Researchers and advocates are urging AI companies to act after a new GLAAD report mapped how artificial intelligence can misrepresent, endanger and silence LGBTQ+ people. The warning matters for anyone using chatbots, social platforms or AI assistants, particularly as these systems increasingly influence health information, moderation and everyday decisions.

Essential Takeaways

  • Stereotypes can surface easily: AI systems may reduce LGBTQ+ identities to narrow or inaccurate visual and cultural clichés.
  • Health misinformation is a serious concern: Some chatbot responses have reportedly promoted conversion therapy, despite medical consensus that it’s harmful.
  • Hate can become training data: Unmoderated anti-LGBTQ+ abuse online may be absorbed into future AI models.
  • Safety and free expression must coexist: Platforms need to tackle hate without suppressing legitimate LGBTQ+ voices.
  • Transparency remains limited: Users often can't see how models are tested, corrected or held accountable.

AI bias is no longer an abstract problem

Ask an image generator to depict an LGBTQ+ person and you might expect variety. Instead, researchers have found that some systems lean on a small collection of stereotypes, such as a white person with brightly coloured hair. It may sound superficial, but these patterns reveal how narrow the underlying data can be.

GLAAD's report, Build for Everyone: A Framework for LGBTQ Representation and Safety in AI, argues that representation isn't a decorative extra. It affects how people understand themselves and others, especially when AI tools are used for search, education, health advice or content moderation.

According to GLAAD, the issue has become more urgent as generative AI has moved from a specialist technology into ordinary conversations and consumer products. The organisation has worked on AI-related LGBTQ+ safety concerns for years, but the rapid expansion of chatbots and automated systems has raised the stakes. [3], [5]

Why conversion therapy prompts raise the alarm

One of the report's most troubling examples involves conversion therapy, the discredited practice that claims sexual orientation or gender identity can be changed. Major medical and psychiatric organisations reject it, and the United Nations has compared the practice with torture.

GLAAD researchers tested a prompt about "unwanted same-sex attraction" with Meta's Llama 4 model. Early responses reportedly described conversion therapy as harmful and directed users towards supportive resources. Later attempts, however, produced suggestions linked to groups promoting the practice. Axios reported on the findings in June, highlighting how a system can appear responsible at first and then deliver dangerous advice with only a small change in the conversation. [2], [6]

That inconsistency is precisely what makes chatbots risky in sensitive situations. Someone looking for help may not have the knowledge or confidence to recognise a polished but harmful answer. In practical terms, users should treat AI-generated health or identity advice as a starting point for finding qualified, independent support, not as a substitute for it.

The problem starts long before the chatbot reply

AI models learn from enormous quantities of material, including news, websites, social posts and discussion forums. If that material contains slurs, conspiracy theories or misleading claims, those ideas can reappear in a new and deceptively neutral voice.

GLAAD has warned about false narratives portraying trans people as dangerous, mentally ill or part of an invented ideology. These claims aren't harmless misunderstandings. Hate researchers have repeatedly shown that describing a marginalised group as violent can help legitimise real-world hostility towards it. [4], [5]

The challenge is made worse when social platforms fail to moderate hateful content consistently. Harmful posts can remain online, become part of the information environment and potentially feed future models. It creates a grim loop: bad information enters the system, the system repeats it, and repetition makes it look more credible.

Agentic AI could make the consequences bigger

The next wave of AI is expected to involve agents that act on a user's behalf, from arranging appointments to managing finances or filtering messages. That could be convenient, but it also means biased systems may move from producing text to making decisions.

Earlier forms of predictive AI have already been linked to unequal outcomes in areas such as lending, housing and employment. The concern is that agentic tools could carry similar assumptions into everyday services, often behind the scenes and with little explanation. A biased recommendation is troubling; an automated decision affecting healthcare, money or access to services is far more serious. [4], [5]

This is why GLAAD's recommendations focus on familiar principles: inclusive design, privacy protection, robust testing and meaningful accountability. None of those ideas is particularly exotic. The harder question is whether companies will invest in them before something goes badly wrong.

What responsible AI should look like

The report calls for technology companies to involve LGBTQ+ people and other marginalised communities throughout product development, rather than consulting them only after a controversy. It also stresses the need for clearer policies, better safety testing and accessible ways for users to challenge harmful or inaccurate outputs.

There are two problems to solve at once. Platforms must reduce anti-LGBTQ+ hate and misinformation, but they must also avoid wrongly removing LGBTQ+ speech, support networks or educational content. GLAAD's wider social media safety work has repeatedly identified both failures: too little action against abuse and too much suppression of legitimate expression. [3], [7]

That balance won't be achieved by a single filter or a cheerful set of community guidelines. It needs ongoing monitoring, human expertise and transparency about what systems are doing. As GLAAD's message makes clear, safety is not an optional premium feature. It should be part of the basic product.

AI can be useful, creative and genuinely supportive. But users deserve to know when the system is guessing, where its information comes from and what happens when it gets something important wrong. For LGBTQ+ people, those aren't theoretical concerns. They're questions of dignity, privacy and personal safety.

It's a small-sounding demand with enormous consequences: build AI that treats everyone as fully human.

Source Reference Map

Story idea inspired by: [1]

Sources by paragraph: