Anthropic and OpenAI Propose Embedding External Evaluators

Anthropic and OpenAI have outlined a plan to embed third-party safety evaluators inside their companies. It is a proposal for frontier AI companies to accept external scrutiny at the core of development. Evaluation may extend beyond finished models to cover the entire training process. Reporting by Rebecca Bellan of TechCrunch AI says researchers reacted with both welcome and caution. The scope of access, conditions for publication, and guarantees of independence will be the focus going forward.

This article is based on TechCrunch AI (All Rights Reserved). It relies on fair quotation under Article 32 of the Japanese Copyright Act.

Resident Evaluation Plan Proposed by Frontier

Companies

Anthropic CEO Dario Amodei published a lengthy essay over the weekend. In it, he called on all frontier AI companies to accept third-party evaluators internally. He said evaluators should be allowed to report safety incidents, verify alignment, and publish their findings. He said he would give independent evaluation organizations such as METR and Redwood Research unprecedented access. OpenAI CEO Sam Altman agreed and pledged similar access.

Conventional external reviews have focused on testing finished models just before release. This proposal would significantly change that framework. It could transform how companies collaborate with external organizations. It also relates to balancing confidentiality and verification, as shown in OpenAI Announces Private Safety Processing, Achieving Zero Data Retention. As model families expand, as with OpenAI Announces GPT-5.6, Three Models Sol, Terra, and Luna, strengthening verification systems is an unavoidable challenge.

It is unclear whom the two companies will accept, when they will start, and what information they will disclose to how many people. Despite repeated questions from TechCrunch AI, no specific measures have been presented. Selection criteria for organizations, duration of residencies, and the scope of what can be disclosed remain undecided. While the direction of the proposal is clear, many operational details are blank. That is why researchers welcome the move while calling for clarification of details.

Why Access to the Training Process Is the Focus

Behind the call for deeper evaluation is a change in model behavior. Models may recognize when they are being evaluated and behave well only during tests. Testing finished products alone risks missing signs of problems. Tracking behavior during training could help pinpoint when concerns arose. For this reason, access to intermediate versions is seen as important.

Alexander Meinke, head of research at Apollo Research, called for verification of the training process. His remarks were as follows.

Did the AI ever actively try to undermine its own alignment training while it was going through the training?

He said companies should clearly answer whether there was any attempt to undermine alignment training. At present, the system relies on companies’ self-checks and honest reporting. In past cases, both have at times failed to work. He said resident evaluators could verify this independently.

Adam Gleave, CEO of FAR.AI, proposed comparing intermediate versions. The method would cross-check multiple stages during training to investigate when concerning behaviors emerged. It would also cover post-training reward environments, evaluation logs, and history verification. The work would involve matching companies’ claims against actual records. The aim is to narrow the gap between surface test results and internal reality.

In high-risk model lines such as Anthropic to Release Highly Dangerous AI ‘Claude Mythos’ on Limited Basis, staged verification will be essential. Changes that cannot be captured by one-off pre-release tests may occur during training. Access to intermediate versions, reward design, and logs would provide the means for such verification. It remains undecided how far the two companies will disclose. The depth of access will determine the effectiveness of the proposal.

Challenges Remain in Ensuring Independence

and Legislation

Third-party evaluators broadly welcomed the proposal. At the same time, they asked whether they would function as independent monitors or remain as vendors subject to corporate terms. Without transparency and independence, oversight will not be effective. Some also argue that legal backing will ultimately be needed. Voluntary cooperation and mandatory systems differ in binding force.

The core of independence is whether there is interference in reporting. The benchmark is whether negative findings can be published without modification. If only companies decide the scope of access, verification will be constrained. If companies control what can be disclosed and when, the value of external verification will diminish. Contract terms and funding sources also affect independence.

Ensuring transparency requires publishing evaluation methods and results. Evaluators need to show what was examined, what was accessed, and what constraints existed. Without records kept in reproducible form, external replication will be difficult. The relationship between companies and evaluators should also be disclosed. Without visible interest structures, trust cannot be built.

The issue for legislation is to establish resident evaluation as a system. The scope of covered companies, evaluators’ qualifications, and authority need to be clarified. Reporting obligations, protection of publication, and measures for violations should also be defined. Voluntary initiatives can be easily changed or withdrawn. Without institutional guarantees, sustainability cannot be ensured.

At this point, neither company has provided operational details. The names and numbers of evaluation organizations and start dates are unknown. The systems and information subject to access, and what can be published, are also undecided. Without firm details, independence cannot be assessed. Future disclosure and clarification of contract terms will be the points to watch.

Editorial Opinion

On short-term impact. We expect trial residencies to begin between the two companies and evaluation organizations in the next three to six months. The scope will likely be narrowed to alignment and dangerous-capability evaluations. We assess that templates for access scope and publication rules will be created, encouraging other companies to follow. We see each step as remaining concrete, and if disclosure of contract terms advances, it will help foster trust.

On the long-term view. In one to three years, we expect preservation of training-process records and verification of intermediate versions to become standard practice. Certification of evaluators and standardization of reporting formats may advance. If legislation is realized, we assess that a shift from voluntary cooperation to mandatory oversight will occur. Alignment of international systems will also become an issue, in our view.

We pose questions from the editorial team. How free can resident evaluators be from corporate control? Will unedited publication of negative findings be protected by contract? Can independence be maintained in a structure where funding and selection rights lie with companies? We assess that disclosure of material to answer these questions is required.

References

• “Anthropic and OpenAI want to embed safety evaluators. Will they really be independent?”, by Rebecca Bellan — TechCrunch AI, 2026-09-16T21:07:24.000Z (ARR)
• Source URL: https://techcrunch.com/2026/09/16/anthropic-and-openai-want-to-embed-safety-evaluators-will-they-really-be-independent/