Anthropic CEO Responds to Safety Criticism in Long Post: The "Race to the Top" Strategy and Its Billion-Dollar Costs

Anthropic CEO Dario Amodei took to X with a rare long-form response, directly addressing sharp criticism from AI safety investigator @_NathanCalvin. This exchange is not just about one company's reputation—it touches on the fundamental tension between safety and commerce at frontier AI labs.

The Core Dispute: Safety Commitments or Commercial Disguise?

The incident began with an investigative report by Nathan Calvin, listing behaviors in which Anthropic allegedly departed from its safety commitments over the past year. Amodei's response opens with an open stance—he praised Calvin's investigative work as coming "from a constructive and honest place," and acknowledged that "a company of Anthropic's scale would have a problem if it weren't being harshly criticized every day." He even admitted that "some of the examples he listed are indeed Anthropic's mistakes."

But the tone quickly shifted. Amodei made clear he "disagrees with the main thrust of the post," arguing that critics overlook a fundamental context: the "race to the top" theory Anthropic has adhered to for years. The core logic of this strategy is that only by being at the very frontier of technology can a company raise industry-wide safety standards by demonstrating responsible practices. According to Amodei's observation, this strategy's effectiveness "is much stronger with frontier companies."

"Race to the top" is a strategic hypothesis in the field of AI governance, whose core logic borrows from the "California effect" in environmental regulation: when a technology leader voluntarily adopts safety standards higher than legally required, this demonstration forces peers to follow through market competition pressure, thereby raising the industry bar overall without relying on government mandates. The theory depends on two premises: that frontier companies indeed have significant market or reputational advantages to "bet," and that later entrants are compelled to emulate in order to compete for the same customers and talent. Critics point out that this strategy lacks verifiable causal evidence and can easily be used by companies as a rhetorical framework to package commercial behavior as public goodwill.

Nine "Costly Safety Decisions"

To substantiate that the "race to the top" is not empty rhetoric, Amodei laid out a detailed list of safety measures Anthropic took over the past year that came at significant commercial cost:

On policy advocacy, Anthropic played a key role in state-level AI legislation such as SB53, at a time when AI regulation was politically highly unpopular, and the company even advocated for stricter provisions than the version ultimately passed. Amodei himself also called in public articles for giving the government the power to block model releases, for which he "came under fierce attack."

In the biosecurity domain, after observing concerning results in biological uplift trials and intercepting potential bioweapon misuse attempts, Anthropic deployed extremely strict biosecurity classifiers. The cost was a large number of false positives—biologists joked that its restrictions were too strict, and competitors even used "more lenient biological capabilities" as a selling point. Amodei candidly admitted that as a biologist he was "particularly frustrated" by this, but insisted "this is the right thing to do."

On model release pacing, the Mythos-class model was ready as early as February but was only fully opened in June, during which time it was provided only to a small number of partners for cyber defense. Amodei acknowledged this "sacrificed a large part of several months of model leadership advantage," but in exchange many enterprises hardened their infrastructure before cyber-capable models proliferated.

At the geopolitical level, Anthropic was "the first and only" frontier company to restrict sales of AI services to Chinese-controlled companies (including entities operating through third countries), and has continued to advocate for chip export controls against China—even though this damaged relationships with key industry partners.

In addition, the company announced this week that it would contribute 5% of its shares to a public interest fund, and has continued to publish numerous safety reports and detailed model cards. Amodei emphasized that these decisions "were not driven by profit-seeking."

Admitting Mistakes but Rejecting the Overall Narrative

Amodei's response was not a blanket defense. He categorized items 7, 8, and 9 in the criticism list as falling under "rhetoric or policy communication that was not prudent enough," admitting "it could have been handled better," and taking responsibility himself: "I'm certainly the CEO, and I'm accountable." Regarding item 5—modifying the Responsible Scaling Policy (RSP)—he acknowledged this was "the most substantive issue," but argued that the old RSP's rigid requirements sometimes forced the team to do a great deal of work that "in hindsight was not important for safety," while the new version emphasizes transparency and external independent review, which instead lays the foundation for the "race to the top."

As for item 6, questioning the coexistence of FCF and RSP, Amodei promised to "look into it further," admitting "we may not have weighed the various considerations in the right way."

But he firmly rejected the overall framing of the criticism. Regarding items 1 through 4, Amodei's response was direct: these are merely examples of "what competing at the frontier means," "transparent and ordinary" commercial behavior, fully consistent with the "race to the top" strategy. He refuted allegations of blocking biological research to favor its own products, calling them "extremely implausible"—the very existence of the life sciences verification program proves Anthropic has no intention of being hostile to biologists.

The Responsible Scaling Policy (RSP) is a tiered safety framework proposed by Anthropic in 2023, whose approach resembles the "safety case" system in aviation or the nuclear industry: it sets several "AI Safety Levels" (ASL) for model capabilities, and each time a new level is reached, the corresponding safety, security, and deployment safeguard requirements must first be met, otherwise expansion may not continue. The old RSP was criticized for adopting relatively rigid trigger conditions, which were seen as potentially over-constraining R&D when capability assessment standards were unclear; the new version shifts to a governance model that emphasizes transparency, external independent review, and continuous iteration. This modification is regarded as "the most substantive" because it directly determines the conditions under which the company allows itself to release stronger models.

Negativity Bias and a More Fundamental Disagreement

Amodei offered a psychological observation: "There is a significant negativity bias that distorts the kind of assessments you do." He argued that any decision can be retroactively interpreted as a conspiracy, but "most don't withstand careful scrutiny."

However, he also acknowledged the legitimacy of the disagreement: if critics believe that loss-of-control risk is far more important than misuse risk, then the balance of assessment will naturally tilt—but "this is not Anthropic's view," as the company regards both as equally serious threats.

At the end, Amodei revealed that he will further elaborate in the near future on how the company practices its values amid rapid change. The value of this public debate may lie precisely here: when a frontier lab is willing to respond item by item to safety questions on a public platform, it itself provides a testable sample for the "race to the top."

And Nathan Calvin's investigation and Amodei's rebuttal together outline the core dilemma of AI safety governance—in the current absence of government regulation, who defines "responsible"? And where is the boundary of corporate self-regulation? This dialogue does not provide a final answer, but at least it makes the debate more concrete.

Background Notes

Biological uplift trials refer to assessing whether AI models can significantly enhance the capabilities of non-experts in tasks related to biological R&D, such as designing pathogens, optimizing experimental procedures, or acquiring dangerous materials. Such assessments are a core part of biosecurity red-teaming at frontier labs, because if a model can substantially lower the technical threshold for bioweapons, it constitutes an "uplift risk." Correspondingly, biosecurity classifiers are content filtering systems deployed at the input and output ends of models to identify and intercept requests related to dangerous biological R&D. The stricter the classifier, the higher the false positive rate, which will erroneously affect legitimate queries from biological researchers—this is the background for biologists complaining about overly strict restrictions and competitors using "more lenient" as a selling point.

One-line Summary

Amodei responds at length to safety criticism, defending nine costly safety decisions with the "race to the top" theory while admitting some mistakes.

Paragraph Summary

Anthropic CEO Dario Amodei published a long post on X responding to criticism from AI safety investigator Nathan Calvin. He admitted that some communications and policy documents did contain mistakes and took responsibility, but disagreed with the overall thrust of the criticism. His core argument is the "race to the top" strategy: only by being at the frontier can one raise industry-wide safety standards by demonstrating responsible practices. To this end, he listed nine commercially costly decisions, including promoting strict legislation such as SB53, deploying biosecurity classifiers with high false positive rates, delaying the release of the Mythos-class model, restricting AI services to China, and donating 5% of shares. He dismissed the first four allegations as transparent and ordinary commercial behavior, explained the RSP modification as a shift from rigid requirements to transparency and independent review, and pointed out that critics have a negativity bias. The article concludes by highlighting the fundamental dilemma of the boundary of corporate self-regulation and responsibility attribution amid regulatory gaps.