Me and My Bot: What Users Talk About in AI Companion Communities on Reddit

It may state two separate claims as one.

Sexual/ERP posts most commonly contained no emotion language, with similar proportions expressing positive and negative emotion and mixed emotion least common.

“Sexual/ERP posts most commonly contained no emotion language, with similar proportions expressing positive and negative emotion and mixed emotion least common”

It may state two separate claims as one.

Every emotion category was more prevalent among posts about one's own bot: positive rose from 5.36% to 32.96%, negative from 8.36% to 18.12%, and mixed from 0.80% to 15.76%.

“Every emotion category was more prevalent among posts about one's own bot, and the increases were substantial: positive rose from 5.36% to 32.96%, negative from 8.36% to 18.12%, and mixed from 0.80% to 15.76%.”

It may state two separate claims as one.

Users often expressed positive emotion about their relationships, yet much of the public discourse in these communities concerned technical problems, platform changes, and governance.

“users often expressed positive emotion about their relationships, yet much of the public discourse in these communities concerned technical problems, platform changes, and governance.”

It may state two separate claims as one.

Of the 5,504 total posts, 3,004 (54.58%) were about bots in general, whereas 2,500 (45.42%) were explicitly about the user's relationship with their bot.

“Of the 5,504 total posts in the final data set, 3,004 were on bots in general (54.58%) whereas 2,500 of the posts were explicitly about the user's relationship with their bot (45.42%), χ²(1) = 46.15,p < .001,w = .09.”

It may state two separate claims as one.

Mixed emotional valence was virtually absent from posts about bots in general but was prominent among posts about the user's own bot and among posts in which the relationship itself was the primary topic, a thirtyfold increase across the three levels.

“Mixed emotional valence, in which a post includes both positive and negative emotion about the user's own state, was virtually absent from posts about bots in general but was prominent among posts about the user's own bot and among posts in which the relationship itself was the primary topic, a thirtyfold increase across the three levels.”

It may state two separate claims as one.

For Continuity posts, when emotions were expressed they were almost equally divided between positive and negative, with relatively low levels of mixed.

“For Continuity posts, when emotions were expressed they were almost equally divided between positive and negative, with relatively low levels of mixed.”

It may state two separate claims as one.

Among the 970 posts that were relationally focused, companionship and romance each accounted for roughly 46% of discussion, and sexual content for 8%.

“Among the 970 posts that were relationally focused, companionship and romance each accounted for roughly 46% of discussion, sexual content for 8%, and emotional valence varied across these subtopics.”

It may state two separate claims as one.

Companionship and romance each account for roughly six times as many posts as sexual/ERP content.

“Companionship and romance each account for roughly six times as many posts as sexual/ERP content.”

It may state two separate claims as one.

In relationally focused posts, users principally discussed companionship and romance and relatively rarely discussed sexual/ERP issues in these public forums.

“users principally discussed companionship and romance and relatively rarely discussed sexual/ERP issues in these public forums.”

It may state two separate claims as one.

Companionship posts were more than twice as unlikely to contain no emotion language at all and correspondingly more likely to express negative or mixed valence.

“Companionship posts were more than twice as unlikely to contain no emotion language at all and correspondingly more likely to express negative or mixed valence.”

It may state two separate claims as one.

Within relationally focused posts, ambivalence was more common than negative emotion and more common than the absence of emotion language altogether, making it the second most frequent valence category after positive emotion.

“Within relationally focused posts, ambivalence was more common than negative emotion and more common than the absence of emotion language altogether, making it the second most frequent valence category after positive emotion.”

It may state two separate claims as one.

Post length and topic each predicted valence after adjusting for the other, indicating independent contributions, and the topics with the highest rates of emotional expression were not the longest ones.

“Post length and topic each predicted valence after adjusting for the other, indicating independent contributions, and the topics with the highest rates of emotional expression were not the longest ones.”

It may state two separate claims as one.

Technical and Story telling posts had elevated no emotion and suppressed mixed valence.

“Technical and Story telling posts had elevated no emotion and suppressed mixed valence.”

It may state two separate claims as one.

Romance posts were not more positive than companionship posts but were more often emotionally flat.

“Romance posts, in other words, were not more positive than companionship posts but were more often emotionally flat.”

It may state two separate claims as one.

The vast majority of posts about bots in general contained no user emotion language, whereas only about one-third of posts about one's own bot did.

“For emotional valence, the vast majority of posts about bots in general contained no user emotion language, whereas only about onethird of posts about one's own bot did.”

It may state two separate claims as one.

Companionship posts were most often positive or mixed, only rarely devoid of emotion, and more frequently negative than romance posts.

“Companionship posts were most often positive or mixed, only rarely devoid of emotion, and more frequently negative than romance posts were.”

It may state two separate claims as one.

Sexual and erotic roleplay is rarely what users write posts about, which is distinct from how often it figures in what they describe.

“sexual and erotic roleplay is rarely what users write posts about, which is distinct from how often it figures in what they describe.”

It may state two separate claims as one.

Romance posts were similarly positive but about equally likely to be mixed or to contain no emotion and were the least negative of the three subtopics.

“Romance posts were similarly positive but about equally likely to be mixed or to contain no emotion and were the least negative of the three subtopics.”

It may state two separate claims as one.

Posts about bots in general were most commonly about technical issues and rarely about relational ones, whereas posts about one's own bot were most commonly relational.

“As Table 3 indicates, posts about bots in general were most commonly about technical issues and rarely about relational ones, whereas posts about one's own bot were most commonly relational.”

It may state two separate claims as one.

Posts about the relationship itself were shorter on average than posts about story telling yet contained emotion language far more often.

“Posts about the relationship itself were shorter on average than posts about story telling yet contained emotion language far more often.”

It may state two separate claims as one.

Relational posts were commonly positive and had the lowest no-emotion percentage of any topic.

“Relational posts were commonly positive and had the lowest no-emotion percentage of any topic.”

It may state two separate claims as one.

The search terms were weighted toward relational and attachment language, so any bias in the retrieved sample should favor the impression that communities are about members' own AI partners.

“Subreddit names can create the impression that these communities are exclusively or predominantly about members' own AI partners, and because our search terms were themselves weighted toward possessive and attachment language, any bias in the retrieved sample should favor that impression.”

It may state two separate claims as one.

The departure from an even split between own-bot and general-bot posts is statistically significant but negligible in magnitude given the sample size.

“Although statistically significant, this departure from an even split is negligible in magnitude given the sample size.”

Probably not a substantial claim.

Chang, Huh-Yoo, and Razi (2026) analyzed over 3,300 self-disclosed romantic AI companion posts from 24 Reddit subreddits spanning 2017 to 2025.

“Most directly relevant is the work of Chang, Huh-Yoo, and Razi (2026) who analyzed over 3,300 self-disclosed romantic AI companion posts from 24 Reddit subreddits spanning 2017 to 2025.”

• cites Chang, Huh-Yoo, and Razi (2026) sourceNot checked yet

Probably not a substantial claim.

Chang et al. found that explicitly positive intimate relationship content constituted only 6% of the discourse.

“Change et al. found that explicitly positive intimate relationship content constituted only 6% of the discourse.”

• cites Chang et al. (2026) sourceNot checked yet

Probably not a substantial claim. It may state two separate claims as one.

Each of the three LLMs received an identical prompt for a given post, with no access to the other models' outputs and no memory of previously coded posts.

“Each of the three LLMs received an identical prompt for a given post.Specifically, each LLM was provided with the full codebook instructions followed by that post's title and body text, with no access to the other models' outputs and no memory of previously coded posts.”

It may state two separate claims as one. The quote only partly states this claim.

The paper's findings replicate the pattern documented by Chang et al. (2026) that explicitly romantic AI companion discourse on Reddit was dominated by technical issues, platform governance, and disturbing bot behavior rather than intimate relationship content, within communities organized around specific platforms rather than the topic of AI romance broadly.

“Our findings replicate that pattern within communities organized around specific platforms rather than the topic of AI romance broadly.”

• cites Chang et al. (2026) sourceNot checked yet

It may state two separate claims as one.

Differences in length work against the reported topic contrasts rather than producing them, though absolute rates of emotional expression should be read as conservative.

“Differences in length therefore work against the reported topic contrasts rather than producing them, though absolute rates of emotional expression should be read as conservative.”

Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

Institutional Review Board approval was obtained prior to data coding, and because the data were publicly available and deidentified, the review was considered exempt from full IRB review.

“Institutional Review Board (IRB) approval was obtained prior to data coding.Because the data were publicly available and deidentified, the review was considered exempt from full IRB review.”

• AssumptionThe authors state this as an assumption: The analyzed Reddit data were publicly available and deidentified.

Probably not a substantial claim.

The low figure represents how often sexual content was the organizing focus of a post, not how often it appeared at all.

“The low figure represents how often sexual content was the organizing focus of a post, not how often it appeared at all.”

Probably not a substantial claim.

The Synthetic Resonance framework proposes that the sense of connection users report emerges from structural alignment built through repeated interaction, not from the AI's sentience or reciprocal care.

“It proposes that the sense of connection users report emerges from structural alignment built through repeated interaction, not from the AI's sentience or reciprocal care.”

• cites Fabes, 2026 sourceNot checked yet

It may state two separate claims as one.

The discourse about AI companion relationships is broad and emotionally complex.

“The findings suggest that users engage with these relationships with AI bots as meaningful, and that the discourse about them is broad and emotionally complex.”

Probably not a substantial claim. It may state two separate claims as one.

Synthetic Resonance does not pathologize user experience or dismiss it as illusion.

“Importantly, Synthetic Resonance does not pathologize user experience or dismiss it as illusion.”

• cites Fabes, 2026 sourceNot checked yet

It may state two separate claims as one.

The finding that users do so commonly with positive emotion, and that sexual content is a small fraction of the discourse, challenges narratives that frame AI companionship primarily in terms of risk, harm, or erotic substitution.

“The finding that users do so commonly with positive emotion, and that sexual content is a small fraction of the discourse, challenges narratives that frame AI companionship primarily in terms of risk, harm, or erotic substitution.”

Probably not a substantial claim. It may be too vague to check against evidence.

When we ask, 'Do AI companions help or harm?', Synthetic Resonance answers, 'It depends on how we design the AI companion.'

“When we ask, “Do AI companions help or harm?”, Synthetic Resonance answers, “It depends on how we design the AI companion.””

• cites Fabes, 2026 sourceNot checked yet

Probably not a substantial claim.

Synthetic Resonance proposes that the sense of connection users report with AI companions emerges not from the AI's sentience, understanding, or reciprocal care, but from structural alignment built through repeated interaction.

“Synthetic Resonance proposes that the sense of connection users report with AI companions emerges not from the AI's sentience, understanding, or reciprocal care, but from structural alignment built through repeated interaction.”

• cites Fabes, 2026 sourceNot checked yet

It may state two separate claims as one.

The emotional stakes attach to the bot, but the volume of conversation reflects the infrastructure that supports it.

“The emotional stakes attach to the bot, but the volume of conversation reflects the infrastructure that supports it.”

Probably not a substantial claim.

Chang et al. found that the overwhelming majority of posts addressed technical issues, platform governance and moderation, and unwanted or disturbing bot behavior.

“The overwhelming majority of posts addressed technical issues, platform governance and moderation, and unwanted or disturbing bot behavior.”

• cites Chang et al. (2026) sourceNot checked yet

Probably not a substantial claim. It may state two separate claims as one.

Synthetic Resonance distinguishes the user's subjective experience of resonance, which is real and psychologically meaningful, from the mechanism producing it, which is architectural rather than relational in the interpersonal sense.

“The theory distinguishes the user's subjective experience of resonance, which is real and psychologically meaningful, from the mechanism producing it, which is architectural rather than relational in the interpersonal sense.”

• cites Fabes, 2026 sourceNot checked yet

Probably not a substantial claim. It may state two separate claims as one.

The public narration of a relationship is not simply a transparent window into private experience; it is also a social act that navigates audience, manages stigma, constructs identity, and draws boundaries.

“The public narration of a relationship is not simply a transparent window into private experience, it is also a social act that navigates audience, manages stigma, constructs identity, and draws boundaries (De Choudhury & De, 2014).”

• cites De Choudhury & De, 2014 sourceNot checked yet

Probably not a substantial claim.

Prior to full-corpus coding, the pipeline and codebook were evaluated through several rounds of testing at increasing scale.

“Prior to full-corpus coding, the pipeline and codebook were evaluated through several rounds of testing at increasing scale.”

It may state two separate claims as one.

The very low frequency of sexual and erotic roleplay posts (less than 8% of relational posts) contradicts the popular framing of AI companions as primarily sexual outlets.

“The very low frequency of sexual and erotic roleplay posts (less than 8% of relational posts) contradicts the popular framing of AI companions as primarily sexual outlets.”

It may state two separate claims as one.

The human validation sample was deliberately stratified to oversample disagreement cases, which means the 87% human-model agreement figure should not be interpreted as a population estimate of accuracy.

“The human validation sample was deliberately stratified to oversample disagreement cases, which means the 87% human-model agreement figure should not be interpreted as a population estimate of accuracy.”

It may state two separate claims as one.

The search terms were weighted toward relationship, attachment, and emotional language, which means the corpus overrepresents relational and emotional content relative to what a random sample of all posts in these communities would contain.

“The search terms were weighted toward relationship, attachment, and emotional language, which means the corpus overrepresents relational and emotional content relative to what a random sample of all posts in these communities would contain.”

Probably not a substantial claim.

Quantifying the proportion of own-relationship posts provides a conservative test of the impression that these communities are predominantly about members' own AI partners.

“Quantifying the proportion therefore provides a conservative test of it.”

Probably not a substantial claim. The paper’s conclusions may not depend on it. The paper may give no reason why it holds.

Moderate kappas relative to high raw agreement rates reflect base-rate compression that occurs when one category dominates, which elevates chance agreement and depresses kappa.

“Moderate kappas relative to high raw agreement rates reflect base-rate compression that occurs when one category dominates, which elevates chance agreement and depresses kappa.”

It may state two separate claims as one.

The user's experience of resonance is real and psychologically meaningful, and asymmetrical, even though the mechanism producing it are architectural rather than interpersonal in the conventional sense.

“The user's experience of resonance is real and psychologically meaningful, and asymmetrical, even though the mechanism producing it are architectural rather than interpersonal in the conventional sense.”

• cites Fabes, 2026 sourceNot checked yet

Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

Body text was absent for 341 posts (about 6%), which were coded from the title alone.

“body text was absent for 341 posts [about 6%] and were coded from the title alone”

It may state two separate claims as one.

OpenAI held the lowest dissent rate on every variable individually, supporting a single tie-break rule applied uniformly rather than a variable-specific one.

“Importantly, OpenAI held the lowest dissent rate on every variable individually, supporting a single tie-break rule applied uniformly rather than a variable-specific one.”

It may state two separate claims as one.

In the final validation batch of 175 posts, there was unanimous agreement on about 60% of codes, about 35% agreement for two of three LLMs, and about 5% no agreement.

“The final batch run consisted of a random sample of 175 posts (excluding posts used in prior pilot batches).There was unanimous agreement on about 60% of the codes, and there was an average of another 35% agreement for two of the three LLMs, leaving about 5% of the codes where there was no agreement at all.”

It may state two separate claims as one.

Cross-platform comparisons are not possible with these data, and findings should not be generalized beyond the specific context of public Reddit discourse.

“Cross-platform comparisons (Reddit vs. Discord vs. private communities vs. inapp behavior) are not possible with these data, and findings should not be generalized beyond the specific context of public Reddit discourse.”

It may state two separate claims as one.

The tie-break procedure for the small percentage of post codings where there was not agreement among the three LLMs (adopting OpenAI's code) was justified by that model's highest agreement with the 2-of-3 majority across all variables, but it does not guarantee accuracy in individual cases.

“The tie-break procedure for the small percentage of post codings where there was not agreement among the three LLMs (adopting OpenAI's code) was justified by that model's highest agreement with the 2-of-3 majority across all variables, but it does not guarantee accuracy in individual cases.”

Probably not a substantial claim.

Both positive and negative outcomes are possible in human-AI relationships.

“Both positive and negative outcomes are possible in human-AI relationships.”

Probably not a substantial claim. It may state two separate claims as one.

Each LLM's response was parsed and checked against automated validity criteria before acceptance: structural validity, candidate grounding for VU, and topic category validity.

“Each LLM's response was parsed and checked against several automated validity criteria before being accepted:”

Probably not a substantial claim. It may state two separate claims as one.

Every post was coded independently by three large language models drawn from three distinct model-development organizations (Anthropic, OpenAI, and DeepSeek), intended to reduce the likelihood that a systematic misreading of the coding instructions would be shared across coders.

“Every post was coded independently by three large language models (LLMs), deliberately drawn from three distinct LLM modeldevelopment organizations (Anthropic, OpenAI, and DeepSeek).This was intended to reduce the likelihood that a systematic misreading of the coding instructions would be shared across coders.”

Probably not a substantial claim. The paper may give no reason why it holds.

If Synthetic Resonance's account is correct, the meaning to the user should be visible in how they talk about their relationships with AI bots.

“If that account is correct, the meaning to the user should be visible in how they talk about their relationships with AI bots.”

It may state two separate claims as one.

Because emotional valence was coded from explicit emotion terms, the amount of text available directly affects the opportunity for such terms to appear, and mixed valence is especially sensitive to length because it requires both positive and negative terms to co-occur within a single post.

“Because emotional valence was coded from explicit emotion terms, the amount of text available directly affects the opportunity for such terms to appear, and mixed valence is especially sensitive to length because it requires both positive and negative terms to co-occur within a single post.”

It may state two separate claims as one.

Users did not reorganize their entire discursive repertoire around whether the post was about their bot or someone else's, but they tended to shift the balance between relational and technical content.

“Users did not reorganize their entire discursive repertoire around whether the post was about their bot or someone else's, but they tended to shift the balance between relational and technical content.”

Probably not a substantial claim. It may state two separate claims as one.

For most variables, the final code was determined by choosing the one that at least two-of-three LLMs agreed upon; where all three models produced different codes with no majority, the post was flagged No Agreement rather than resolved by an automated tie-breaking rule.

“For most variables, the final code was determined by choosing the one that at least two-of-three LLMs agreed upon.Where all three models produced different codes with no majority, the post was flagged No Agreement rather than resolved by an automated tie-breaking rule.”

It may state two separate claims as one.

Users who post publicly about their AI relationships are a self-selected subset of users who have them, and the selection dynamics are unknown and likely nonrandom.

“Users who post publicly about their AI relationships are a self-selected subset of users who have them, and the selection dynamics are unknown and likely nonrandom.”

Probably not a substantial claim. The paper’s conclusions may not depend on it.

Chang et al. documented significant temporal drift: over time discussions moved away from experiential intimacy toward governance concerns, technical infrastructure, and personal consequences.

“They also documented significant temporal drift: over time as discussions moved away from experiential intimacy toward governance concerns, technical infrastructure, and personal consequences.”

• cites Chang et al. (2026) sourceNot checked yet

It may state two separate claims as one.

AI companion subreddits are not just the medium through which relationships are discussed, refined, and reflected upon, but a place where users discuss AI bots as a topic itself.

“it appears that these communities are not just the medium through which the relationship is discussed, refined, and reflected upon, but a place where users discuss AI bots as a topic itself.”

It may state two separate claims as one.

Users may come for the relationship, but what they talk about is frequently the platform itself.

“Users may come for the relationship, but what they talk about is frequently the platform itself.”

Probably not a substantial claim. It may be too vague to check against evidence. It may state two separate claims as one.

These communities represent a natural laboratory in which millions of people are publicly working out what it means to form a relationship with an entity that cannot reciprocate and discovering that it means something, nonetheless.

“these communities represent a natural laboratory in which millions of people are publicly working out what it means to form a relationship with an entity that cannot reciprocate and discovering that it means something, nonetheless.”