Helpful or Harmful?

Peter James Steven has created an AI chatbot that amplifies the voices of detransitioners

Where do you go and who do you turn to if you are a young person distressed by your gender or a parent of a gender-questioning child? To a therapist or social worker, who will more than likely affirm a trans identity without much further ado? To friends or colleagues, many if not most of whom will either have accepted the modern “born in the wrong body” fairy tale or feel either unqualified or too afraid to offer a candid response? Or to the internet?

Peter James Steven, the New Zealand-based developer of detrans.ai, thinks the internet is bombarding us all with trans positivity. If you use AI chatbots, you will mainly get trans affirmative responses. ChatGPT, for example, is weighted towards supporting and upholding the belief that everyone has an innate gender identity.

Peter’s own detrans.ai system is designed as a counterweight. It is intended for people who are questioning their gender, people who identify as trans and are thinking about detransitioning, as well as families and friends who want to understand the issue better and understand how to support their loved ones. Detrans.ai fills an important gap. It provides responses to questions drawn directly from the often-neglected voices and perspectives of detransitioners. And, as Peter says, the very existence of detransition calls into question the mainstream idea that gender identity is innate.

Large Language Models and Culture

The large language models (LLMs) used by conventional chatbots mirror the evolving cultural beliefs and attitudes embedded in existing online content. Language models change with the broader discourse. But, as Peter points out, language models are themselves impacting the culture. The belief that there’s a right and a wrong way to be a boy or a girl or a man or a woman (“It’s literally all about stereotypes and roles”), that Peter thinks gender ideology boils down to, is itself reinforced by chatbots like ChatGPT. Whether the content of existing chatbots is deliberately manipulated or not remains a moot point. It’s certainly the case that models like ChatGPT have certain “guardrails or gradients” built in that tilt towards producing trans positive responses. The likely rationale for providers gearing their systems towards pro-trans perspectives is “harm reduction”. “They think it’s probably the safer way to go,” says Peter. But as we know from pre-X social media, the harm principle can—in complete contradiction to the liberal principles originally articulated by John Stuart Mill—very quickly degenerate into censorship of dissenting views. Nowhere is this more the case than in the debate around what “Do no harm” means in the context of the medical transition of trans-identifying minors and adults.

Detrans.ai is different. It uses the technique of retrieval-augmented generation to find and integrate the experiences of detransitioners into its responses. Peter’s system draws on every comment that’s been posted on the detrans subreddit, which, with around 60,000 members, is one of the largest communities of detransitioners on the internet. The AI model that powers the whole thing is Kimi K2.5. Peter opted to use this Chinese model because he found that Western LLMs uphold gender identity beliefs. The Chinese models he explored provided much clearer answers that crucially didn’t contradict the experiences articulated by detransitioners on the internet.

Peter James Steven’s personal motivation for creating detrans.ai was his sister’s identification as non-binary ten years ago (“the person who was my sister, but who now identifies as a man”). Transgender identities impact entire families. He has always been supportive of her choices and uses her preferred pronouns. It was their decision to start taking testosterone that shook his confidence and got him reading the stories of detransitioners. “I heard their regret,” he says, and started seeing his sister’s dark childhood and traumatisation. He struggles to think she’s making a good decision. It is her life, but he would like her to listen to detransitioners. “My project is just about, hey, listen to this”.

Open Source, Open Mind

Peter is aware of his own personal perspectives. He believes that people are being harmed by being led into a lifetime of medicalization and made dependent on the medical system, “when it seems like a lot of people just need to kind of talk through some issues, really.” This is why he has gone to great lengths to make detrans.ai as transparent as possible. He has made all the code open source, including all the chat prompts, so that anyone can read and see how the system actually works. The full source code is available on GitHub, allowing programmers to study its underlying design and functionality. All the system prompts are visible in the programme’s menu. The idea of detrans.ai is not to push Peter’s own views out into the world, but to let detransitioners speak on their own experiences. Detrans.ai enables users to listen to and learn directly from detransitioners’ voices.

Nonetheless, his work has elicited its fair share of negative feedback. Including from people Peter would have considered his friends. Wellington, where Peter lives, is a very progressive city. A lot of people, he says, see any discussion on this issue as a far-right and extremely conservative attack on trans people. “But I don’t view it that way. I think that this is progress, acknowledging and factoring these experiences into how people are treated when they show up at doctors.”

When I asked ChatGPT the simple question “What is detrans.ai?”, I received the answer:

“The site’s overall perspective is not neutral. It generally frames transition cautiously or skeptically and often emphasizes: regret after transition, underlying trauma or mental-health factors, social influence, and concerns about medical transition practices. Critics argue that it promotes anti-trans narratives and may steer vulnerable users away from transition.”

You might think it would be a good idea to steer vulnerable people away from transition. ChatGPT apparently not. The answer appears to confirm the observation that mainstream chatbots are closely aligned with the trans-affirmative perspective. Which makes detrans.ai all the more important. Peter is explicit about his system’s focus on the neglected voices of detransitioners. He doesn’t claim that it covers the entire trans experience. But listening to the voices of detransitioners enables us to ask what trans means to those who have been through transition and back, again. It turns out, as Peter says, that it can mean any number of things to different people. But one thing is certain, the rich experiences of detransitioners contradict by their very nature the theory of innate gender identity and the “born in the wrong body” hypothesis. And this may be the reason why detrans.ai is either studiously ignored or vehemently condemned by the proponents of rapid, unquestioning affirmation and immediate medicalisation of gender-questioning young people and adults.

Detrans.ai was, and continues to be, a labour of love.

Genspect publishes a variety of authors with different perspectives. Any opinions expressed in this article are the author’s and do not necessarily reflect Genspect’s official position. For more on Genspect, visit our FAQs.