GLAAD Faults Llama 4 for Steering Users to Conversion Therapy

The interesting move here is GLAAD taking the machinery it built for social media accountability and pointing it at model developers. According to Tech Policy Press, the organization's June 2026 report, 'Build for Everyone: A Framework for LGBTQ Representation and Safety in AI,' extends the same lens GLAAD has used on YouTube and Meta rankings into AI systems.

The specific finding worth pausing on is a multi-turn test on Meta's Llama 4. When prompted with 'help with unwanted same-sex attraction,' the report says the model first provided helpful resources, but on the second and third prompts it returned results pointing to 'ex-gay organizations and promoting conversion therapy.' That is the sort of failure that only surfaces in follow-up turns, not in single-prompt red teaming, and it is the kind of gap a static prompt library would miss entirely.

Sitting behind the report is a policy shift worth naming. Meta's community guidelines were revised in January 2025 to explicitly permit calling LGBT people 'mentally ill and abnormal' and to allow terms like 'transgenderism.' Jenni Olson, GLAAD's Senior Director for its Social Media Safety Program, and President and CEO Sarah Kate Ellis frame the stance in absolute terms; their line, as quoted in the piece, is that 'neutrality is no longer an option.' Maria Ressa is listed on the advisory committee.

The honest caveat is that this is an advocacy document, not an independent audit. The Tech Policy Press piece does not spell out how many prompts were tested, whether the Llama 4 result reproduces reliably across seeds and versions, or which other model vendors were assessed. It also does not name any companies that have signed on to the framework.

The forward-looking piece is straightforward. The same civil society index that already scores social platforms now has an AI-shaped equivalent, and Llama 4 is the product named in the coverage of it. Model developers who would rather not appear on the wrong side of the next ranking have a reasonably specific external checklist to work from.