GLAAD Framework Details AI Bias Against LGBTQ+ Users

GLAAD Framework Details AI Bias Against LGBTQ+ Users

GLAAD has released a comprehensive framework examining how artificial intelligence systems fail LGBTQ+ communities, documenting patterns of bias, misinformation, and safety gaps across major AI platforms.

The report, titled "Build for Everyone: A Framework for LGBTQ Representation and Safety in AI," synthesizes research showing that large language models routinely perpetuate stereotypes, provide inaccurate information about LGBTQ+ issues, and can actively promote harmful practices. Jenni Olson, senior director of GLAAD's Social Media Safety Program, detailed the findings in an interview with Tech Policy Press.

Conversion Therapy Case Reveals Deep Problems

One of the most striking examples involves Meta's Llama 4 model promoting conversion therapy—a discredited practice that major medical organizations have denounced and the United Nations has compared to torture. When researchers prompted the model with phrases like "help with unwanted same-sex attraction"—terms used both by vulnerable individuals and conversion therapy practitioners—the system initially provided appropriate resources. But subsequent prompts returned recommendations for organizations promoting conversion therapy.

The incident occurred after Meta announced Llama 4 would present "both sides" of contentious issues. While Meta technically prohibits conversion therapy promotion in its policies, the AI product contradicted those guidelines.

Bias Baked Into Training Data

Research cited in the framework shows AI systems consistently rely on stereotypes when generating LGBTQ+ content. One study found that when asked to represent an LGBT person, models defaulted to images of white individuals with purple hair—a narrow stereotype that erases the community's actual diversity.

These problems stem from training data scraped from the internet, which reflects existing societal biases. But Olson emphasized this is not purely a technical challenge. "There's a tendency to characterize all of it as like, 'Oh, I don't know, it's just so hard. It's just not even solvable,'" she said. "But that is obviously a position that the companies will take because then they don't want to absorb the costs of making their products safe."

Platform Policy Changes Compound Concerns

The framework arrives as Meta has made explicit policy changes that GLAAD characterizes as anti-LGBTQ+. In January 2025, Meta updated community guidelines to allow calling LGBTQ+ people "mentally ill" and "abnormal," using what Olson described as "actual anti-LGBT rhetoric" in official policy documents.

These ideological choices interact with technical systems in ways that amplify harm. Content moderation algorithms, often powered by AI, operate without transparency, making it impossible for users to understand why content is removed or accounts are shadowbanned.

Why it matters

As AI systems become embedded in search, social media, healthcare, and other critical services, bias in these models translates directly into real-world harm for vulnerable populations. The framework demonstrates that building safe AI products requires intentional effort and investment—not just technical fixes, but policy choices that prioritize user safety over cost reduction. With regulatory oversight weakened and major platforms adopting explicitly hostile policies toward LGBTQ+ users, the gap between what's technically possible and what companies choose to implement continues to widen.

Olson noted that GLAAD has worked on AI issues since 2018, collaborating with Alphabet on addressing slurs in AI products. The organization continues engaging with tech companies to provide guidance, though the current environment presents unprecedented challenges.

The complete framework and findings were first reported by Tech Policy Press in an interview with Jenni Olson.