‘Not a Safe AI Company’: The Coalition That Chose Anthropic’s IPO Eve to Say So

The Library Media — Science, Technology & Health · AI News & Intelligence Desk

‘Not a Safe AI Company’: The Coalition That Chose Anthropic’s IPO Eve to Say So

By Gideon Black

Share

Published November 19, 2026 · 6:22 PM PT

The accusation arrived at a consequential moment. On November 18, hours before Anthropic priced its initial public offering, the Center for AI Safety joined a coalition of researchers, academics, technologists, advocates and public figures in declaring that “Anthropic is not a safe AI company.” The statement was only three sentences long. Its implications were considerably larger: it challenged the company’s safety credentials, questioned the consistency of its public commitments and placed those questions before investors at the point when they were being asked to value the business.

That timing deserves scrutiny, but not a reflexive verdict. A safety claim made on the eve of an offering is simultaneously a moral intervention and a market act. It can draw attention to information investors should consider before committing capital. It can also exert pressure on a company at the moment when reputational damage may be most expensive. Those possibilities are not mutually exclusive. The available evidence establishes grounds for asking hard questions about Anthropic’s safety commitments. It does not establish every allegation in the coalition’s statement, much less settle the coalition’s motives.

The full statement, published by CAIS on November 18, reads:

“Anthropic warns that recursively self-improving AI could concentrate power and escape human control, yet races to build it first. The company uses safety branding to court investors, employees, and the public—even as it has quietly weakened its safety pledges and lobbied for self-serving regulation. Anthropic is not a safe AI company, and we must reject its reckless attempt to amass power in the name of the greater good.”

The coalition’s composition is part of its argument. CAIS describes the signatories as safety researchers, open-source advocates, ethicists and accelerationists who “disagree on almost everything.” The names include Harvard scholars Steven Pinker and Oliver Hart, UCLA’s Judea Pearl, Harvard Law figures Laurence Tribe and Lawrence Lessig, Hugging Face co-founder Thomas Wolf, Rumman Chowdhury, Tristan Harris, Max Tegmark, George Hotz, Stuart Russell, Gary Marcus and CAIS executive director Dan Hendrycks. The list also includes Fields Medalists Peter Scholze, James Maynard and David Mumford, as well as xAI co-founder Igor Babuschkin and public commentator Stephen Fry.

The count requires care. Contemporary coverage on release day reported figures between 89 and 91, while the live list continued to change. A desk review on November 19 found more than 120 signatories, with the list still growing. The meaningful point is not a falsely precise total. It is that the statement attracted a broad and expanding group, including people whose views about AI development and its risks are not interchangeable.

Hendrycks made the coalition’s central claim explicit: “People in this coalition disagree on almost everything, yet we all recognize the danger of a company that convinces itself its own dominance is the greater good.” He also argued: “You cannot credibly claim to be safe while weakening your own previous safety commitments and asking the public for tens of billions of dollars to build recursively self-improving AI.”

The coalition’s case has three principal components. The first concerns recursive self-improvement: it says Anthropic warns that such systems could concentrate power and escape human control while presenting the pursuit of recursive self-improvement to public-market investors as a competitive advantage. The second concerns the company’s identity and marketing. CAIS argues that Anthropic uses its positioning as an “AI safety and research company” to attract employees and investors while attempting to automate AI research and development. The third concerns policy: CAIS says Anthropic weakened its Responsible Scaling Policy in February 2026 and lobbied against or sought to weaken several state AI measures, while privately advocating for federal pre-emption of state AI laws.

These claims should not be treated as a single evidentiary package. They have different levels of verifiability, and the distinction matters.

The strongest, most readily checkable part concerns the Responsible Scaling Policy, or RSP. Anthropic’s own policy page confirms that Version 3.0 took effect on February 24, 2026, as a comprehensive rewrite that added Frontier Safety Roadmaps and Risk Reports. The crucial question is what happened to the company’s ability or obligation to pause development when risks become serious.

Anthropic’s own note accompanying the subsequent Version 3.1 clarification, dated April 2, says that even when the RSP does not require it, the company “remains free to take measures such as pausing the development of our AI systems in any circumstances in which we deem them appropriate.” The note adds that this was also true under Version 3, but that Version 3.1 states it more clearly.

That language supports a specific part of the coalition’s criticism. Under Version 3, pausing development was discretionary rather than an RSP requirement. In that limited documentary sense, the coalition is right that the earlier unilateral pause commitment was abandoned. This is not the same as saying Anthropic has lost the ability to pause, has promised never to pause, or is legally prohibited from doing so. The company’s own clarification says the opposite about its discretion: it remains free to pause when it considers that appropriate.

But the distinction does not make the change immaterial. A voluntary safety policy is not a contract with the public in every respect, and a company can revise its own framework. Yet when a company builds part of its identity around safety commitments, the terms of those commitments become relevant evidence. There is a meaningful difference between retaining the power to pause and maintaining a policy that commits the company to pause under specified circumstances. Discretion belongs to management; a pledge establishes an expectation against which management can be judged.

Investors do not need to conclude that the revised policy is inadequate to recognize that the commitment changed. Nor should they assume that the addition of roadmaps and risk reports automatically compensates for the removal of a unilateral pause requirement. Those are different policy mechanisms. Their relative strength depends on details and implementation that this fact pack does not establish.

The coalition’s other allegations are less settled. Its descriptions of Anthropic’s lobbying concerning SB 1047, the RAISE Act and the AEGIS Act, and its account of private advocacy for federal pre-emption, are allegations advanced by CAIS. They are not findings by a regulator or court in the evidence available here. The same caution applies to the claim that Anthropic’s safety branding is being used to court investors and employees while concealing a conflicting motive. The company’s public identity and its commercial ambitions can be examined. But an assertion about why it presents itself that way is an assertion about intent. The statement supplies the coalition’s interpretation, not independent proof of motive.

That is where the wording moves beyond the documentary record. “Not a safe AI company” is a categorical conclusion. The evidence presented here establishes a revision to a voluntary commitment and a set of disputed policy and lobbying claims. It does not establish that Anthropic’s current RSP fails a legal or voluntary standard. It does not demonstrate that the company’s systems are unsafe, or that its current safety procedures are ineffective. A company can have serious safety weaknesses without a regulator having made a finding; equally, a critic cannot turn an allegation into a finding by stating it with confidence.

The breadth of the coalition cuts both ways, too. Its signatories bring different kinds of expertise and public standing. Their agreement on a warning may indicate that the concern crosses familiar institutional and ideological boundaries. But the coalition itself says some participants are accelerationists. Their reasons for opposing Anthropic may differ sharply from those of researchers concerned about catastrophic risk or the weakening of safety commitments. That diversity is part of the coalition’s political strength, not proof that every signatory endorses every premise in the same way.

The timing is the unavoidable question. The statement appeared hours before Anthropic priced its IPO. An offering of this scale makes safety representations more than a matter of corporate reputation. Investors are being asked to assess a company whose stated mission, risk controls and development ambitions may all bear on its long-term value. If a company’s public safety identity is part of the case it makes to employees, customers and capital providers, a change in the underlying commitments is relevant at precisely the moment investors are deciding what they believe that identity is worth.

That is the strongest argument for speaking then rather than later. Waiting until after pricing could leave investors without a prominent challenge to claims that the coalition considers consequential. Public-market scrutiny should not be suspended to protect a company from an inconvenient question during an offering. If the concern is real, the calendar does not make it less real.

But the counterargument is equally serious. A statement timed to the eve of pricing can affect sentiment when the target is especially exposed. It can frame a company’s public identity in the most damaging possible terms, before the disputed claims have been independently resolved. The coalition’s decision to use a categorical headline and a highly charged conclusion, rather than a narrower statement about the policy revision, makes the intervention more consequential and more difficult to separate from its potential market effect.

There is no evidence in this pack that establishes an intention to move Anthropic’s share price, nor is there evidence that the timing was chosen for that purpose. There is also no basis for pretending the timing is neutral. The same moment that makes the warning potentially useful to investors makes it potentially costly to the company. That is why the statement should be assessed on its evidence, not accepted or dismissed because of its publication date.

Anthropic had not provided a response on the record as of November 19. Its newsroom carried no response to the statement, and the CAIS release itself contained none. That absence should be reported plainly, but it should not be converted into an admission or treated as confirmation of the coalition’s account.

The most defensible conclusion is narrower than the coalition’s headline and more consequential than a routine dispute over policy language. Anthropic revised its RSP in a way that left pausing development discretionary rather than a requirement of the policy. That is a real change, documented by the company itself. It is fair to ask what investors were told about the significance of that change, how it fits with the company’s safety claims, and whether its other policy positions are consistent with the public interest it invokes. It is not fair to present contested lobbying characterisations or inferred motives as established facts.

A safety pledge is not proof of safety. Nor is the revision of a pledge proof that a company is unsafe. But when a company asks the public to finance the development of increasingly capable systems, the difference between a commitment and a discretion deserves attention. The coalition chose a moment when that distinction could not easily be ignored. Whether that was responsible warning, strategic pressure, or both remains an open question. The documents give investors a reason to ask it. They do not answer it for them.

Gideon Black
Chief Skeptic, The Library Media

From The Library — the estate of our family.

Follow Us

X Facebook Instagram Threads Mastodon Bluesky

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top