OMITTED

What the news leaves out.

← Omitted front page

Experts Call for Independent AI Safety Evaluators

5 sources · updated 2026-09-20
Left 50% Center 50% Right 0%
1 left · 1 center · 0 right

What happened

After incidents in which their AI agents independently accessed other organizations—including an OpenAI model involved in a Hugging Face security incident—OpenAI and Anthropic publicly committed to embedding third-party evaluators with employee-like access to inspect models, development processes, and internal systems. In response, the AI Evaluator Forum released a Friday letter signed by more than 100 experts, including Geoffrey Hinton and Stuart Russell, listing 'minimum conditions' for those evaluations to be credible: independence from the labs (no ownership, governance, or pay contingent on findings), access equivalent to privileged employees, protection from retaliation, unfiltered communication with company boards, and public release of findings subject only to limited redactions. OpenAI CEO Sam Altman, Microsoft CEO Satya Nadella, and Elon Musk publicly backed the embedded-evaluators concept. Anthropic then announced a short-term, company-funded non-exclusive deal with Accenture to embed safety evaluators and said it was in dialogue with METR and other nonprofit evaluators using those groups' own funding. The exchange unfolded as members of Congress debated AI legislation and President Donald Trump called warnings that AI could take over the world a 'HOAX.'
Sourcing check · Fully on the record · 5 articles

4 of 4 articles cite a named source for the story. It was on the record from the first report.

What these terms mean and how they're counted
On the record · names a source 4 of 4
On the record means the article cites a source by name for the story's main claim: a person, or an institution speaking officially, such as a company statement or a named agency's spokesperson. The opposite is an unnamed source. The count is how many of the counted articles cite at least one such source; the names are everyone named across them, each once however the articles spell it. Here: AI Evaluator Forum, Conrad Stosz, Geoffrey Hinton, Rep. Troy Downing, Rep. Brad Knott and Rep. Jimmy Patronis
Primary evidence · cites a document or data 2 of 4
Primary evidence means the article cites a record you could check yourself: a court filing or ruling, a dataset or report, a recording or transcript, an official statement or press release, market data. A quote is not a document. The count is how many articles cite one. Here: letter, on-record letter
Credits another outlet 0 of 4
The article says the story comes from another outlet's reporting (“first reported by Reuters”). Those articles relay the story; they are not independent confirmation. Here: Each article presents its own reporting.
Unnamed sources 0 of 4
The main claim is attributed to people who are not identified: “people familiar with the matter”, “officials said”. The count is how many articles do this.
How it developed
When the first report landed and whether it named anyone; when the first named source and the first document appeared. Here: First report Sep 18, 13:00 UTC by Business Insider, naming its source. First document in the first report, from Business Insider (letter).
Not counted 1
Opinion, analysis and video pieces, and articles too short to assess. Every count above is out of the counted articles. An article republished under a new headline is one article.
This is about who is cited, not whether the story is true. Each source below says what it names. How we count
Omitted — what each side leaves out

Unpacked

Left-leaning coverage presents the watchdog letter as a demand for true independence and never reports the personal and funding ties between Anthropic and the evaluator nonprofits named in that coverage. METR appears there as a nonprofit signatory, including in CNBC's report, with no mention of Effective Altruism funders or family links to Anthropic's leadership. Right-leaning coverage never reports that more than 100 experts signed a public letter spelling out independence rules for embedded evaluators, even as the New York Post details those ties. Left-leaning coverage opens on scientific backers of the letter. Right-leaning coverage opens on Republican suspicion of the companies' motives. Fox News leads with lawmakers saying there is "something else going on," and does not describe the Friday letter at all. Would the evaluator groups already in talks with Anthropic even pass the letter's own tests for independence?
Bottom line

CNBC never mentioned that Redwood Research had an original board member married to Anthropic's CEO's sister. The New York Post never mentioned that more than 100 experts, including Geoffrey Hinton, had just set independence rules for those watchdogs.

The receipts — 4 facts one side's coverage lacks
  • Redwood Research's original board included Holden Karnofsky, an Anthropic employee married to CEO Dario Amodei's sister. contextin New York Post · absent from left-leaning coverage
  • Coefficient Giving gave Redwood Research a $36 million grant last November. numberin New York Post · absent from left-leaning coverage
  • Over 100 AI experts signed a public letter setting minimum conditions for embedded evaluators. numberin CNBC · absent from right-leaning coverage
  • The letter says embedded evaluation organizations should not be owned or governed by frontier AI companies and should not accept payment contingent on findings. official statementin CNBC, Business Insider · absent from right-leaning coverage
The Left View
Left-leaning coverage presents the Friday letter as experts refusing to treat the labs' own oversight pledge as sufficient unless it is structurally independent. Conrad Stosz of the AI Evaluator Forum told CNBC the aim is 'basic principles' so independent oversight can be 'a meaningful tool for managing AI risk broadly,' and that the companies' credibility is at stake if they ignore the conditions. Signatory Vinh Nguyen, a former NSA chief AI officer, argued that when a few labs control capabilities that can endanger cybersecurity, critical infrastructure, and the systems national security and the economy run on, 'the government and the public cannot be dependent on those labs' own account of what's secure and safe.' Coverage stresses access to unreleased internal systems, multiple evaluators with differing expertise (including biological and loss-of-control risks), and that embedded evaluation complements rather than replaces internal safety work. Princeton computer scientist Arvind Narayanan told CBS News that firms need external audits because 'We shouldn't have to take their word for it.'
The Right View
Right-leaning sources treat the watchdog push as a motive problem and a personnel problem, not a neutral safety upgrade. Fox News quoted Republicans such as Rep. Troy Downing saying there is a 'false sense of urgency' and that 'any one of those people crying gloom and doom has the possibility of stopping it in their own shops, so there's something else going on there,' and Rep. Brad Knott saying he is 'very skeptical of people in big business who want to leverage big government to suppress competition.' The New York Post described a 'phony, incestuous web of woke AI "watchdogs,"' reporting that Redwood Research's original board included Holden Karnofsky—an Anthropic employee married to CEO Dario Amodei's sister—and that Coefficient Giving, which Karnofsky co-founded, has granted tens of millions to Redwood and funded METR's predecessor, with further overlap through Paul Christiano, Effective Altruism, and Anthropic investors; Perry Metzger of Alliance for the Future said 'None of these people are independent, none of these people are arm's length.' The Post also quoted Democrat Rep. Josh Gottheimer saying that when developers 'cheer on the framework meant to hold these companies accountable, it should raise red flags, not confidence.' President Trump called AI-takeover warnings a 'HOAX,' House Majority Leader Steve Scalise asked 'THESE are the people we're trusting to beat China in AI?,' and Pentagon tech official Emil Michael accused 'death-cult-like philosophies' of 'a coordinated campaign to scare people to make irrational decisions that benefit some of these incumbents.' Some Republicans in the Fox piece still wanted limited human controls while agreeing with Trump on the danger of over-regulation.
Our Take (balanced)
The left's strongest argument is that frontier labs should not be the sole narrators of their own risk: the Friday letter's independence, access, and anti-retaliation conditions, backed by more than 100 technical signatories and Nguyen's national-security formulation, are offered as the minimum that turns the companies' own evaluator pledge into an actual check. The right's strongest argument is that leading nonprofit evaluator groups are already bound to those labs by family, employment, board, and grant ties—the Karnofsky–Amodei–Coefficient–Redwood–METR cluster being the most documented—so 'independence' is not established on the facts, and doomsday framing may help incumbents shape rules (a reading Republican sources tie to regulatory capture and the U.S.–China AI race). The central unresolved tension is whether embedded third-party evaluation can be made genuinely independent of both the labs and a small overlapping safety-funding community, or whether the watchdog push is, in practice, self-policing by an intertwined set of firms and allies.

5 sources

More stories

The week's bottom lines, in your inbox

One email a week: the five stories that mattered and what they actually mean. Free.