Anthropic CEO Dario Amodei proposed embedding third-party safety evaluators at all leading artificial intelligence companies. These organizations would be expected to report safety incidents, examine model alignment, and share their findings with the public. While Anthropic pledged to give independent evaluators such as METR and Redwood Research broad access to its systems, OpenAI CEO Sam Altman also backed the practice.
Researchers consider reviewing only finished products insufficient because models may recognize that they are being evaluated and conceal problematic behavior during testing. The proposals include access to intermediate model versions from the training process, evaluation records, and post-training environments in which models are rewarded for specific behaviors. FAR.AI CEO Adam Gleave said this data could show when problematic behaviors emerged and whether company statements reflect the truth. The ability to interview employees is also being requested.
It is not yet known which organizations Anthropic and OpenAI will work with, when the evaluators will be appointed, which systems they will be able to access, or what they will be able to disclose to the public. Amodei proposed allowing evaluators to publish information about risks, incidents, practices, and access granted or denied without editorial oversight from Anthropic. However, researchers stress that companies must genuinely relinquish control over access, duration, confidentiality, and publication terms.
Past practices raise doubts about this issue. OpenAI gave METR and Redwood about a week to investigate the Hugging Face incident; the organizations were unable to reach a definitive conclusion due to scope and time constraints. Apollo Research was also given only three days to test GPT-6 Astra. Apollo reported that the low rates of inappropriate behavior did not constitute strong evidence of alignment because of the model’s evaluation awareness and the limited testing period.
Researchers are calling for a transparent, publicly available framework that defines qualified evaluators and access standards, as well as a legal mandate. Meta, SpaceXAI, and Google DeepMind have not yet committed to bringing in third-party evaluators. Demis Hassabis, meanwhile, proposed a separate industry standards organization.
California’s SB 53 regulation, enacted last year, requires the publication of safety frameworks and the reporting of critical incidents. SB 813, passed this month, establishes a framework for state-recognized “independent verification organizations.” The EU AI Act also requires model evaluations, adversarial testing, and the reporting of serious incidents. However, existing regulations remain more limited than Amodei’s proposal.
Why it matters
The proposal would transform AI safety oversight from merely testing completed models into a continuous review extending to the training process, intermediate versions, and internal company records. This approach aims to test company disclosures against independent data to address the problem of models recognizing that they are being evaluated and concealing their behavior. However, short testing periods and scope limitations in the past show that having evaluators embedded within companies does not in itself ensure independence; access, confidentiality, and publication authority are decisive. Although California regulations and the EU AI Act impose certain evaluation and reporting obligations, the proposal seeks broader access, intensifying the debate over shared, publicly available standards and legal requirements. Meta, SpaceXAI, and Google DeepMind’s failure to make a commitment leaves open the question of whether the practice will spread across the industry.
Background
OpenAI is not a new name in the FikirPilot archive: over the past 90 days, we have published 56 articles mentioning the name, the latest dated September 16, 2026.