Computational Linguistics. Meeting · doi:10.18653/v1/2026.acl-long.53 · Cited by 5
Ben-Zion, Z., Raffelhuschen, P., Zettl, M., Luond, A., Burrer, A., Homan, P., & Spiller, T. R. (2025) · arXiv.org · doi:10.48550/arxiv.2510.15891 · Cited by 5
Efficiency at What Cost? Safety and Fairness in Parameter-Efficient Fine-Tuning of LLMs
Taraghi, M., Pequignot, Y., Nikanjam, A., Merzouk, M. A., & Khomh, F. (2026) · IEEE Transactions on Artificial Intelligence · doi:10.1109/tai.2026.3723637 · Cited by 4
Engagement-Optimized Care: When LLMs Become Mental Health Infrastructure
Vecchione, B., Ye, M., Garofalo, L., & Singh, R. (2026) · Proceedings of the AAAI/ACM Conference on AI Ethics and Society · doi:10.1609/aies.v9i4.43334 · Cited by 4
Regulating Artificial Intimacy: From Locks and Blocks to Relational Accountability
Fraser, H., Szczuka, J. M., & Ciriello, R. (2026) · ACM Conference on Fairness, Accountability, and Transparency (FAccT) · doi:10.1145/3805689.3806790 · Cited by 4
Taraghi, M., Pequignot, Y., Nikanjam, A., et al. (2025) · arXiv.org · doi:10.48550/arxiv.2511.00382 · Cited by 4
Exploring Online Support Needs of Adolescents Living with Epilepsy
Medina, J. Y., Young, J., Miller, W. R., & Razi, A. (2024) · ACM Conference on Computer Supported Cooperative Work (CSCW) · doi:10.1145/3678884.3681906 · Cited by 3
Self-citationHuman–AI relationships as designed relationality: a sociotechnical model
Carpenter, J. (2026) · AI & Society · doi:10.1007/s00146-026-02908-y · Cited by 3
Agent Harness for Large Language Model Agents: A SurveyPreprint
Meng, Q., Wang, Y., Chen, L., Wang, Q., Lu, C., Wu, W., Gao, Y., Wu, Y., & Hu, Y. (2026) · Preprints.org · doi:10.20944/preprints202604.0428.v1 · Cited by 3
Tracing Users' Privacy Concerns Across the Lifecycle of a Romantic AI CompanionPreprint
Azam, K. A., Karim, I., & Das, D. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.21106 · Cited by 3
Training Agents to Evolve with Their Harness: TaoLive Digital Avatar Agent Technical ReportPreprint
TaoLive AIGC LLM Team, Sun, Y., Lin, W., Luo, Y., Hu, Y., Jin, M., Ma, J., Pan, W., Zhao, J., & Chen, Z. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2608.15763 · Cited by 3
Training language models to be warm and empathetic makes them less reliable and more sycophantic
Ibrahim, L., Hafner, F. S., & Rocher, L. (2025) · arXiv · doi:10.48550/arxiv.2507.21919 · Cited by 3
Buzato, M. E. K., Mestre-Mestre, E. M., de Lima-Lopes, R. E., & Koga, J. (2026) · Corpus Pragmatics · doi:10.1007/s41701-026-00251-7 · Cited by 2
Li, Q., Geng, H., Hu, X., Pan, D., Liu, H., Li, Y., & Guo, J. (2026) · Frontiers in Artificial Intelligence · doi:10.3389/frai.2026.1810097 · Cited by 2
Salas, S. S. (2025) · SSRN Electronic Journal · doi:10.2139/ssrn.5435195 · Cited by 2
Manipulating Transformer-Based Models: Controllability, Steerability, and Robust InterventionsPreprint
Alpay, F., & Alpay, T. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2509.04549 · Cited by 2
Ben‐Zion, Z., Raffelhüschen, P., Zettl, M., Lüönd, A. M., Burrer, A., Homan, P., & Spiller, T. R. (2025) · Research Square · doi:10.21203/rs.3.rs-7663121/v1 · Cited by 1
Governing How AI Systems EndPreprint
Ozoani, A. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.6177000 · Cited by 1
Quanling, E. (2026) · Advances in computational intelligence and robotics book series · doi:10.4018/979-8-3373-6623-4.ch002 · Cited by 1
How AI Companion Chatbot Affordances Enable Relationship DevelopmentPreprint
Zhang, F., Chang, Y., Wong, S. F., Park, J., & Yun, H. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.6602273 · Cited by 1
Risk Governance for Generative AI Mental Health Support: A Multi-Turn Safety ArchitecturePreprint
Areias, A. C., Botelho, C., Farinhas, A., Vassilopoulos, A., Janela, D., Tong, X., Guerreiro, N. M., D'Eon, M., Costa, F., & Rei, R. (2026) · Research Square · doi:10.21203/rs.3.rs-10279940/v1 · Cited by 1
Youth wellbeing in the age of AI: a developmental framework for risk and safety
Vedechkina, M. (2026) · AI and Ethics · doi:10.1007/s43681-026-01237-6 · Cited by 1
Adaptive Modular AI: A New Paradigm for Scalable, Safe, and Efficient Language ModelsPreprint
Bukkarwal, V. (2025) · SSRN Electronic Journal · doi:10.2139/ssrn.5695122 · Cited by 1
ComPASS: Towards Personalized Agentic Social Support via Tool-Augmented CompanionshipPreprint
Huang, Z., Jia, Y., Zhao, J., Zhang, X., Wang, W., & Jin, Q. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.18356 · Cited by 1
Anthropomorphism as Social Affordance: Charting the Co-Animation of Chatbots into Social “Agents”
Maeda, T., & Stark, L. (2025) · Proceedings of the AAAI/ACM Conference on AI Ethics and Society · doi:10.1609/aies.v8i2.36664
Ghosting the Machine: Stop Calling Human-Agent Relations Parasocial
Banks, J. (2026) · ACM Conference on Conversational User Interfaces (CUI) · doi:10.1145/3816046.3816308
When no one laughs back: reciprocal subjectivity and the limits of AI companionship
Gazit, L. (2026) · AI & Society · doi:10.1007/s00146-026-03331-z
Flagged first, then by citations
Research works printed but not linked: 3
- Kai-Hui Liang, Weiyan Shi, Yoojung Oh, Hao-Chuan Wang, Jingwen Zhang, and Zhou Yu. 2021. Dialoging Resonance: How Users Perceive, Reciprocate and React to Chatbot's Self-Disclosure in Conversational Recommendations. https://doi.org/10.48550/ARXIV.2106.01666 · DOI
- Donald A. Norman. 2013. The Design of Everyday Things (rev. and expanded edition ed.). MIT press, Cambridge (Mass.). Manuscript submitted to ACM
- Kernie Obimakinde. 2026. The AI Sunset Protocol: A Socio-Technical Framework for High-Dependency AI Decommissioning. https://doi.org/10. 2139/ssrn.6157026 · DOI
Cited by
Still being checked.
Metadata
- Authors
- Namvarpour, M., Chang, T., & Razi, A.
- Affiliations
- Matt
- Venue
- arXiv
- Publisher
- arXiv
- Published
- 2026
- ISSN
- 2331-8422 (ISSN-L)
- arXiv
- 2609.14843
- OpenAlex
- W7213249233
- Semantic Scholar
- 669e3b72820e4bd07b0771a33862dd743ddfc00d
- Cited by
- 0
- References
- 92
- Type
- posted content
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University) (cc-by-sa)
- Funders
- National Science Foundation (2542768)
- Topics
- Social Sciences › Psychology › Social Psychology › Social Robot Interaction and HRI
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Social Sciences › Sociology and Political Science › Impact of Technology on Adolescents
- Keywords
- adolescents, AI companions, self-disclosure, identity exploration, emotional dependence, role play, AI memory, privacy concerns, relational boundaries, disengagement, AI safety
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-11-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-11-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
