Loading page…
We are in Beta, collecting meaningful, anonymous input to help shape how we approach AI safety models. Our methods and safeguards will evolve openly as we learn together.
Evidence, expertise, experience, and disagreement belong alongside one another.
AISafetyModels is in Beta. We are collecting meaningful, anonymous contributions and building this site together to explore how we should approach AI safety models. As we learn, we will continue to update the website and our approach safely and transparently.
Answer anonymously. No account required. Changing an answer replaces your previous choice. Polls show your saved answer, not aggregate public results.
Anonymous contributions help us understand participants’ experiences, priorities, and concerns. Participation is voluntary and should not be treated as a representative sample, expert consensus, or proof that an option is safe. We will explain the scope and limitations of any results we publish.
Every poll uses explicit answer options and shows its status and any closing date. During Beta, we are developing clearer records of question wording, collection periods, eligibility, and how questions are developed so participants can understand how their input is interpreted.
Participation is anonymous and does not require an account. Changing an answer replaces your previous choice. We are setting up and refining safeguards against repetitive votes and spam during Beta. These safeguards do not establish verified unique participants; we will explain their scope and limitations as they evolve.
Profiles name their scope and link the primary documentation used to describe it. Provider-stated behavior and capabilities are distinguished from our practical interpretation. We have not independently reproduced model benchmarks, measured latency or assigned safety scores. A product family is not one fixed deployment.
Explainers and perspectives are site editorial content informed by linked research and documentation. Citations identify supporting sources. We distinguish editorial interpretations and proposed design choices from research findings.
Any company involvement, sponsorship, or paid question development will need to be disclosed alongside the relevant content and results.
During Beta, we are developing contribution rules and moderation safeguards to reduce spam and abuse while protecting good-faith disagreement. We will explain review processes and how concerns can be raised as these features become available. Popularity should not erase minority experiences or substitute for technical analysis.
Participate anonymously. No account required. Please avoid including private or identifying information in written answers. External source links open their respective sites.
We prefer original model cards, provider documentation, research papers and guidance from the organization that produced it. Documentation establishes what a provider says; a research result establishes only what its study tested. Review dates indicate when we checked the material, not when every claim was independently verified.
When revising an article, preserve the relevant release and study scope, recheck the supporting link, and make clear when a conclusion is editorial. If a source changes or contradicts a profile, the profile should be corrected rather than silently extending an old claim to a new release. Linked sources do not independently validate participation data or establish that a popular choice is safe.