Amodei says AI industry 'lied' about risks and opens models

Anthropic CEO Dario Amodei told "CBS Sunday Morning" on September 13 that "for too long the industry has lied to people about the fact that this technology has risks." In the same interview, he announced that Anthropic will allow independent evaluators permanent access, "as if they were employees," to its models. Amodei compared the mechanism to a "sanitary inspector" who can verify whether companies comply with their own security commitments.

What Exactly Was Announced

Amodei proposed that competitors adopt the same protocol: continuous access for third parties to frontier models, with the right to inspect behaviors that today only the company itself can see. This is the first time that one of the three major labs has publicly offered a continuous audit pipeline as opposed to one-off pre-launch assessments. Amodei also stated support for federal regulation and acknowledged that independent evaluators, on their own, "are hardly enough."

The statement came as a direct response to Jacob Coxon, a former Anthropic researcher who left the company days earlier warning about existential risks of self-improving superintelligence. Amodei admitted that he "had not fully assessed what it would be like when progress came so quickly." This statement contrasts with his position from January, when Amodei described AI advancement as "exponential" and likened the moment to a "teenage phase" of the technology.

The Chinese Dilemma, in the Parallel Talk with CNBC

On the same day, Amodei told CNBC that China represents the "hardest dilemma" for his deceleration proposal. The logic is simple: if American labs moderate their pace while Chinese firms continue at full throttle, the gain in safety comes with a loss of strategic advantage. Amodei's response was that it is still worth holding back, nonetheless. "I’m not saying the problem is easy," he stated. "I’m saying the alternative is worse."

This argument contradicts the stance that David Sacks, the White House AI czar, has been publicly advocating. Sacks rejected the idea of any pact between American labs this week and highlighted the risk of "unilateral disarmament." The distance between the two positions is the reason conversations between the Trump administration and companies have stalled since July, as reported by The Information.

Global Reading: Three Geographies, Three Different Problems

For the United States, Amodei’s proposal shifts the cost of safety internally within the industry and away from Congress. If three labs accept external evaluators with access equivalent to that of employees, the battleground shifts from defining risk (currently done by the company producing the model) to governing the metric. A direct consequence: enterprise contracts requiring "AI safety attestation" gain an auditable requirement.

For the European Union, the announcement complicates the interpretation of the AI Act. Since August 2, the Commission can fine providers of general-purpose models up to 3% of global revenue. If Anthropic voluntarily accepts a regime resembling the conduct code audit described in Article 55 of the AI Act, the European AI Office loses part of the argument that it needs to enforce access via regulatory processes. At the same time, it gains a technical baseline for comparing OpenAI and Google.

For China, the proposal is rhetorical: no Chinese lab will open its models to a Western evaluator. What changes is the political cost of stating this. If Anthropic institutionalizes third-party access, Zhipu, Moonshot, and DeepSeek will operate with an implicit "not evaluated" label when entering RFPs from American or European multinationals. The practical effect is smaller than the headline suggests but is real in corporate purchasing decisions.

What Weakens Amodei’s Argument

Amodei himself stated that continuous auditing "is hardly enough." This phrase lends credibility but limits expectations. The discussion about who pays these evaluators, who accredits them, and who is responsible for failure to notify is still open. A softer version of the arrangement could become a pro forma committee like the one Meta maintained until 2023.

The other gap is temporal. Amodei proposes slowing down at a moment when Anthropic has just locked in 14.8 GW of computing capacity and reported annualized revenue exceeding $30 billion, outpacing OpenAI in enterprise revenue per seat across various verticals. Rivalry and security discourse coexist, but the side that is already ahead benefits more from a coordinated pause. For the C-level buyer, the test remains the same: the audit SLA stipulated in the contract, not the Sunday interview.