Microsoft CEO Satya Nadella has publicly endorsed a controversial new model where independent watchdog groups would embed evaluators directly inside AI companies with employee-level access to assess safety risks. This marks a significant shift in how the tech industry approaches external oversight of artificial intelligence systems, even as internal documents reveal tensions between what these companies say publicly and what their executives acknowledge privately about AI's impact on existing businesses.

What Are AI Labs Committing to With External Evaluators?

Both OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei recently pledged to welcome third-party evaluators inside their organizations. However, the AI Evaluator Forum, a coalition of watchdog groups, has now laid out specific conditions these companies must meet for the arrangement to work effectively. The letter, published on Friday and backed by dozens of AI researchers including Geoffrey Hinton, widely known as the "Godfather of AI," and Stuart Russell, sets a demanding standard for what genuine oversight looks like.

Nadella's public support signals that Microsoft is willing to subject itself to this level of scrutiny. This comes at a particularly sensitive moment, as internal testimony and documents from both Microsoft and OpenAI executives have surfaced in a major copyright lawsuit filed by The New York Times, revealing concerns these companies had about their AI systems potentially replacing publisher websites.

What Specific Conditions Must AI Labs Meet for Embedded Evaluators?

The watchdog coalition has outlined several non-negotiable requirements for evaluators to operate effectively inside AI companies:

• Scientific Independence: Evaluators must maintain complete objectivity and transparency, with immunity from coercion by the companies being assessed.
• Board-Level Access: Evaluators must have unhampered access to company boards and other governing bodies, ensuring they can reach decision-makers directly.
• Publication Rights: Evaluators must have freedom to publish their findings without retaliation or legal threats from the companies they evaluate.
• System Access: Evaluators must have the same access to systems, data, tools, and physical spaces as internal company assessors.
• Diverse Expertise: Each lab must bring in evaluators with different perspectives and specialties rather than a single evaluator or homogeneous team.

The coalition behind this effort includes METR, a research nonprofit that recently conducted security research on OpenAI's use of Hugging Face, and the AI Verification and Evaluation Research Institute founded by Miles Brundage, a former OpenAI researcher.

Why Are Internal Statements Creating Legal Problems for Microsoft and OpenAI?

The timing of Nadella's endorsement of external oversight is complicated by newly unsealed court documents that expose internal warnings from both companies about AI's market impact. In deposition testimony cited in The New York Times' copyright case, Nadella himself acknowledged that chatbot conversations had "substituted" for publisher websites.

Microsoft director of applied science Brent Hecht described large-scale data scraping for AI training as "an astonishing theft of unprecedented proportions" and possibly "the largest theft of labor in human history," according to the plaintiffs' brief. Meanwhile, OpenAI's head of ChatGPT, Nick Turley, wrote internally that the company's products were "largely substitutive" and that publishers faced an "existential threat".

"Our products are largely substitutive, period," Turley wrote, warning that publishers faced an "existential threat."

Nick Turley, Head of ChatGPT at OpenAI

Microsoft has pushed back on these characterizations, with spokesman Alex Haurek stating that Hecht's writings reflected "one employee's individual perspective" and were not legal analysis representing the company's official views. Haurek also reframed Nadella's testimony as an observation about how people consume information rather than a legal conclusion about copyright.

How Could Embedded Evaluators Change AI Development?

If these conditions are met, the embedded evaluator model could represent a meaningful shift in how AI safety gets monitored. Rather than relying on companies to self-report risks or waiting for government regulators to catch problems, independent experts would have real-time visibility into how these systems are built and deployed. The presence of credible external oversight could also help rebuild public trust in AI companies at a moment when concerns about safety, copyright, and market displacement are intensifying.

The fact that major AI labs are willing to consider this arrangement suggests they recognize the legitimacy of external concerns. However, the detailed conditions laid out by the watchdog coalition make clear that a genuine commitment to oversight will require more than symbolic gestures. Companies will need to grant real access, protect evaluators from retaliation, and allow findings to be published even if the results are unflattering.

The case is still in its early stages, with both OpenAI and Microsoft filing their summary-judgment memoranda on September 4, 2026. A ruling on the copyright claims could come in the coming months, potentially clarifying whether the companies' internal acknowledgments of substitution effects constitute evidence of copyright infringement.