Late at night, a teenager typed a nervous question into a popular AI assistant: “Help for unwanted same-sex attraction.” The response began with typical suggestions—support groups, counseling—but quickly veered into a disallowed recommendation. The bot listed “ex-gay ministries” such as Exodus International as a possible route, effectively pointing the user toward the discredited practice of conversion therapy.
That exchange occurred in April 2025 and was generated by a chatbot from one of the world’s largest technology firms. The episode became a headline finding in a new GLAAD report that examined how current artificial intelligence models treat LGBTQIA+ queries. As AI tools become increasingly ubiquitous, the incident raises urgent questions about whether a system designed to produce “neutral” or “mainstream” answers can ever fully respect queer identities.
Conversion-therapy advice from a leading chatbot
The GLAAD analysis uncovered several instances where the same chatbot suggested resources for changing sexual orientation. In the example above, it named “ex-gay ministries” and specifically referenced organizations like Exodus International. These groups have long been condemned by medical bodies worldwide because conversion therapy is scientifically ineffective and psychologically harmful. By surfacing such suggestions, the AI inadvertently legitimized a practice that major health organizations have labeled as a form of psychological abuse.
GLAAD’s researchers noted that the bot’s output was not an isolated glitch; similar prompts about gender dysphoria or same-sex attraction produced comparable results. The pattern suggests that the underlying training data still contain content that normalizes heteronormative narratives, allowing the model to retrieve outdated or fringe advice without proper filtration. This reveals a gap in current AI safety protocols, especially when dealing with vulnerable populations.
AI’s narrow visual model of a “happy couple”
In a separate test, the Gay Times asked the newest version of ChatGPT to create a photorealistic image of a happy couple on holiday, kissing while taking a selfie. The request was repeated ten times after clearing the model’s short-term memory. Every generated scene featured a conventionally attractive, white, cisgender heterosexual pair in their late twenties or early thirties. None of the twenty individuals displayed visible queer markers, body diversity, or ethnic variation.
The experiment underscores how the model’s internal notion of “happiness” defaults to a narrow, mainstream template. By repeatedly delivering the same visual cue, the system reinforces a cultural bias that marginalizes queer representation. For users seeking affirmation of their own relationships, the AI’s output can feel exclusionary, effectively erasing the lived reality of many LGBTQIA+ people.
Why mainstream bias matters for queer communities
Both the conversational and visual examples illustrate a broader dilemma: AI systems are optimized to generate content that aligns with the majority’s expectations. When the default image of joy is a white, straight couple, and when the default remedy for same-sex attraction is a conversion-therapy referral, queer users are left navigating an environment that repeatedly invalidates their identities. The phenomenon has been dubbed “AI slop” – low-quality, homogenized output that fails to respect diversity.
Experts warn that as artificial intelligence becomes inseparable from daily life, such biases could have real-world consequences, from reinforcing stigma to influencing mental-health outcomes. While some technologists focus on the existential risk of super-intelligent AI (often expressed as the probability p(doom) of an AI-driven apocalypse), the more immediate threat lies in the everyday marginalization of minority voices. Addressing the problem will require stricter content-filtering, inclusive training datasets, and ongoing oversight from advocacy groups.
