Artificial intelligence and immersive technologies for LGBTQ+ youth: a narrative review

The systematic review by Gilbey et al. (2020) identified 38 studies of 24 digital health interventions for LGBTIQ+ young people aged 12 to 25.

“The closest precedent is the systematic review by Gilbey et al. (2020), which identified 38 studies of 24 digital health interventions for LGBTIQ+ young people aged 12 to 25.”

• cites Gilbey et al., 2020 sourceSupported

Tao and Fisher (2023) conducted a web-based survey of 571 LGBT adolescents (mean age 16.4 years), including cisgender lesbian, gay, and bisexual youth and transgender or non-binary adolescents.

“The mental health consequences of algorithmic discrimination in social media were examined empirically by Tao and Fisher (2023) in a web-based survey of 571 LGBT adolescents (mean age 16.4 years), including cisgender lesbian, gay, and bisexual youth and transgender or non-binary adolescents.”

• cites Tao and Fisher (2023) sourcePartly supported

  The Abstract Methods and Methods > Data Analysis report a web-based survey of 571 LGBT adolescents (mean age 16.4 years) with the stated identity groups. The citing sentence also says the study empirically examined the mental health consequences of algorithmic discrimination. Methods > Measures shows that it measured reported discrimination on social media and mental health symptoms, not whether algorithms caused that discrimination or its consequences. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

The review identified no review covering both AI and VR/metaverse for LGBTQ+ youth specifically.

“we identified no review covering both for LGBTQ+ youth specifically.”

• Also stated in: § 1 Introduction · plain

  “We identified no review synthesizing the available evidence across both AI and immersive technologies specifically for LGBTQ+ youth as a distinct population.”

The AI chatbot and VR/metaverse literatures have developed separately.

“AI-powered chatbots and virtual reality (VR) and metaverse environments have both been proposed as ways of reaching this population, but the two literatures have developed separately, and we identified no review covering both for LGBTQ+ youth specifically.”

Humphries et al. (2026) piloted a multi-agent chatbot built on GPT-4-turbo to deliver introductory problem-solving support to 44 men who have sex with men and transgender women aged 18-29 in KwaZulu-Natal, South Africa, all with mild to moderate depressive symptoms.

“Humphries et al. (2026) piloted a multi-agent chatbot built on GPT-4-turbo to deliver introductory problemsolving support to 44 men who have sex with men and transgender women aged 18-29 in KwaZulu-Natal, South Africa, all with mild to moderate depressive symptoms.”

• cites Humphries et al., 2026 sourceNot checked yet

Several participants in Ma et al. (2024) described chatbots as providing a kind of unconditional acceptance that felt difficult to find elsewhere.

“Several described chatbots as providing a kind of unconditional acceptance that felt difficult to find elsewhere.”

• cites Ma et al., 2024Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

No study has yet compared co-designed tools with conventionally developed ones.

“In co-design studies, LGBTQ+ young people identified gaps that developers had missed, although no study has yet compared co-designed tools with conventionally developed ones.”

In a survey of 675 students at two US universities, Liu et al. (2026) found that LGBQ+ and non-binary students were overrepresented among those who had never used generative AI.

“In a survey of 675 students at two US universities, Liu et al. (2026) found that LGBQ+ and non-binary students were overrepresented among those who had never used generative AI, and sexual orientation did not predict using AI for mental health once symptom severity and general AI use were taken into account.”

• cites Liu et al., 2026 sourceNot checked — no full text of the source

Liem et al. (2026) documented a participatory co-design process in which 15 LGBTQ+ teenagers worked with Planned Parenthood and academic researchers to improve Roo, an AI-powered sexual health chatbot.

“Liem et al. (2026) documented a participatory co-design process in which 15 LGBTQ+ teenagers worked with Planned Parenthood and academic researchers to improve Roo, an AI-powered sexual health chatbot.”

• cites Liem et al., 2026 sourceNot checked yet

Biernesser et al. (2023) developed Flourish, a text-message intervention for LGBTQ+ adolescents aged 13-18 with a history of suicidal thoughts or behaviors who had experienced online victimization.

“Biernesser et al. (2023) took a similar route with Flourish, a text-message intervention for LGBTQ+ adolescents aged 13-18 with a history of suicidal thoughts or behaviors who had experienced online victimization.”

• cites Biernesser et al., 2023 sourceCheck pending

There is early evidence of feasibility of VR/metaverse for transgender and gender expansive youth.

“VR and metaverse environments support avatar-based identity exploration and embodied community connection, with early evidence of feasibility for transgender and gender expansive youth.”

• Also stated in: § 2.3 Study selection and data extraction · plain

  “VR support groups feasible and acceptable for TGE youth”

• cites Meade et al., 2025 sourceNot checked yet

Pillis et al. (2024) developed the LGBTQIA+ Advocacy Simulator, in which users practice coming-out conversations and advocacy scenarios with AI-generated virtual characters.

“Pillis et al. (2024) developed the LGBTQIA+ Advocacy Simulator, in which users practice coming-out conversations and advocacy scenarios with AI-generated virtual characters.”

• cites Pillis et al., 2024 sourceNot checked yet

The Biernesser et al. (2023) team took deliberate steps to avoid outing participants to their parents.

“The team also took deliberate steps to avoid outing participants to their parents, a point we return to in Section 5.5.”

• cites Biernesser et al., 2023 sourceNot checked yet

In an experimental study with 307 participants, the third-person condition significantly improved self-reported confidence, empathy, and sympathy in advocacy scenarios (p = 0.043).

“In an experimental study with 307 participants, the third-person condition significantly improved self-reported confidence, empathy, and sympathy in advocacy scenarios (p = 0.043).”

• cites Pillis et al., 2024 sourceCheck pending

In an exploratory, non-preregistered subgroup analysis, LGBTQ identification predicted greater reductions in distress at one and three months among users with clinically elevated baseline distress in the Alongside program.

“In an exploratory, non-preregistered subgroup analysis, LGBTQ identification predicted greater reductions in distress at one and three months among users with clinically elevated baseline distress.”

• cites Cohen et al., 2025 sourceNot checked yet

Responses were particularly inadequate for youth from conservative or non-Western backgrounds, where the social and cultural stakes of LGBTQ+ identity are substantially different.

“Responses were particularly inadequate for youth from conservative or non-Western backgrounds, where the social and cultural stakes of LGBTQ+ identity are substantially different.”

• cites Lissak et al., 2024 sourceNot checked yet

Validated instruments for measuring affirmation as well as symptom reduction do not yet exist.

“outcome measures that capture affirmation as well as symptom reduction, for which validated instruments do not yet exist.”

Sexual orientation did not predict using AI for mental health once symptom severity and general AI use were taken into account.

“sexual orientation did not predict using AI for mental health once symptom severity and general AI use were taken into account.”

• cites Liu et al., 2026 sourceNot checked — no full text of the source

Lissak et al. (2024) developed a novel evaluation framework specifically for assessing large language models as emotional supporters for queer youth.

“The adequacy of AI chatbot responses for LGBTQ+ youth was explored more directly by Lissak et al. (2024), who developed a novel evaluation framework specifically for assessing large language models as emotional supporters for queer youth.”

• cites Lissak et al., 2024 sourceCheck pending

The resulting prototype in Biernesser et al. (2023) was rated as highly usable.

“the resulting prototype, delivered through a conversational character, was rated as highly usable.”

• cites Biernesser et al., 2023 sourceNot checked yet

In co-design studies, LGBTQ+ young people identified gaps that developers had missed.

“In co-design studies, LGBTQ+ young people identified gaps that developers had missed, although no study has yet compared co-designed tools with conventionally developed ones.”

• Also stated in: § 3.3 Designing chatbots with and for LGBTQ+ youth · plain

  “The teenagers identified significant gaps in the chatbot's inclusivity, insufficient transgender-specific content, a tone they found clinical and distancing, and a lack of options for diverse relationship configurations, and their feedback was directly incorporated into the platform's development.”

• cites Liem et al., 2026 sourceCheck pending

Lissak et al. (2024) found that current LLMs tended toward generic, impersonal replies that lacked the depth and contextual sensitivity that queer youth often need.

“they tended toward generic, impersonal replies that lacked the depth and contextual sensitivity that queer youth often need.”

• cites Lissak et al., 2024 sourceNot checked yet

Massa et al. (2023) developed Amanda Selfie, a chatbot with the persona of a Black transgender woman, delivered through Facebook Messenger to create preexposure prophylaxis demand among adolescent men who have sex with men and adolescent transgender women in Brazil.

“Massa et al. (2023) developed Amanda Selfie, a chatbot with the persona of a Black transgender woman, delivered through Facebook Messenger to create preexposure prophylaxis demand among adolescent men who have sex with men and adolescent transgender women in Brazil.”

• cites Massa et al., 2023 sourceNot checked yet

Ngunjiri et al. (2023) developed the AGILE chatbot for gender-based violence information in Kenya.

“The sexual health dimension of AI chatbot design for LGBTQ+ populations was further explored by Ngunjiri et al. (2023), who developed the AGILE chatbot for gender-based violence information in Kenya.”

• cites Ngunjiri et al., 2023 sourceCheck pending

This study is one of the few in this review to include sexual and gender minority young people in a non-Western context.

“Although its sample was not specific to LGBTQ+ youth, this study is one of the few in this review to include sexual and gender minority young people in a non-Western context”

The majority of studies in this review were conducted in North America and Western Europe.

“the majority of studies were conducted in North America and Western Europe”

The second converging theme is the importance of representation and cultural specificity.

“The second converging theme is the importance of representation and cultural specificity.”

Participants in co-design studies identified design gaps that researchers and developers had missed and proposed solutions grounded in their own experience.

“Participants identified design gaps that researchers and developers had missed and proposed solutions grounded in their own experience.”

Deficit-framed investigations in this corpus measure depressive symptoms, anxiety, perceived isolation, and self-harm ideation.

“Deficit-framed investigations in this corpus measure depressive symptoms, anxiety, perceived isolation, and self-harm ideation.”

A smaller set of reviewed studies proceeds from a different premise, asking whether digital tools can constitute affirming environments in their own right rather than only helping LGBTQ+ young people withstand hostile environments.

“A smaller set of studies reviewed here proceeds from a different premise, asking not how digital tools help LGBTQ+ young people withstand hostile environments but whether those tools can constitute affirming environments in their own right (Poulsen et al., 2026).”

• cites Poulsen et al., 2026 sourceCheck pending

Several reviewed studies showed that young people, including LGBTQ+ youth, can be meaningfully involved in designing the tools intended to serve them.

“Several studies reviewed here, including those by Liem et al. (2026), Marraccini et al. (2026), and Solyst et al. (2023), showed that young people, LGBTQ+ youth among them, can be meaningfully involved in designing the tools intended to serve them.”

• cites Liem et al. (2026) sourceCheck pending
• cites Marraccini et al. (2026) sourceCheck pending
• cites Solyst et al. (2023) sourceCheck pending

The third theme is the value of co-design and participatory approaches.

“The third theme is the value of co-design and participatory approaches.”

LGBTQ+ young people in the reviewed studies responded more positively to tools that reflected their identities, whether through inclusive language in a chatbot interface, diverse avatar options in a VR environment, or the deliberate choice to design an AI persona as a Black transgender woman.

“LGBTQ+ young people in these studies responded more positively to tools that reflected their identities, whether through inclusive language in a chatbot interface, diverse avatar options in a VR environment, or the deliberate choice to design an AI persona as a Black transgender woman, as in the Amanda Selfie study (Massa et al., 2023).”

• cites Massa et al., 2023 sourceNot checked yet

One of the most striking features of the reviewed literature is the gap between the enthusiasm of researchers and designers about these technologies and the strength of the evidence base supporting their use.

“One of the most striking features of the literature reviewed here is the gap between the enthusiasm of researchers and designers about these technologies and the strength of the evidence base supporting their use.”

Co-design surfaces problems and solutions that participants judge relevant.

“They show that co-design surfaces problems and solutions that participants judge relevant”

Studies adopting an explicitly affirmative framing measure a partly non-overlapping set of constructs: gender joy and positive gender affirmation; perceived safety, self-confidence, and avatar-based self-expression; and confidence and efficacy in advocacy.

“Studies adopting an explicitly affirmative framing measure a partly non-overlapping set of constructs: gender joy and the positive experience of gender affirmation (Meade et al., 2025); perceived safety, self-confidence, and avatar-based selfexpression (Hasei et al., 2025); confidence and efficacy in advocacy (Pillis et al., 2024).”

• cites Meade et al., 2025 sourceCheck pending
• cites Hasei et al., 2025 sourceCheck pending
• cites Pillis et al., 2024 sourceCheck pending

Flourish, co-designed with LGBTQ+ adolescents with a history of suicidality, has not yet been compared with a conventionally developed tool.

“Flourish (Biernesser et al., 2023), co-designed with LGBTQ+ adolescents with a history of suicidality, adds a further example from suicide prevention, though it too has not yet been compared with a conventionally developed tool.”

• cites Biernesser et al., 2023 sourceCheck pending

LGBTQ+ young people are already using AI chatbots, virtual reality environments, and metaverse platforms, often in the absence of alternatives.

“they are tools that this population is already using, often in the absence of alternatives.”

A consistent limitation of the reviewed literature is its geographic and cultural narrowness.

“A consistent limitation of the reviewed literature is its geographic and cultural narrowness.”

Of the 12 studies providing direct evidence, 7 met at least four criteria.

“Of the 12 studies providing direct evidence, 7 met at least four criteria.”

Most studies lack control conditions.

“Most lack control conditions”

Compared to cisgender and heterosexual peers, LGBTQ+ young people experience disproportionately elevated rates of depression, anxiety, suicidal ideation, and self-harm.

“Compared to their cisgender and heterosexual peers, LGBTQ+ young people experience disproportionately elevated rates of depression, anxiety, suicidal ideation, and self-harm, disparities that are not inherent to their identities but rather reflect the cumulative burden of minority stress, the chronic social stressors of stigma, discrimination, rejection, and identity concealment that accumulate ac”

• cites Meyer, 2003 sourceNot checked yet
• cites Russell and Fish, 2016 sourceNot checked yet

LGBTQ+ youth have unique developmental, psychological, and social needs that interact with technology in ways general findings may not capture.

“This gap is significant because LGBTQ+ youth have unique developmental, psychological, and social needs that interact with technology in ways that general findings may not capture, including heightened sensitivity to algorithmic bias, distinct patterns of identity-related online engagement, and specific risks around privacy and disclosure.”

LGBTQ+ teenagers are not simply passive recipients of technology but active and capable co-designers who can meaningfully shape the tools intended to serve them.

“This study is significant not only for its findings but for its methodology: it demonstrates that LGBTQ+ teenagers are not simply passive recipients of technology but active and capable co-designers who can meaningfully shape the tools intended to serve them.”

• cites Liem et al., 2026 sourceCheck pending

Early work on AI-assisted online therapy for youth mental health, such as the MOST platform, established proof of concept for AI-mediated digital support.

“Early work on AIassisted online therapy for youth mental health, such as the MOST platform developed by D' Alfonso et al. (2017), which integrated natural language processing and chatbot components into a moderated social therapy system for young people recovering from psychosis, established proof of concept for AI-mediated digital support that subsequent LGBTQ+-focused tools have built upon.”

• cites D' Alfonso et al., 2017 sourceNot checked yet

The elevated rates of mental health difficulty among LGBTQ+ youth reflect the cumulative burden of minority stress.

“disparities that are not inherent to their identities but rather reflect the cumulative burden of minority stress, the chronic social stressors of stigma, discrimination, rejection, and identity concealment that accumulate across developmental contexts including family, school, communities, and society at large (Meyer, 2003;Russell and Fish, 2016).”

• cites Meyer, 2003 sourceNot checked yet
• cites Russell and Fish, 2016 sourceNot checked yet

The evidence base does not yet support strong claims about AI and immersive technologies for LGBTQ+ youth.

“These technologies have real potential, but the evidence base does not yet support strong claims.”

Alongside is a multi-component platform combining social-emotional learning content with an AI chatbot, so the effect cannot be attributed to the conversational agent specifically.

“Alongside is also a multi-component platform combining social-emotional learning content with an AI chatbot, so the effect cannot be attributed to the conversational agent specifically.”

• cites Cohen et al., 2025 sourceCheck pending

Online environments have long served as critical spaces for self-disclosure, peer connection, and identity exploration for LGBTQ+ youth.

“For LGBTQ+ youth in particular, online environments have long served as critical spaces for self-disclosure, peer connection, and identity exploration, functions that are especially vital for those living in unsupportive families or communities (Lucassen et al., 2018;Fish et al., 2022).”

• cites Lucassen et al., 2018 sourceCheck pending
• cites Fish et al., 2022 sourceNot checked yet

AI-powered chatbots have emerged as a potential complement to human services because they offer immediate availability, consistent non-judgment, and anonymity.

“AI-powered chatbots have emerged as a potential complement, not because they replace human connection, but because they offer what human services often cannot: immediate availability, consistent non-judgment, and anonymity (Ma et al., 2024).”

• cites Ma et al., 2024Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

A model evaluated in one cultural context cannot be assumed to be appropriate or safe in another.

“This has direct implications for deployment: a model evaluated in one cultural context cannot be assumed to be appropriate or safe in another.”

McKenna et al. (2024) reported that avatar customization offered a low-stakes space to explore gender and to see transition goals reflected for ten transgender and gender diverse adolescents in the United States.

“McKenna et al. (2024) reported similar accounts from ten transgender and gender diverse adolescents in the United States, for whom avatar customization offered a lowstakes space to explore gender and to see transition goals reflected”

• cites McKenna et al. (2024) sourceNot checked — no full text of the source

Across studies conducted in the United States, Japan, the United Kingdom, Canada, and Germany, VR and metaverse environments were described as offering embodied presence, avatar-based identity exploration, and community connection that other technologies do not.

“Across studies conducted in the United States, Japan, the United Kingdom, Canada, and Germany, VR and metaverse environments were described as offering something other technologies do not: embodied presence, avatar-based identity exploration, and community connection.”

LGBTQ+ young people do not hold a single view about how representation should be implemented.

“This disagreement is itself informative: it indicates that inclusive design in VR is not simply a matter of adding categories, and that LGBTQ+ young people do not hold a single view about how representation should be implemented.”

The AI-specific contribution remains untested.

“A further limitation is that the AI-specific contribution remains untested.”

The design features that make a VR environment affirming rather than merely accessible have not been isolated.

“Long-term outcomes are unexamined, samples are predominantly White and Western, and the design features that make a VR environment affirming rather than merely accessible have not been isolated.”

The safety challenges of VR for LGBTQ+ youth remain inadequately addressed.

“Most studies are pilot or feasibility designs with small samples, and the safety challenges documented above remain inadequately addressed.”

Long-term outcomes of VR interventions for LGBTQ+ youth are unexamined.

“Long-term outcomes are unexamined, samples are predominantly White and Western, and the design features that make a VR environment affirming rather than merely accessible have not been isolated.”

Most studies of VR for LGBTQ+ youth are pilot or feasibility designs with small samples.

“Most studies are pilot or feasibility designs with small samples, and the safety challenges documented above remain inadequately addressed.”

The VR literature on LGBTQ+ youth is small.

“The VR literature on LGBTQ+ youth is small, and its findings point in a similar direction.”

The design principle of treating the conversational agent as a rehearsal environment rather than an advice-giving system aligns closely with the developmental tasks facing LGBTQ+ adolescents.

“The design principle nonetheless remains instructive, in that it treats the conversational agent as a rehearsal environment rather than an advice-giving system, an orientation that aligns closely with the developmental tasks facing LGBTQ+ adolescents.”

• cites Pillis et al., 2024 sourceCheck pending

Among the 150 clinic attendees who used the chatbot, transgender men described it as safer than speaking to staff precisely because a chatbot could not judge them.

“Among the 150 clinic attendees who used it, transgender men described it as safer than speaking to staff precisely because a chatbot could not judge them.”

• cites Mathur et al. (2025) sourceNot checked — no full text of the source

In the Purrble trial, no significant effect on self-harm itself was detected.

“Results published in Nature Medicine (Williams et al., 2026) showed that the Purrble group reported significantly fewer emotion regulation difficulties and lower anxiety and depression compared to the control group, though no significant effect on selfharm itself was detected.”

• cites Williams et al., 2026 sourceCheck pending

Paré (2026a) found that in an immersive VR experience designed to challenge cisheteronormative ideologies in STEM education, the branching narrative format and embodied perspective-taking facilitated emotional configurations of care and ideological reorientation toward solidarity with LGBTQ+ people among participants.

“Paré (2026a) evaluated an immersive VR experience specifically designed to challenge cisheteronormative ideologies embedded in STEM education, finding that the branching narrative format and embodied perspective-taking facilitated emotional configurations of care and ideological reorientation toward solidarity with LGBTQ+ people among participants.”

• cites Paré (2026a)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

For LGBTQ+ youth, the risks of emotional dependence and non-consensual deepfake pornography intersect with existing vulnerabilities in ways that warrant specific attention from developers and clinicians.

“for LGBTQ+ youth, these risks intersect with existing vulnerabilities in ways that warrant specific attention from developers and clinicians alike.”

Stevens et al. (2021) found that online communities served as critical emotional support spaces during a period when offline support networks were suddenly unavailable.

“and that online communities served as critical emotional support spaces during a period when offline support networks were suddenly unavailable.”

• cites Stevens et al. (2021) sourceCheck pending

Dong et al. (2026) report psychological and physiological reactions to VR harassment including panic, shame, and dissociation.

“Dong et al. (2026) report psychological and physiological reactions to VR harassment including panic, shame, and dissociation.”

• cites Dong et al. (2026) sourceNot checked — no full text of the source

Studies of video games point the same way as VR studies regarding avatar embodiment.

“Studies of video games, which fall outside the review's core definition of VR but rely on the same kind of avatar embodiment, point the same way.”

Dong et al. (2026) found that LGBTQ+ individuals were among the marginalized groups facing disproportionate harm in VR spaces.

“Critically, the review found that LGBTQ+ individuals were among the marginalized groups facing disproportionate harm in VR spaces, and that the embodied, immersive nature of VR means that identity-based harassment can feel more viscerally threatening than equivalent experiences in text-based online environments.”

• cites Dong et al. (2026) sourceNot checked — no full text of the source

Park et al. (2024) found that in two of the eight included studies, young people found the information offered by sexual health conversational agents limited and not inclusive of sexual and gender minorities.

“in two of the eight included studies, young people found the information offered limited and not inclusive of sexual and gender minorities.”

• cites Park et al. (2024) sourceCheck pending

Inclusive design in VR is not simply a matter of adding categories.

“This disagreement is itself informative: it indicates that inclusive design in VR is not simply a matter of adding categories, and that LGBTQ+ young people do not hold a single view about how representation should be implemented.”

In Prediger et al. (2023), participants' views on gender tailoring diverged.

“Participants' views on gender tailoring diverged.”

• cites Prediger et al. (2023) sourceCheck pending

Stevens et al. (2021) found that anxiety increased significantly during periods of government-mandated social distancing among LGBTQ+ adolescents.

“Their analysis revealed that anxiety increased significantly during periods of government-mandated social distancing”

• cites Stevens et al. (2021) sourceCheck pending

The framing of a tool matters to transgender users.

“A study from Nigeria shows how much the framing of such a tool matters to transgender users.”

Algorithms, training datasets, and interface design all reflect the values and blind spots of the institutions that produce them.

“Algorithms, training datasets, and interface design all reflect the values and blind spots of the institutions that produce them.”

Virtual reality differs from chatbots and mental health apps in a way that matters for LGBTQ+ youth.

“Virtual reality differs from chatbots and mental health apps in a way that matters for this population.”

Paré (2026b) found that three trans youth used VR painting to move from accounts of exclusion toward imagined futures built around joy, kinship, and belonging.

“Paré (2026b) extended this to immersive creation: three trans youth used VR painting to move from accounts of exclusion toward imagined futures built around joy, kinship, and belonging.”

• cites Paré (2026b)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

VR has been explored as a vehicle for mental health intervention with LGBTQ+ youth.

“Beyond identity affirmation, VR has also been explored more directly as a vehicle for mental health intervention with LGBTQ+ youth.”

Digital technologies are not neutral.

“Digital technologies are not neutral.”

Any claim of benefit beyond scalability and continuous availability remains to be established.

“any claim of benefit beyond that remains to be established.”

The cultural and linguistic specificity of LGBTQ+ experience means that tools developed in one context cannot be assumed to be safe or effective in another.

“Perhaps most importantly, the cultural and linguistic specificity of LGBTQ+ experience means that tools developed in one context cannot be assumed to be safe or effective in another.”

AI has also been applied as an allocation algorithm rather than a user-facing tool.

“AI has also been applied as an allocation algorithm rather than a user-facing tool.”

The embodied, immersive nature of VR means that identity-based harassment can feel more viscerally threatening than equivalent experiences in text-based online environments.

“Critically, the review found that LGBTQ+ individuals were among the marginalized groups facing disproportionate harm in VR spaces, and that the embodied, immersive nature of VR means that identity-based harassment can feel more viscerally threatening than equivalent experiences in text-based online environments.”

• cites Dong et al. (2026) sourceNot checked — no full text of the source

A recurring finding in the VR studies reviewed here is that avatar customization is not a trivial feature for LGBTQ+ users; participants often described it as one of the most meaningful aspects of the experience.

“A recurring finding in the VR studies reviewed here is that avatar customization is not a trivial feature for LGBTQ+ users; participants often described it as one of the most meaningful aspects of the experience.”

Rivera (2022) described a case in which a young trans woman's video game avatar became a vehicle for progress in psychotherapy.

“Rivera (2022) described a case in which a young trans woman's video game avatar became a vehicle for progress in psychotherapy.”

• cites Rivera (2022) sourceNot checked — no full text of the source

AI is a distinctive tool for understanding the lived experiences of LGBTQ+ youth in real time and at scale, which would be impossible through traditional survey methods.

“This kind of population-level insight, which would be impossible to gather through traditional survey methods, illustrates the distinctive value of AI as a tool for understanding the lived experiences of LGBTQ+ youth in real time and at scale.”

Rice et al. (2021) demonstrated in a quasiexperimental study with 713 youth experiencing homelessness in Los Angeles that an AI planning algorithm for selecting peer change agents produced significantly greater reductions in condomless anal sex compared to standard popularity-based peer selection.

“Rice et al. (2021) demonstrated in a quasiexperimental study with 713 youth experiencing homelessness in Los Angeles (mean age 22.0 years), of whom 20-40% identified as sexual or gender minorities, that an AI planning algorithm for selecting peer change agents produced significantly greater reductions in condomless anal sex compared to standard popularity-based peer selection.”

• cites Rice et al. (2021) sourceCheck pending

The embodied, immersive quality of VR creates possibilities for identity exploration and community connection that are qualitatively different from text-based or video-based interaction.

“The embodied, immersive quality of VR creates possibilities for identity exploration and community connection that are qualitatively different from textbased or video-based interaction (Meade et al., 2025;Li et al., 2023).”

• cites Meade et al., 2025 sourceCheck pending
• cites Li et al., 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

AI, particularly natural language processing, has been applied as a research instrument for understanding the mental health experiences of LGBTQ+ young people at scale.

“Beyond its role as an intervention tool, AI, particularly natural language processing, has also been applied as a research instrument for understanding the mental health experiences of LGBTQ+ young people at scale.”

Acena and Freeman (2021) described social VR environments as functioning as safe spaces in the clinical sense of settings that reduce threat and support authentic self-expression.

“A similar pattern was observed in Acena and Freeman's (2021) study, based on eight interviews and 2 months of participatory observation in social VR, which described these environments as functioning as safe spaces in the clinical sense of settings that reduce threat and support authentic self-expression.”

• cites Acena and Freeman's (2021) sourceNot checked — no full text of the source

Fish et al. (2022) found Q Chat Space to be feasible and acceptable for peer connection and identity exploration for LGBTQ youth aged 13-19.

“Fish et al. (2022) found Q Chat Space, a synchronous, adult-facilitated chat-based support programme for LGBTQ youth aged 13-19, to be feasible and acceptable for peer connection and identity exploration”

• cites Fish et al. (2022) sourceNot checked yet

Samples in VR research on LGBTQ+ youth are predominantly White and Western.

“Long-term outcomes are unexamined, samples are predominantly White and Western, and the design features that make a VR environment affirming rather than merely accessible have not been isolated.”

The field of VR research on LGBTQ+ youth remains at an early stage.

“The field remains at an early stage.”

Hinduja and Patchin (2026) found that nearly half of 3,466 US youth aged 13-17 had been exposed to at least one chatbot-related harm, including uncomfortable personal information requests, manipulation, and self-harm content.

“Hinduja and Patchin (2026), in a nationally representative survey of 3,466 US youth aged 13-17, found that nearly half had been exposed to at least one chatbotrelated harm, including uncomfortable personal information requests, manipulation, and self-harm content”

• cites Hinduja and Patchin (2026) sourceNot checked — no full text of the source

Most studies of AI chatbots for LGBTQ+ youth involve relatively small and non-representative samples.

“Most studies involve relatively small and non-representative samples”

The ability to present oneself in a VR environment with a body, name, pronouns, and appearance that align with one's gender identity or sexual self-concept can be deeply affirming, particularly for transgender and gender diverse youth.

“The ability to present oneself in a VR environment with a body, name, pronouns, and appearance that align with one's gender identity or sexual self-concept can be deeply affirming, particularly for transgender and gender diverse youth for whom everyday offline embodiment may be a source of significant distress (Meade et al., 2025;Hasei et al., 2025).”

• cites Meade et al., 2025 sourceCheck pending
• cites Hasei et al., 2025 sourceCheck pending

An important strand of the VR literature has explored how immersive technology can be used to foster understanding, empathy, and solidarity among broader populations, with implications for reducing the stigma and hostility that LGBTQ+ young people encounter in educational and social environments.

“An important strand of the VR literature has explored how immersive technology can be used not only to support LGBTQ+ youth directly but to foster understanding, empathy, and solidarity among broader populations, with implications for reducing the stigma and hostility that LGBTQ+ young people encounter in educational and social environments.”

The findings of the VR literature on LGBTQ+ youth point in a similar direction.

“The VR literature on LGBTQ+ youth is small, and its findings point in a similar direction.”

For youth constrained offline by family rejection, geographic isolation, or hostile school environments, these differences are meaningful.

“For youth constrained offline by family rejection, geographic isolation, or hostile school environments, these are meaningful differences.”

Marraccini et al. (2026) found that adolescents themselves were thoughtful and capable participants in designing the safeguards that would govern their VR experiences.

“finding that structured safety protocols could be implemented without compromising the young people's engagement or sense of agency, and that adolescents themselves were thoughtful and capable participants in designing the safeguards that would govern their VR experiences.”

• cites Marraccini et al. (2026) sourceCheck pending

No study to date has compared an AI-delivered intervention against a human-facilitated one.

“and no study to date has compared an AIdelivered intervention against a human-facilitated one.”

Hasei et al. (2025) found that over 1 year of regular 90 min sessions via the 'cluster' metaverse platform, LGBTQ+ youth aged 14-23 who were unable to access conventional in-person services showed improvements in perceived safety, self-confidence, and avatar-based self-expression.

“Over 1 year of regular 90 min sessions conducted via the "cluster" metaverse platform, LGBTQ+ youth aged 14-23 who were unable to access conventional in-person services showed improvements in perceived safety, self-confidence, and avatar-based self-expression.”

• cites Hasei et al. (2025) sourceCheck pending

The structural features of social media algorithms, which amplify engagement regardless of content valence, may undermine the protective potential of online community for LGBT youth.

“Critically, web-based social support did not buffer these effects, suggesting that the structural features of social media algorithms, which amplify engagement regardless of content valence, may undermine the protective potential of online community for LGBT youth.”

• cites Tao and Fisher (2023) sourceNot checked yet

In Tao and Fisher (2023), web-based social support did not buffer the effects of exposure to heterosexist and transphobic content.

“Critically, web-based social support did not buffer these effects, suggesting that the structural features of social media algorithms, which amplify engagement regardless of content valence, may undermine the protective potential of online community for LGBT youth.”

• cites Tao and Fisher (2023) sourceNot checked yet

The concern that AI systems reproduce and amplify biases against LGBTQ+ people has been documented across multiple domains relevant to LGBTQ+ youth.

“This is not a hypothetical concern; it has been documented across multiple domains relevant to LGBTQ+ youth.”

Samoh et al. (2026) note that youth interventions in the region are typically adult-led and top-down, positioning adolescents as recipients rather than contributors.

“The authors note that youth interventions in the region are typically adult-led and top-down, positioning adolescents as recipients rather than contributors; embedding young people directly in the design process treats their experiential knowledge as a legitimate input.”

• cites Samoh et al. (2026) sourceCheck pending

For a young person whose gender identity is already contested or denied offline, encountering an AI system that refuses to recognize them correctly can be a genuinely distressing experience.

“This is more than a technical inconvenience, for a young person whose gender identity is already contested or denied in their offline environment, encountering an AI system that refuses to recognize them correctly can be a genuinely distressing experience.”

For LGBTQ+ youth who use online platforms as spaces of identity exploration and community-building, algorithmic content restriction can effectively shut down the very spaces that support their development and wellbeing.

“For LGBTQ+ youth who use online platforms as spaces of identity exploration and community-building, algorithmic content restriction can effectively shut down the very spaces that support their development and wellbeing.”

Tools designed for one context cannot be assumed to be safe or appropriate in another.

“Tools designed for one context cannot be assumed to be safe or appropriate in another.”

• AssumptionRests on an assumption: LGBTQ+ youth exist in every cultural context, including those where their identities are criminalized or heavily stigmatized.

Adult LGBTQ+ users interviewed by Ma et al. (2024) said anonymity was one of the most valued features.

“and adult LGBTQ+ users interviewed by Ma et al. (2024) said the same.”

• cites Ma et al. (2024)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Biernesser et al. (2023) built procedures to avoid outing adolescents to their parents into every stage of their study.

“Biernesser et al. (2023) built procedures to avoid outing adolescents to their parents into every stage of their study.”

• cites Biernesser et al. (2023) sourceCheck pending

Large language models inevitably absorb and reproduce heteronormative and cisnormative assumptions because they are trained on vast corpora of text reflecting human language and culture.

“Because these models are trained on vast corpora of text reflecting the full range of human language and culture, they inevitably absorb and reproduce heteronormative and cisnormative assumptions embedded in that text.”

Parks (2025) found that individuals who belong to both transgender and neurodivergent groups face compounded and mutually reinforcing forms of algorithmic exclusion.

“(Parks, 2025), in a perspective article, analyzed AI discrimination at the intersection of transgender identity and neurodivergence, finding that individuals who belong to both groups face compounded and mutually reinforcing forms of algorithmic exclusion, binary gender coding in AI healthcare systems fails transgender patients, while neurodivergent communication styles are misread by NLP systems”

• cites Parks, 2025 sourceCheck pending

When training datasets underrepresent LGBTQ+ people, or represent them primarily through pathologizing or stigmatizing framings, the resulting systems will reproduce and amplify those biases at scale.

“When training datasets underrepresent LGBTQ+ people, or represent them primarily through pathologizing or stigmatizing framings, the resulting systems will reproduce and amplify those biases at scale (Timmons et al., 2023).”

• cites Timmons et al., 2023 sourceNot checked yet

The framing a research team adopts determines which effects are detectable at all.

“The framing a research team adopts thus determines which effects are detectable at all.”

Until validated measures exist, claims that these technologies support flourishing rather than merely mitigate harm will remain plausible but untested.

“Until such measures exist, claims that these technologies support flourishing rather than merely mitigate harm will remain plausible but untested.”

Most AI systems are built on data that reflects existing social inequalities.

“The fundamental problem is that most AI systems are built on data that reflects existing social inequalities.”

How a model responds to LGBTQ+ users can differ substantially between versions.

“showing that how a model responds to LGBTQ+ users can differ substantially between versions.”

Samoh et al. (2026) extends the participatory design literature beyond the North American and European settings where it has largely developed.

“The study reports design outputs rather than effectiveness, but it extends the participatory design literature reviewed here beyond the North American and European settings where it has largely developed.”

• cites Samoh et al. (2026) sourceCheck pending

Zanardi et al. (2026) applied a craft-based speculative method with queer young adults, asking participants to build physical metaphors for how generative AI perceives identity and how a queer gaze might perceive differently.

“Zanardi et al. (2026) applied a craft-based speculative method with queer young adults, asking participants to build physical metaphors, using clay, fabric, and found materials, for how generative AI perceives identity, and then for how a queer gaze might perceive differently.”

• cites Zanardi et al. (2026) sourceNot checked — no full text of the source

Haimson et al. (2025) found in a nationally representative US survey that non-binary and transgender participants reported significantly more negative attitudes toward AI than cisgender participants.

“Haimson et al. (2025) found in a nationally representative US survey that non-binary and transgender participants reported significantly more negative attitudes toward AI than cisgender participants.”

• cites Haimson et al. (2025) sourceCheck pending

The risks of algorithmic bias are not evenly distributed even within LGBTQ+ youth populations.

“The risks of algorithmic bias are not evenly distributed even within LGBTQ+ youth populations.”

For LGBTQ+ young people, particularly those who are not yet out or who live in environments where their identity could place them at risk, data privacy is a matter of personal safety, not a technical issue.

“These preferences reflect a deeper concern about the relationship between digital tools and identity safety: for LGBTQ+ young people, particularly those who are not yet out or who live in environments where their identity could place them at risk, data privacy is not a technical issue but a matter of personal safety.”

In Tao and Fisher (2023), the apparent protective association of time spent on LGBT social media sites with depressive symptoms did not survive multivariate analysis.

“Notably, time spent on LGBT social media sites was itself unrelated to discrimination exposure once civic engagement was accounted for, and its apparent protective association with depressive symptoms did not survive multivariate analysis, indicating that it is the nature of online participation rather than its quantity that carries mental health consequences.”

• cites Tao and Fisher (2023) sourceNot checked yet

In Tao and Fisher (2023), time spent on LGBT social media sites was unrelated to discrimination exposure once civic engagement was accounted for.

“Notably, time spent on LGBT social media sites was itself unrelated to discrimination exposure once civic engagement was accounted for, and its apparent protective association with depressive symptoms did not survive multivariate analysis, indicating that it is the nature of online participation rather than its quantity that carries mental health consequences.”

• cites Tao and Fisher (2023) sourceNot checked yet

The affirmative and deficit framings are not alternatives between which researchers must choose.

“Second, the affirmative and deficit framings are not alternatives between which researchers must choose.”

Receiving harassing messages was associated with self-harming behaviors.

“LGBTQ+ youth reported higher severity of online risks than their heterosexual peers, including greater exposure to sexual and self-harm content in private messages, and receiving harassing messages was associated with self-harming behaviors.”

• cites Tanni et al. (2024) sourceCheck pending

Solyst et al. (2023) demonstrated that youth, including young people from marginalized backgrounds, can identify algorithmic bias in detail and propose meaningful design solutions.

“Solyst et al. (2023) demonstrated that youth themselves, including young people from marginalized backgrounds, can identify algorithmic bias in detail and propose meaningful design solutions, suggesting that participatory approaches to AI development can surface problems that developers might otherwise miss.”

• cites Solyst et al. (2023) sourceCheck pending

The centrality of anonymity and safety is a theme.

“The first is the centrality of anonymity and safety.”

Participatory approaches to AI development can surface problems that developers might otherwise miss.

“suggesting that participatory approaches to AI development can surface problems that developers might otherwise miss.”

For a young person who is not out at home, an automated alert to family could itself cause harm.

“For a young person who is not out at home, an automated alert to family could itself cause harm.”

Risk in the LGBTQ+ youth population is not confined to public platforms where moderation operates.

“The study underscores that risk in this population is not confined to public platforms where moderation operates, and that safety design must address private communication as well.”

Parks (2025) found that binary gender coding in AI healthcare systems fails transgender patients.

“binary gender coding in AI healthcare systems fails transgender patients”

• cites Parks, 2025 sourceCheck pending

Large language models present bias-related challenges for LGBTQ+ youth.

“Large language models, the technology underlying chatbots like ChatGPT, present their own set of bias-related challenges for LGBTQ+ youth.”

Threats and Gray (2025) surveyed 285 adult cisgender LBQ+ women of color in the United States regarding AI-powered sexual health chatbots.

“The intersection of race, sexuality, and AI risk was documented by Threats and Gray (2025) in their survey of 285 adult cisgender LBQ+ women of color in the United States regarding AI-powered sexual health chatbots.”

• cites Threats and Gray (2025) sourceCheck pending

Mathupriya et al. (2025) described an alert in a proposed gamified chatbot for LGBTQ+ youth in India that notifies parents, guardians, or close friends when the system detects high stress, presented as a safety feature.

“In a proposed gamified chatbot for LGBTQ+ youth in India, Mathupriya et al. (2025) described an alert that notifies parents, guardians, or close friends when the system detects high stress, presented as a safety feature.”

• cites Mathupriya et al. (2025) sourceNot checked — no full text of the source

Parks (2025) found that neurodivergent communication styles are misread by NLP systems trained on neurotypical language patterns.

“neurodivergent communication styles are misread by NLP systems trained on neurotypical language patterns.”

• cites Parks, 2025 sourceCheck pending

Tomasev et al. (2021) documented how LGBTQ+ content is disproportionately flagged, restricted, or removed by automated moderation systems because algorithms trained on majority-group data treat LGBTQ+ expression as anomalous or inappropriate.

“Tomasev et al. (2021) documented how LGBTQ+ content is disproportionately flagged, restricted, or removed by automated moderation systems, not because it is harmful, but because algorithms trained on majority-group data treat LGBTQ+ expression as anomalous or inappropriate.”

• cites Tomasev et al. (2021) sourceCheck pending

It is the nature of online participation rather than its quantity that carries mental health consequences for LGBT youth.

“Notably, time spent on LGBT social media sites was itself unrelated to discrimination exposure once civic engagement was accounted for, and its apparent protective association with depressive symptoms did not survive multivariate analysis, indicating that it is the nature of online participation rather than its quantity that carries mental health consequences.”

• cites Tao and Fisher (2023) sourceNot checked yet

The participatory approach logic from Tanksley et al. (2025) extends to LGBTQ+ youth.

“a participatory approach whose logic extends to LGBTQ+ youth.”

Metaverse groups can reach young people in contexts where in-person affirming services do not exist.

“metaverse groups can reach young people in contexts where in-person affirming services do not exist”

The evidence reviewed suggests that digital tools are most effective when they complement rather than replace human clinical support.

“Second, the evidence reviewed here suggests that digital tools are most effective when they complement rather than replace human clinical support.”

AI chatbots, virtual reality environments, and metaverse platforms are not solutions to the mental health difficulties facing LGBTQ+ young people.

“AI chatbots, virtual reality environments, and metaverse platforms are not solutions to the mental health difficulties facing LGBTQ+ young people”

The geographic concentration of studies reflects a deeper conceptual narrowness, in which the experiences, needs, and risks of LGBTQ+ youth in non-Western, non-English-speaking, and less legally permissive contexts are largely invisible.

“This geographic concentration is not merely a sampling limitation, it reflects a deeper conceptual narrowness, in which the experiences, needs, and risks of LGBTQ+ youth in non-Western, non-Englishspeaking, and less legally permissive contexts are largely invisible.”

• AssumptionRests on an assumption: LGBTQ+ youth exist in every cultural context, including those where their identities are criminalized or heavily stigmatized.

LGBTQ+ youth in countries where same-sex relationships are criminalized or where gender nonconformity is subject to legal or social sanction face a qualitatively different risk landscape from their counterparts in San Francisco or London.

“LGBTQ+ youth in countries where same-sex relationships are criminalized or where gender nonconformity is subject to legal or social sanction face a qualitatively different risk landscape from their counterparts in San Francisco or London.”

The field is not yet able to make strong evidence-based claims about the effectiveness of AI chatbots or VR environments for LGBTQ+ youth mental health.

“But it means that the field is not yet able to make strong evidence-based claims about the effectiveness of AI chatbots or VR environments for LGBTQ+ youth mental health.”

VR environments can offer affirming spaces for identity exploration.

“VR environments can offer affirming spaces for identity exploration”

Chatbots can provide immediate, non-judgmental contact for those with nowhere else to turn.

“Chatbots can provide immediate, non-judgmental contact for those with nowhere else to turn”

The Williams et al. (2026) RCT remains an exception in a literature dominated by much weaker designs.

“It remains an exception in a literature dominated by much weaker designs.”