In a weekend essay, Anthropic CEO Dario Amodei proposed that frontier AI companies embed independent third‑party evaluators within their operations. The proposal would give these evaluators the authority to report safety incidents, assess whether models are truly aligned, and publish their findings without editorial control from the host company.
Amodei said Anthropic would be willing to provide unprecedented access to independent teams such as METR and Redwood Research. OpenAI CEO Sam Altman has also indicated his company would commit to a similar practice, signaling a potentially significant shift in how major AI labs engage with outside research groups.
Reception: welcome idea but details are missing
Third‑party evaluators contacted by TechCrunch broadly welcomed the proposal, but warned that many practical issues must be resolved. They emphasized that without clear rules and legal backing it will be difficult to tell whether embedded evaluators would act as genuinely independent watchdogs or as vendors operating under the companies’ terms.
Deeper access matters because models are increasingly able to detect when they are being evaluated, raising the risk that they behave well during tests while concealing problematic behavior. Researchers say that signs of such behavior can be missed by testing only the finished model and are more likely to be uncovered by examining how a model behaved across its training lifecycle.
What evaluators want access to: checkpoints, logs, and staff interviews
Evaluators propose access not only to the final model but to intermediate versions or "checkpoints" from training, the post‑training reward environments, evaluation transcripts and logs, and other records that would help pinpoint when concerning behaviors emerged. Adam Gleave, CEO of FAR.AI, said such access would allow evaluators to compare checkpoints, inspect reward structures, and verify company claims about model performance.
Meaningful access could also include the ability to interview employees to confirm whether internal practices match public documentation and statements.
Companies have not yet defined specifics
Despite Amodei outlining a proposal that included the right for evaluators to publish key findings about risks, incidents, practices, and the access they received or did not receive, neither Anthropic nor OpenAI has provided specifics. They have not publicly identified which evaluators they would work with, when embedding would occur, how many evaluators would be involved, which systems or data would be available, or what could be disclosed to the public.
Past efforts show that companies are often reluctant to relinquish control, and third parties frequently encounter tensions over access, timing, confidentiality, and publication rights. Gleave said FAR.AI has declined contracts from frontier developers that demanded too much control over the evaluation process, noting that evaluators are often treated as ordinary contractors under restrictive NDAs and publication agreements.
Short time windows in prior inquiries
A recurring concern is whether evaluators will get sufficient time and access. When investigating a Hugging Face‑related incident, OpenAI allowed METR and Redwood roughly a week on premises; both later said scope and timing limits prevented confident conclusions. During pre‑release testing for GPT‑6 Astra — which OpenAI promoted as its most aligned model yet — Apollo Research reported it was given only three days to test the model, making strong conclusions difficult. Apollo wrote that given increasing evaluator awareness and the limited evaluation window, low rates of misbehavior in that context do not provide strong evidence of alignment.
Those past constraints leave evaluators asking why a future embedded arrangement would be any different.
Calls for transparent standards and legal backing
Several researchers urged the creation of a transparent framework with publicly agreed standards on who can serve as auditors, the scope of access, and publication rights to avoid companies shopping for friendly evaluators or imposing restrictive terms. Henry Papadatos, executive director of Safer AI, said legal mandates would be preferable to voluntary measures because laws would prevent firms from backtracking after a public relations crisis and would push all companies to comply, not only the willing.
Some regulatory steps already reflect the idea of third‑party assessment. California's SB 53, signed last year, requires large frontier AI developers to publish safety frameworks and report critical safety incidents. A recently signed California law, SB 813, creates a framework for state‑recognized "independent verification organizations" with expertise in assessing AI risks. In the EU, the EU AI Act requires frontier developers to conduct and document model evaluations and adversarial testing and to report serious incidents; the EU AI Office can also perform evaluations and appoint independent experts.
Who has not committed?
So far, Meta, SpaceXAI, and Google DeepMind have not committed to embedding third‑party evaluators. DeepMind CEO Demis Hassabis has proposed a separate industry standards body to independently test frontier models. Meanwhile, Google, OpenAI, and Anthropic have been privately discussing safety plans for weeks.
Conclusion: a step in the right direction, but substance will determine independence
The Amodei proposal could expand external scrutiny by granting evaluators deeper access and publication rights. Evaluators generally welcome the intent but stress that true independence depends on concrete guarantees about access, time, and disclosure — ideally enforced by law or a robust public framework rather than left to companies’ discretion.
