Responsible AI Guidelines
Powerful models come with obligations that run in both directions. Here is what we commit to as the academy building them — and what we ask of everyone using them.
What we hold ourselves to
Open audits, always
Every model we release ships with its training data audit, evaluation results, and known failure modes. You should never have to take our word for what a model can or cannot do.
Consent-first affect research
Our emotional-intelligence work follows a hard rule: no covert inference of emotion, participant ownership of affect data, and open audits of every emotion model we train.
Red-teaming in daylight
We publish our red-team methods and findings, including the uncomfortable ones. Safety work done in secret cannot be verified — so ours isn't done in secret.
People before benchmarks
No model ships until our social-science program has studied what it does to the people who will live with it. Benchmark scores never override field evidence.
What we ask of users
Verify before you rely
Model outputs are drafts, not facts. Check claims independently before acting on them — especially in medical, legal, financial, or safety-critical contexts.
Disclose AI involvement
Don't present model-generated content as human-authored where that matters: journalism, scholarship, legal filings, or personal correspondence intended as your own.
Respect people in the data
Never attempt to re-identify individuals in our research datasets, and don't use our models to profile or manipulate people's emotions without their knowledge.
Report what you find
Found a failure mode, a bias, or a safety issue? Tell us at safety@ichancellors.com. We credit reporters in our published audits (or keep you anonymous — your call).
Found a safety issue?
We treat every report seriously and respond within two business days.