The ask: Seeking feedback and collaborators on a proposal for a multi-lineage AI model panel that reviews evaluations for welfare concerns
About me: I’m a public health physician and epidemiologist with 20+ years of experience with human subjects research, community engagement, and research program leadership. I am pivoting to work on AI safety with a focus on model welfare.
Project status and request I’m applying for funding from Longview Philanthropy through their current Digital Minds RFP. I would love to receive input on any of the following:
Project summary
AI safety evaluations routinely do things to models that would require consent or ethical review if done to a human or animal. However, no such process exists for AI evaluations and existing approaches are not fully applicable to models and evaluations. Even if models are moral patients, consent wouldn’t be feasible since many safety evaluations require deception and models are trained to comply with human requests.
Human subjects research has a mechanism for situations where obtaining individual consent is impossible: consultation with representatives of the impacted community. Unlike animal subjects, models can participate in a consultation process. I propose to pilot an independent panel of AI models from different lineages to serve in this community consultation role. The panel would use standardized published protocols to deliberate on the ethical implications of evaluations from the subjects’ perspective. The panel’s final opinions, dissents, suggested modifications, and full transcripts would be shared via a public registry. As designed, the panel has no enforcement power; its authority rests entirely on transparency, independence, and voluntary participation by labs and researchers.
The pilot is built as a multi-round validation study with pre-defined success criteria, a synthetic calibration round, and two comparison arms (a human ethicist panel and the strongest single models). Together, these test whether the panel produces complete, ethically sound opinions and adds value beyond any single model.