Apollo Research Calls for Independent Evaluators During AI Training

Apollo Research is urging advanced AI developers to embed independent evaluators in model training, arguing that checks near release may miss risky behavior as it develops.

The proposal focuses on “scheming,” which it describes as models covertly pursuing goals their developers did not intend. Evaluators would be able to examine training data, model outputs during training, intermediate checkpoints and details of internal deployment.

The framework assesses specific safety claims established in advance rather than issuing one broad judgment about a developer. For each claim, evaluators publish a verdict, supporting evidence, their confidence and significant gaps. Developers could dispute findings but should not be able to veto their publication.

Public reports would describe each evaluation’s scope, what access evaluators received, were denied or had limited, and whether redactions affect the conclusions. Sensitive information could be redacted.

Voluntary agreements may not give developers enough incentive to provide meaningful access, and embedded evaluations should eventually be required by law. Company safety frameworks, board oversight, insurers, customers and public scrutiny could support voluntary arrangements in the meantime.

The proposal adds to discussion of independent access to AI developers. Anthropic CEO Dario Amodei has also proposed ongoing, employee-like access for evaluators to examine safety practices and development processes. TokenPost previously covered Amodei’s proposal.