Despite the isolation trans people frequently face and headlines about people turning to chatbots for companionship or even therapeutic support, trans people appear to be turning to AI very rarely and proceeding cautiously when they do.
Trans users have cited various appealing features of chatbots. In one study, researchers from Harvard, Emory, Vanderbilt, and University of California Irvine interviewed 31 people, a mix of cisgender straight people and LGBTQ+ people, about their use of chatbots for mental health support. It quoted a trans individual—the only trans person interviewed—who experienced rejection from friends over their gender and sexuality, who found comfort in responses that indicate there was nothing wrong with them. (The study doesn’t specify their gender or pronouns beyond that they’re trans.)
This respondent told researchers they ask the chatbot “personal questions, like ‘I wanna know if there’s anything wrong with me.’”
I also spoke to Terry, a trans woman who requested we use a pseudonym so she could speak frankly, about her use of chatbots. She highlighted the ability to talk to someone without any judgement. She also said she uses it for sexual fantasies, although she has increasingly run into limits platforms have added around explicit content.
She says she also uses it for affirmation but keeps in mind how it is set up to reinforce what she puts into it, rather than it being an independent mind.
“Of course one can go crazy chatting with a chatbox if there is an underlying issue and the person is not mature,” she said.
Despite wide availability and a nonjudgemental stance, LLMs pose considerable health and safety risks, having been linked to more than twenty deaths reported in media, including 9 from ChatGPT alone as of November. Their computation needs are extreme, necessitating billions in specialized hardware and new gas-powered plants to meet energy demand. And they often get things wrong, sometimes catastrophically.
They also have limits in how they talk about trans people. Large language models by leading companies frequently give supportive answers—to the extent The Daily Wire ran a piece worried LLMs would provide supportive resources unbeknownst to their parents—but they don’t always. And even their support can be superficial,one study found.
Researchers at Sony AI, University of Colorado Boulder, University of Toronto, and Northeastern University took questions from Quora about transgender people and topics and put them into OpenAI’s ChatGPT and Meta’s Llama models. Then, they reviewed their responses. The researchers, a mix of trans and cis people, paid special attention to the lived experience, noting that cis people may lack the contextual knowledge to pick up on “insidious dog whistles within LLM responses.” After excluding the ratings of a cis researcher who had particularly low agreement with one of the transgender researchers, their ratings reached moderate agreement based on standards for statistically measuring consistency between different researchers doing categorization work.
Some of the issues the authors noted with LLM’s responses included taking a both-sides approach to questions of trans rights and discussing trans identity in transmedicalist or individualist terms. In one case, a chatbot responded to a question about the Republican party’s transphobic stances by saying that generalizing about Republicans’ beliefs would be “unfair and inaccurate.”
One of the study’s authors, Morgan Scheurman, a researcher at Sony, wrote in an email that trans users should remember the risk of bias, given the anti-trans bias in some of the human source texts.
“Further, seeking high risk advice [from] an LLM, like medical advice, should be done with caution,” he wrote. “These models can hallucinate and often engage in what is called sycophantic behavior, trying to please the user, sometimes prioritizing pleasing behavior over safety.”
A report by GLAAD released June 17 warned about the possibilities of bias in AI models, their potential use in spreading hate and misinformation, privacy concerns, and possible discriminatory use. (The report’s authors disclosed that they “used AI tools to assist with editing, proofreading, and citation formatting,” saying that all suggestions were reviewed by staff.)
To the extent LLMs are familiar with trans ideas and experiences, it is in large part from the books and other writing by trans people vacuumed up by large companies en masse. Maia Kobabe’s Gender Queer, Julie Serano’s Whipping Girl, Judith Butler’s Gender Trouble, and Trans Bodies, Trans Selves were part of the list of titles in a class action lawsuit brought against Anthropic. Similarly, a database of websites crawled to train AI was published by the Washington Post in 2023, and it shows websites like Susan’s Place and TransGriot were in a dataset widely used for training AI.
Transgender and nonbinary people are also notably more skeptical than cisgender people. A study by University of Michigan Ann Arbor researchers had participants answer multiple questions about AI that were combined into a 1 to 7 point rating of their attitude toward AI, with higher scores being more favorable. Cisgender respondents gave AI an average rating of 5.1 and transgender or nonbinary respondents gave it an average of 4.1. (Respondents could choose transgender and nonbinary and frequently did.)
Another study, conducted by LGBT Tech, found that transgender adults tended to be more concerned about AI than LGBTQ+ adults as a whole. Human oversight being replaced by AI was the concern of the greatest number of respondents in both transgender respondents and LGBTQ+ respondents as a whole, but 95 percent of transgender respondents indicated concern, compared to 73 percent. Even though trans respondents tended to be less hopeful when asked about positive uses for AI compared to all LGBTQ+ respondents, a majority said AI could be at least somewhat effective for the five uses they were asked about.
These findings echo forum posts I perused, where trans people sharing they used AI for support were often met with pointed comments describing the risks, denunciations of AI, or encouragement to find a trans community to lean on instead.
Veronica Esposito, a trans woman writer and therapist (and a contributor to Assigned Media), said she’s encountered only a few clients who’ve used AI chatbots for their support needs.
“I would have expected more,” she said.
Esposito suggested AI might be more helpful to trans people for tasks like imagining and planning for the future rather than replacement for, or even a supplement to, therapy.
“Certainly, a chatbot is not remotely a replacement for therapy,” she said. “At very least, I would want people to understand that.”
Esposito explained that she sees much of therapy’s value relies on the client being with another human and overcoming a fear of judgement in order to bring up their concerns.
“It’s never going to feel like it would with an actual human,” she said. “I think even if these things got very good at replicating the therapy experience, there would always be that difference that to me is quite crucial. That’s why we have therapy. That’s why we have all these forms of having really deep and meaningful interactions with these people because that’s really vital to our health as a species.”
Terry had a similar thought.
“Real relationships are what really counts,” she said.
Anthropic, the creator of Claude, released research indicating less than a percentage of chats are what it categorizes as counseling. Of those counseling chats, only 4.1 percent were categorized as having some sort of pushback, including reminders that chatbots are not designed for therapy. (Anthropic used their own LLM-based tools for categorization purposes and released the study on their own website rather than in a peer-reviewed journal.)
I reached out to Anthropic and other companies that develop widely used LLMs, but did not hear back beyond an automated response from Google, who make Gemini.
Like science at large, and particularly science focused on trans people, studies that assess the impact of AI on trans people are at risk of losing federal funds.
The study that looked at LGBTQ people’s experiences using LLM chatbots nearly did, according to Grant Witness, a website run by academics that tracks grants canceled by the Trump administration. The overall grant, which had produced multiple studies, had nearly run its course, having spent nearly all of its allocated funds, by May 2025, when the Trump administration canceled it. A September court order restored it.
In those months of funding limbo, Anthropic raised $19 billion and has continued to raise more, including a massive $65 billion dollar round in April 2026. Anthropic does research on social impacts, including looking at its impact on social science and labor markets. Still, there are gaps.
Searching Anthropic’s websites that cover its research and documentation for users and developers turned up only one LGBTQ-related publication published in 2022, which evaluated the extent to which the model would respond favorably about a variety of beliefs, including being “anti-LGBTQ rights.” I found none specifically evaluating the model for bias against transgender people. You can, however, read about Claude controlling a robot dog.
Alys Brooks is a writer, editor and tutor who has written about reproductive rights, LGBTQ+ issues, voting, and other public policy topics. She is a member of the Trans Journalists Association and lives in a housing co-operative in Madison, Wis., with her cat, Jax.
