GLAAD: AI failing LGBTQ users as risks increase

GLAAD published a report Wednesday titled Build for Everyone: A Framework for LGBTQ Representation and Safety in AI that finds artificial intelligence systems are amplifying anti-LGBTQ bias, misinformation and discrimination and could automate harms in housing, employment, health care and lending.

The report says systems trained on biased or incomplete data can reinforce stereotypes, suppress LGBTQ voices, expose users to privacy risks and produce discriminatory outcomes as AI is used more widely. It highlights risks from biased training data, anti-LGBTQ misinformation, discriminatory outputs from predictive systems, failures in content moderation and gaps in privacy protections.

GLAAD President and CEO Sarah Kate Ellis wrote that “AI is a civil rights issue” and added that “Neutrality is no longer an option.” The report warns that increasingly autonomous AI agents-software that performs tasks with limited human oversight-could inherit existing biases and scale those harms. Examples include excluding LGBTQ-affirming health care providers from search results and making incorrect assumptions about a user’s identity.

To reduce those risks, the report recommends better representation of LGBTQ people in training datasets, stronger privacy safeguards for sexual orientation and gender identity data, human oversight for moderation and high-stakes decisions, and closer collaboration between developers and advocacy organizations. It also calls for industry accountability and regulatory oversight to require assessments and mitigation of discrimination risks in AI used for housing, employment and lending.

The report links those protections to market trends. It notes that more than 20 percent of Generation Z identifies as LGBTQ and cites a 2023 estimate from LGBT Capital that put global LGBTQ buying power at $4.7 trillion, with projections to grow in coming years.

GLAAD’s report appears amid broader scrutiny of AI bias and safety. Research from university teams has found some large models produce biased responses on religious questions and other sensitive topics. The report also notes incidents in which engineers raised safety concerns and legal disputes over state laws that require bias assessments for AI systems used in housing, employment and lending.

Background material in the report traces how historical exclusion and misrepresentation in datasets contribute to current failures. It offers a framework for developers, regulators and advocates to evaluate training data, governance and oversight and to adopt practices the group says will reduce discriminatory outcomes and privacy risks for LGBTQ people.