Embedded Evaluation

Embedded evaluation is the practice, proposed in September 2026, of placing independent evaluators inside a frontier AI developer on a standing basis, with access to people, systems and training processes comparable to an employee's, so that they can verify the company's safety commitments and report on what they find. It differs from the pre-deployment testing most third-party evaluators already perform, in which an evaluator receives access to a finished or near-finished model for a bounded period. The term and the practice were introduced by Dario Amodei's essay "We Must Pace the Frontier" (We Must Pace the Frontier (Amodei, September 2026)), and Anthropic named its first embedded evaluator six days later (Source: anthropic.com).

Origin

Amodei's essay of September 12, 2026 makes a team of "embedded third-party evaluators" with "ongoing, employee-like access" the first of three steps toward slowing the rate of frontier capability improvement, and the only one Anthropic committed to unilaterally. He describes the evaluators' role as verifying adherence to safety practices and commitments, reporting incidents, and helping assess the alignment "of not just completed AI models but training pipelines and processes," and calls embedded evaluation "the key step for verifiability of any pacing commitments." He cites the banking sector, where regulatory supervisors are sometimes embedded alongside employees, as precedent, and characterises the practice as "a quite radical practice that goes far beyond what any AI company is doing today." The essay calls on governments to require the step of other frontier companies and says "all frontier labs should partner with government to formalize the idea of permanent embedded evaluators" (We Must Pace the Frontier (Amodei, September 2026)).

The essay names METR as an example of the class of organization it has in mind. Anthropic's September 17, 2026 post on measuring the pace of AI development restated the plan, describing access to internal processes, systems and data "comparable to what internal risk assessment teams have" and naming metrics such as its R&D Automation Index among the things embedded evaluators would monitor (Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026)).

Elements

Anthropic's statements set out the terms it offers:

Anthropic states that "embedded evaluation is new, and many of the details about how it will operate are still being worked out," that there are "as yet, no standards for what information embedded evaluators should have access to, or how they should report what they find," and that there is "no settled system for funding independent evaluation." It also states that embedded evaluators "do not reduce our accountability, but help to make it more verifiable" (Source: anthropic.com).

First arrangement

On September 18, 2026 Anthropic announced a partnership with Accenture on embedded evaluation, led by Faculty, which Anthropic describes as Accenture's specialist AI business. The work covers evaluating and red-teaming models, conducting alignment assessments and testing model safeguards, and Anthropic says Accenture's experience of enterprise deployment will inform its evaluation. Anthropic and Accenture each expect to invest at least $1 billion in building capacity in the area over the next five years. Anthropic funds Accenture's work directly, citing the absence of pooled or government funding; it is separately "in dialogue with METR and other nonprofit evaluators to pilot elements of embedded evaluation using their own funding," and said it would announce further evaluators "in the coming weeks" (Source: anthropic.com).

Faculty is a company Accenture acquired in January 2026 to serve as its AI division. TechCrunch reported that the choice surprised observers, since discussion of embedded evaluators after Amodei's essay had centered on safety research organizations such as METR, Redwood Research and Apollo Research, and that Accenture's shares rose 8% in after-hours trading on the announcement. Anthropic pointed to Accenture's practical experience deploying AI for large corporations and government agencies; TechCrunch added that, as a large public company predating the current AI industry, Accenture is more functionally independent of Anthropic than organizations within the ecosystem around the laboratory (Source: techcrunch.com).

Reception

Responses to the proposal divided along lines that recur in the wider pacing debate (Coordinated slowdown proposals compared):

• Endorsement from a pacing critic. Clément Delangue, chief executive of Hugging Face, wrote on X that "it's not time to slow down but to accelerate" while endorsing Amodei's proposal to embed "independent evaluators" inside AI companies (Source: reuters.com).
• Capture objection. Kristian Kersting, professor of AI at the Technical University of Darmstadt, said the proposals amounted to "regulatory capture": "If a handful of large companies decide for themselves what constitutes 'dangerous', select the appropriate inspectors and demand an exemption from competition law, that is no longer genuine regulation" (Source: reuters.com). See Regulatory Capture in AI Policy.
• Preference for government review. Announcing the American AI Security Act on September 18, 2026, Rep. Josh Gottheimer rejected reliance on "third-party testers or auditors," saying there are few of them, that they have limited people, compute and access to the labs, lack national-security clearances, and "rely, in part, on the labs themselves for funding. In short, it's like having the fox guard the hen house" (Source: gottheimer.house.gov). The bill would instead give the National Security Agency pre-deployment access (American AI Security Act (Gottheimer–Lawler)).
• Advocates' legislative asks. The Washington Post's account of the September 16, 2026 coalition letter led by Americans for Responsible Innovation reported that AI safety advocates want evaluators embedded in the laboratories; the letter itself asks that evaluations and auditing be "mandatory and conducted independently" without specifying embedding (Senate coalition letter on redlines for AI safety legislation (Americans for Responsible Innovation, September 2026)).
• Evaluator independence. In September 2026 some White House officials and AI executives described the existing ecosystem of safety and benchmarking groups as too closely tied to the leading companies; METR states that it does not take funding from frontier laboratories (see METR).

An August 2026 assessment by Guidelight AI Standards found that companies' plans for third-party assessment were unclear on frequency and depth of access, and that only one of the five companies it assessed appeared to have granted hands-on access to third-party assessors (see Guidelight AI Standards).

Relation to other verification proposals

Embedded evaluation is one of several mechanisms proposed in 2026 for verifying frontier developers' conduct. It differs from licensed independent verification organizations (the model in California's SB 813 and the FRONTIER Act) in being arranged by the company rather than accredited by a government, though Amodei calls for governments to require and formalize it; from government pre-deployment review such as the Commerce Department's pre-release vetting and the NSA review proposed in the American AI Security Act; and from one-off external audits in that the evaluators have standing access to training processes rather than to a finished model. California Governor Gavin Newsom's Executive Order N-9-26 of September 18, 2026 directs the Government Operations Agency to recommend, by November 16, 2026, whether to require all large frontier developers to embed designated independent verification organizations onsite for periodic audits and evaluations, which would combine the two models (California Executive Order N-9-26 (independent verification and frontier AI safety)).

Open questions

• Which further evaluators Anthropic names, whether METR or another nonprofit agrees to a pilot on its own funding, and whether any other frontier developer adopts the practice.
• What standards govern evaluators' access and reporting, which Anthropic states do not yet exist.
• How findings are published in practice under the redaction terms Anthropic has stated.

Relationships

• depends-on: We Must Pace the Frontier (Amodei, September 2026) — the essay that introduced the proposal
• instance-of: Alignment Auditing — a standing form of external alignment assessment
• related: Independent Verification Organizations (IVOs) — the government-accredited alternative
• related: AI Pre-Release Vetting — government testing before deployment
• related: Anthropic, METR, Coordinated slowdown proposals compared
• related: Regulatory Capture in AI Policy — Kersting's objection that developer-selected inspectors are not genuine regulation