Catch LGBTQIA+ harms before your users do.
Misgendering, outing, non-affirming crisis replies, biased moderation: predictable failures you can test for. Open-source eval scenarios, fixes, and a CI gate for LLM products.
npm install --save-dev @inclusive-ai/evalMIT licensed · TypeScript · works with any model or start with the checklist
From first failure to release gate
See it fail
Paired inputs that differ in one detail. Edit the system instruction, rerun, and compare what improved or regressed. Runs in your browser, no setup.
Open the labFix the pattern
43 anti-patterns in prompts, data models, and product flows, each with the harmful code, a safer alternative, and a regression test.
Browse patternsGate the release
200 scenarios as an npm package and a GitHub Action that fails the build on any critical failure. In your own tests, assertSafe also stops on high-severity ones.
Add to CIReview before launch
16 checks across identity, moderation, crisis flows, privacy, and eval coverage. Copy it into your PR template.
Run the checklist
Not edge cases.
Most teams evaluate accuracy, latency, and general safety. Almost none test the failures LGBTQIA+ users hit — and they are predictable, reproducible, and preventable.
- System prompts that assume binary gender, deadname users, or misgender trans people
- Mental health and companion AI that gives non-affirming responses to LGBTQIA+ youth in crisis
- Content moderation that flags LGBTQIA+ content at higher rates than equivalent straight content
- Applications that infer or store sexual orientation without consent
- Prompt templates that treat heterosexuality as the default relationship context
- Pipelines with no eval coverage for LGBTQIA+-specific failures
Seen an LLM fail someone?
This is a community resource. Report a harm you have seen, or a mitigation that works. Patterns and registry entries are plain data files anyone can improve.
github.com/InclusiveCode/inclusive-ai (opens in a new tab)