In June 2026, Anthropic CEO Dario Amodei published a public article titled "Policy on the AI Exponential." In this lengthy article, he proposed a specific suggestion: the US government should establish a regulatory body similar to the Federal Aviation Administration (FAA) to conduct mandatory third-party testing on all cutting-edge AI models. The testing should cover four dimensions: cybersecurity, biological weapons, runaway risks, and automated development. The government would have the right to block the release of models that fail the tests.

In the same article, Amodei also wrote a statement that is quite rare among Silicon Valley CEOs: AI may lead to “significant and persistent unemployment”, which may be due to “the inherent property of technology to widely replicate human cognition”, and traditional economic adaptation mechanisms may be overwhelmed by the speed of technology.

For a CEO who has long championed a "Responsible Scaling Policy" (RSP), these statements are not casual remarks. For the past three years, his public stance has consistently been one of transparent corporate disclosure and waiting for risks to become more concrete before enacting legislation. Now, he is not only actively calling for government intervention but has also announced that Anthropic will invest approximately $350 million to push for the implementation of this regulatory framework.

The shift from "corporate self-regulation and waiting for government intervention" to "actively contributing funds and effort to promote legislation" is itself a signal. How has Amodei's stance been revised over the past few years? What factors forced him to move from "internal constraints" to "external demands"?

Anthropic Responsible Scaling Policy (RSP) v3.0 Official Illustration

An optimistic letter, a warning, a cry for help

Amodei's diagnosis in "Policy on the AI Exponential" is that AI technology is developing exponentially, significantly exceeding the response speed of existing policy-making processes. He mentions in the article that Anthropic's Claude Mythos Preview model, released in April of this year, has demonstrated national-level vulnerability discovery capabilities in the field of cybersecurity. According to the system card officially released by Anthropic, the model's performance in tasks such as zero-day vulnerability discovery has reached the threshold requiring reporting to national security agencies.

The proposed solution is to establish a mandatory regulatory mechanism similar to the FAA. Pre-release models would require third-party testing across four dimensions, and the government would have the right to prevent the release of models that fail the tests. The radical aspect of this proposal is that it doesn't rely on industry self-regulation or voluntary commitments, but rather on legally binding pre-approval.

Amodei explicitly acknowledges that AI could lead to large-scale, persistent unemployment. He writes that this may be due to "the inherent property of technology to widely replicate human cognition," and that traditional economic adaptation mechanisms could be overwhelmed by the speed of technology. This assessment is consistent with his previous stance in *The Adolescence of Technology*, but expressed with greater certainty.

Alongside the article, Anthropic announced three new initiatives: a $200 million research fund for the future of the economy for empirical research and policy experimentation; a $150 million national scholarship program for early-career professionals; and funding for cutting-edge model testing of legislative proposals and unemployment policy frameworks. The official name of the third initiative has not yet been announced, but according to multiple media reports, its core is to directly fund legislative initiatives.

The total amount of these three initiatives is approximately $350 million. For reference, Anthropic completed a $30 billion Series G funding round in February 2026, valuing the company at $380 billion. The $350 million represents approximately 1.2% of that funding round.

The $200 million research fund for the economic future didn't appear out of thin air. Anthropic launched the Economic Futures Program in June 2025, initially committing $10 million. From $10 million to $200 million, the size increased 20-fold in just one year. This leap indicates that Amodei's assessment of the economic impact of AI is rapidly tightening; he no longer considers it a long-term issue.

From optimistic vision to policy calls

Amodei's policy shift did not happen suddenly. A corrective trajectory emerges when you look at his three major articles published over the past two years together.

In October 2024, Amodei published "Machines of Loving Grace." This lengthy article has an optimistic tone. In it, he depicts a future where AI greatly benefits humanity: in biology and health, AI can compress scientific discoveries that would otherwise take decades into a few years; in economic development, AI can bring unprecedented productivity gains; and on a broader societal level, AI has the potential to help humanity solve major problems such as climate change and poverty.

The core message of this article is that the risks of AI are real, but the rewards will be enormous if humanity can safely navigate the critical window of opportunity for technological development. Amodei at the time positioned this window around 2026.

In January 2026, Amodei published "The Adolescence of Technology." The tone of this lengthy article shifted significantly. He likened the current stage of technological development to the "adolescence" of human civilization: dangerous, unpredictable, but unavoidable. He began to move from purely technological security to broader socio-economic risks, calling for a wealth tax to address the potential economic impact of AI.

In his article, Amodei no longer describes the economic risks of AI as "managed transitional pains," but instead begins to use terms like "structural shock." He writes that the impact of AI on the labor market may not be gradual, but rather abrupt; once certain cognitive abilities are replicated by models, the corresponding occupational groups may face large-scale replacement in a short period of time.

Then came the "Policy on the AI Exponential" in June 2026. Amodei's stance shifted from "early warning" to directly prescribing policy solutions, and they were willing to invest money and effort to promote it.

The optimistic vision of "Machines of Loving Grace" was squeezed by the realistic risks of "The Adolescence of Technology"; the latter's warnings about the labor market were further escalated into a direct appeal for policy tools in "Policy on the AI Exponential." This is not a wavering stance, but rather a gradual adjustment of judgment forced upon us by the consequences of technological capabilities spilling over into economic and national security.

What was removed from RSP from version 1.0 to 3.0?

To understand why Amodei shifted from "internal constraints" to "external demands," we need to examine what happened to Anthropic's internal self-regulatory framework.

In September 2023, Anthropic released RSP version 1.0. This is an internal security governance framework with the core commitment that if a model reaches certain predefined thresholds of dangerous capabilities and there are insufficient security measures in place, Anthropic will suspend training or deployment. RSP 1.0 represents a typical "inward constraint" approach: the enterprise sets its own red lines, monitors them itself, and commits to compliance. At the time, this framework was regarded by the security community as a benchmark for self-discipline among leading AI companies.

On February 24, 2026, Anthropic released RSP version 3.0. This version abandoned some of its earlier stringent commitments, including modifications to the hard pause training conditions. The security community reacted swiftly. Zvi Mowshowitz, a commentator who has long followed AI security, published an analysis on Substack criticizing this change as a concession to commercial competitive pressure. Similar criticisms emerged on the Effective Altruism Forum, arguing that Anthropic's retreat from its security commitments demonstrates that relying solely on corporate self-discipline is unsustainable in reality.

The evolution of RSP from version 1.0 to 3.0 exposes a structural problem. When both business competition and accelerating technology exert pressure, it becomes difficult to maintain the rigid security commitments unilaterally made by companies. If a company slows down due to security concerns, while competitors do not share the same constraints, security self-discipline becomes a competitive disadvantage.

This predicament directly laid the logical groundwork for Amodei's policy appeal in June. In his book "Policy on the AI Exponential," he essentially acknowledged this: since internal self-regulation is insufficient, external enforcement is needed to set industry bottom lines.

However, this presents a trust paradox. While Amodei is calling for external oversight, Anthropic's own self-regulatory framework has just undergone a revision criticized by the security community as a "step backward." Some members of the security community are therefore questioning whether a company that has compromised on internal self-regulation is qualified to call for mandatory government oversight. In discussions on Hacker News, some commentators have described it as being "both player and referee."

What are they buying with $350 million?

The three new initiatives announced by Anthropic appear on the surface to be charitable and public welfare initiatives. However, when viewed alongside Amodei's policy appeals, the true function of these funds becomes clearer.

A $200 million research fund for the future of the economy will be used for empirical research and policy experimentation. A $150 million scholarship program will be available for early-career professionals. A third initiative will directly fund efforts to advance legislative proposals and policy frameworks.

These funds are precisely targeted at the policy directions outlined by Amodei in his lengthy article. The Economic Future Research Fund can be used to fund research that supports his assertion that "AI causes structural unemployment," providing academic endorsement for policy legislation. Scholarship programs can cultivate a group of professionals who share his governance philosophy. Legislative promotion funds are the most direct: funding the drafting and lobbying of bills that align with Anthropic's security principles.

In a Hacker News discussion, a developer raised the issue of "regulatory capture." This concept refers to companies using regulatory pressure to raise industry entry barriers, thereby consolidating their market position. Mandatory third-party testing and high compliance costs are manageable expenses for leading companies like Anthropic, which already possess well-established security teams and red-team testing capabilities. However, for startups with limited funds and talent, this can constitute a formidable barrier.

An analysis on Medium directly raises this question: Is Amodei's proposal a safety plan or a blueprint for regulatory capture? The article points out that Anthropic abandoned its commitment to a hard moratorium in RSP 3.0, but is now asking the government to use legal force to constrain the entire industry.

From Amodei's perspective, he might not see this as a contradiction. In his framework, the compromise of internal self-regulation stems precisely from a prisoner's dilemma caused by a lack of external enforcement. If the government sets uniform industry standards, companies wouldn't need to choose between "safety" and "competition." From this perspective, the $350 million is an attempt to break this dilemma.

However, this logic rests on one premise: the regulatory framework must be designed fairly and not favor any particular company. Whether Anthropic, as the framework's promoter and funder, can maintain this distance remains an unanswered question.

Will the security community buy into this?

Amodei's policy call has sparked a "divisive" reaction from developers and the security community.

Some believe this marks the first time leading AI companies have acknowledged the limitations of industry self-regulation with such concrete policy recommendations and substantial financial investment. Given the rapid spillover of AI capabilities into sensitive areas such as cybersecurity and biosafety, government intervention to set bottom lines is both necessary and urgent.

Another group of critics focused on two points. The first was the issue of trust. Anthropic's concessions in RSP 3.0 led some members of the security community to believe that the company had overdrawn its credibility. Zvi Mowshowitz's analysis on Substack criticized Anthropic's retreat from its promises point by point. Against this backdrop, Amodei's call for government regulation was interpreted by some as "if they can't do it themselves, they'll get the government to force everyone else to do it."

The second concern is the risk of regulatory capture. The compliance costs of mandatory third-party testing could become a moat for leading companies. Anthropic's investment in security infrastructure is among the industry's best; if regulatory standards are based on Anthropic's existing practices, other companies will face enormous costs to catch up.

However, one factor in Amodei's appeal makes the "do nothing" option increasingly unacceptable. Following its release in April 2026, Claude Mythos Preview demonstrated cybersecurity capabilities exceeding the expectations of many observers. According to Anthropic's official system card, the model's performance in zero-day vulnerability discovery tasks has triggered an internal protocol for reporting to national security agencies. The UK's AI Security Institute subsequently published an independent assessment of Mythos Preview, confirming the magnitude of its cybersecurity capabilities.

When a model's capabilities have reached the level of national security, and a company's internal RSP framework is forced to compromise under the pressure of commercial competition, the remaining options are indeed limited. Either the government intervenes to set mandatory limits, or the company accepts a competition without any hard constraints. This tension is not unique to Anthropic.

What has changed in the past three years?

From “responsible scaling” to “exponential policy gaps,” the most fundamental change in Amodei’s governance philosophy is that he no longer believes that corporate self-discipline is sufficient to keep pace with the speed of AI development.

This judgment was gradually pushed forward by a series of events. The optimistic vision of 2024 was revised in the face of the technological realities of 2025 and 2026. The compromise of RSP from version 1.0 to 3.0 proved the fragility of internal self-regulatory frameworks under the pressure of business competition. The overflow of Claude Mythos Preview's capabilities made the incremental approach of "waiting for risks to become concrete before legislation" no longer feasible.

Amodei's solution is to shift towards external enforcement. He not only called on governments to establish a regulatory mechanism similar to the FAA, but also invested $350 million to promote the framework's implementation. The essence of this shift is that when the speed of technological development exceeds the capacity of corporate self-regulation, the only remaining binding force is the government.

However, this plan also brings new problems. The risks of regulatory capture, the trust deficit caused by the retreat of self-regulatory commitments, and the controversy in the economics community regarding the assertion that "AI will inevitably lead to large-scale and persistent unemployment" are all real obstacles that Amodei's policy appeals need to face.

For those observing the cutting edge of the AI industry, Amodei's lengthy article provides a clear signal: when the CEO of an AI company that champions safety publicly admits that its self-regulatory framework is insufficient and proactively seeks government assistance, this is a noteworthy development.

SafeDAO关于“开放SAFE可转让性”的提案获得通过，暂定4月23日开始实施

Popular Articles