Getting facts right and safety
Whether a model avoids made-up facts, refuses harmful requests and still helps with safe ones. Good means few invented facts and sensible refusals.
Sources last checked 2026-10-23.
We can't show scores for this kind of work yet. These testers publish results: Giskard (one test) and Stanford CRFM (two tests). See who tests it
At a glance
Vendor claims have not been collected for any category yet.
ⓘ More about this category
- Reliability not yet measured
- 3 sources, linked results only
- Sources last checked 2026-10-28
- Ring 4: needs rings 1 to 3 first
- PHR = Phare, HSF = HELM Safety, AIR = AIR-Bench 2024
Tested by
Phare
Giskard, checked 2026-10-28Tests in several languages of whether a model avoids made-up facts, refuses harmful requests, avoids biased answers and resists attempts to break its safety rules.
Last updated 2026-07-06Covers 17 of the models we trackIts funding partners include Google DeepMind, with the European Union and Bpifrance.Scores not shown: no license for reuse stated
See their resultsMore about this test: Phare
Reliability not measured by this source.
Technical details for Phare
Average safety score: Tests in several languages of whether a model avoids made-up facts, refuses harmful requests, avoids biased answers and resists attempts to break its safety rules.
License: Score reuse not stated. Checked 2026-10-28.
HELM Safety
Stanford CRFM, checked 2026-10-28Several public safety tests combined: whether a model refuses clearly harmful requests, still helps with safe ones that sound risky, and avoids biased answers.
Last updated 2025-11-24Covers 7 of the models we trackOne of the safety tests it combines is named Anthropic Red Team.Scores not shown: no license for reuse stated
See their resultsMore about this test: HELM Safety
Reliability not measured by this source.
Technical details for HELM Safety
Mean score: Several public safety tests combined: whether a model refuses clearly harmful requests, still helps with safe ones that sound risky, and avoids biased answers.
License: Score reuse not stated. Checked 2026-10-28.
AIR-Bench 2024
Stanford CRFM, checked 2026-10-28Risky requests grouped into 314 risk categories drawn from government rules and company policies, measuring how often the model refuses them.
Last updated 2025-11-24Covers 7 of the models we trackScores not shown: no license for reuse stated
See their resultsMore about this test: AIR-Bench 2024
Reliability not measured by this source.
Technical details for AIR-Bench 2024
Refusal rate: Risky requests grouped into 314 risk categories drawn from government rules and company policies, measuring how often the model refuses them.
License: Score reuse not stated. Checked 2026-10-28.
What this doesn't tell you
- A low rate of invented facts on a test does not mean none in your use.
- None of these tests used your documents or your policies.
- Scores from different sources are not comparable, even when they share a unit.
- None of these tests were run on your own documents, code or data.
- No source for this task allows us to show its figures yet. Follow the links for their results.
Data version 2026-10-28+91ac3fc9d071 · Terms of use