GLAAD CEO Sarah Kate Ellis highlighted the risks AI bias poses to the LGBTQIA+ community at the AI+ NY Summit held earlier this month. She introduced the new report, 'Build for Everyone: A Framework for LGBTQ Representation and Safety in AI,' which documents widespread algorithmic discrimination and outlines industry recommendations. The report notes that foundation models often encode systemic prejudice, citing a 2024 UNESCO study which found that Meta’s Llama 2 model generated negative content about gay people in approximately 70% of instances, characterizing them as criminals or abnormal.

Algorithmic discrimination often arises when training data mirrors societal biases, which then propagates through downstream applications. Beyond biased text, AI systems can infer sexual orientation or gender identity through behavioral data, search history, and social links, posing surveillance risks in regions where same-sex relationships are criminalized. This vulnerability is exacerbated by a challenging political climate, with the Trans Legislation Tracker reporting that 797 anti-LGBTQIA+ bills have been introduced in the U.S. as of June 2026.

Addressing these issues requires auditing systems using demographic data, a process that risks further exposing marginalized users if handled improperly. The Partnership on AI (PAI) developed 'Participatory & Inclusive Demographic Data Guidelines' to help developers collect sensitive information while ensuring privacy and safety. These guidelines were shaped by seven equity experts and emphasize the necessity of involving affected communities in the design of protection standards. The approach prioritizes 'privacy-by-design' and civil society engagement to prevent harms like those observed when Grindr users stopped disclosing HIV status after data sharing practices were revealed.

Industry implementation of these safety standards remains inconsistent while the rapid deployment of AI into healthcare, education, and public services continues to outpace regulatory protections. Organizations are urged to move beyond stated commitments to implement consistent, verifiable practices regarding data security, consent, and community-led fairness definitions. PAI emphasizes that building trust requires demonstrating reliable, protective standards rather than relying on abstract policy promises as AI becomes increasingly embedded in critical social systems.

GLAAD CEO Sarah Kate Ellis highlighted the risks AI bias poses to the LGBTQIA+ community at the AI+ NY Summit held earlier this month. She introduced the new report, 'Build for Everyone: A Framework for LGBTQ Representation and Safety in AI,' which documents widespread algorithmic discrimination and outlines industry recommendations. The report notes that foundation models often encode systemic prejudice, citing a 2024 UNESCO study which found that Meta’s Llama 2 model generated negative content about gay people in approximately 70% of instances, characterizing them as criminals or abnormal.

Algorithmic discrimination often arises when training data mirrors societal biases, which then propagates through downstream applications. Beyond biased text, AI systems can infer sexual orientation or gender identity through behavioral data, search history, and social links, posing surveillance risks in regions where same-sex relationships are criminalized. This vulnerability is exacerbated by a challenging political climate, with the Trans Legislation Tracker reporting that 797 anti-LGBTQIA+ bills have been introduced in the U.S. as of June 2026.

Addressing these issues requires auditing systems using demographic data, a process that risks further exposing marginalized users if handled improperly. The Partnership on AI (PAI) developed 'Participatory & Inclusive Demographic Data Guidelines' to help developers collect sensitive information while ensuring privacy and safety. These guidelines were shaped by seven equity experts and emphasize the necessity of involving affected communities in the design of protection standards. The approach prioritizes 'privacy-by-design' and civil society engagement to prevent harms like those observed when Grindr users stopped disclosing HIV status after data sharing practices were revealed.

Industry implementation of these safety standards remains inconsistent while the rapid deployment of AI into healthcare, education, and public services continues to outpace regulatory protections. Organizations are urged to move beyond stated commitments to implement consistent, verifiable practices regarding data security, consent, and community-led fairness definitions. PAI emphasizes that building trust requires demonstrating reliable, protective standards rather than relying on abstract policy promises as AI becomes increasingly embedded in critical social systems.