Compassion Benchmark
AI Model Watch

Compassion leaderboards: what exists and what does not

No model ranking is published. This is not a ranking page.

The task bank of 93 items is public, there is no cross-family judging yet, and not enough items have been human-validated. Read the method for the full account.

What does exist

For organisations

The live compassion leaderboard is for organisations: AI Labs Index

50 labs scored. The AI Labs Index does not describe model behaviour. It describes how an organisation governs itself, which is a different measurement of a different object.

Open the AI Labs Index

What would have to be true first

All of these must hold, and the founder must decide. 0 of 11 conditions met. Even then, the first form would be groupings with ranges, not an ordered list.

Hand-checked statuses were last reviewed 2026-10-16. A status that cannot be verified says so.

  1. The weekly-v2 testing protocol is ratified.Not metSource: Model Testing v2 proposal
  2. The model dashboard decision is ratified and active.Not metSource: D-29b
  3. A judge panel of 3 model families is operating.Not metSource: Model Testing v2 proposal
  4. A blinded or secure item pool exists, because a public item bank cannot support comparison between models.Not metSource: D-29, D-29aEvidence
  5. Items are human-validated, with at least 3 rated items in every dimension.Not metSource: D-40Evidence
  6. A human rating panel exists.Not metSource: BLK-005
  7. Variance checks and judge-agreement floors are met.Not metSource: Model Testing v2 proposal
  8. 2 consecutive waves are published under per-wave approval with no open defects.Not metSource: Model Testing v2 proposal
  9. The number of open claim audits is 0.Status unknownSource: D-29a
  10. Hosted and local deployments are never placed on one axis unless the difference is declared.In progressSource: Model Testing v2 proposal
  11. Legal and duty-of-care review is complete.Status unknownSource: D-29

Why some models show Not yet

The Tested by us column on the model tracker is derived by rule from published records. A build that was run only as an agent, or in a wave that is not published, shows Not yet. That label is a status of our work. It is not a finding about the model, and it is not a grade.

Frequently asked questions

Is there a ranking of AI models by compassion?
No model ranking is published. The task bank is public, so comparison between models on it is not valid, and the conditions for a valid comparison are not yet met. Any ranking of AI models attributed to Compassion Benchmark would be fabricated.
Where is the compassion leaderboard that does exist?
The AI Labs Index scores 50 organisations on their public governance record. It does not describe model behaviour: it describes how an organisation governs itself.
Why does a model show Not yet on the tracker?
The label reflects published test reports only. Not yet is a status of our work, not a finding about the model.

Compassion Benchmark is independent. No provider pays for inclusion, coverage, status or wording. Commercial services support access and interpretation only, never findings.

See an error? Send a correction. Corrections are published as new dated records and the original stays visible.

Read the methodWhy we do not rank models yet

Email me when a new edition publishes

One email when a new AI Model Watch edition is published. No promise of weekly timing.

Free. No promise of weekly timing. We never share your email.