For years, Anthropic positioned itself as the responsible counterweight to the breakneck pace of artificial intelligence development. Founded by former OpenAI researchers who left partly over safety concerns, the company built its brand on a simple but powerful pledge: if its AI models ever showed signs of being capable of causing catastrophic harm, Anthropic would stop deploying them until the risks were mitigated. Now, the company has walked that promise back — and the implications for the entire AI industry are significant.
As first reported by Time, Anthropic has revised its Responsible Scaling Policy (RSP), the framework it introduced in 2023 to govern how it develops and releases increasingly powerful AI systems. The original policy contained what amounted to a hard commitment: if internal evaluations determined that a model had reached dangerous capability thresholds — specifically around biological weapons, cybersecurity attacks, or autonomous self-replication — the company would halt deployment until adequate safeguards were in place. The updated version, released quietly alongside other corporate communications, softens that language considerably, replacing firm commitments with more flexible, discretionary language.
From Hard Lines to Soft Guidelines
The original RSP was built around a system of “AI Safety Levels,” or ASLs, modeled loosely on the biosafety level classifications used in laboratories handling dangerous pathogens. Each level corresponded to a tier of model capability and a corresponding tier of required safety measures. ASL-1 covered models with no meaningful dangerous capabilities. ASL-2, the current classification for Anthropic’s Claude models, covered systems that might provide some uplift for malicious actors but could be managed with existing safeguards. ASL-3 and beyond were reserved for models that could materially increase catastrophic risks — and it was at these thresholds that the company’s deployment pause commitment was supposed to kick in.
Under the revised policy, according to Time, Anthropic no longer frames these thresholds as automatic tripwires. Instead, the company has given itself considerably more latitude in how it interprets evaluation results and what actions it takes in response. The language has shifted from prescriptive rules to principles-based guidance, a move that safety researchers and industry observers say effectively removes the teeth from the policy. Where the original RSP said the company “will not” deploy models above certain capability thresholds without corresponding safeguards, the new version uses softer formulations that leave room for judgment calls by company leadership.
Why Anthropic Says the Change Was Necessary
Anthropic has defended the revision, arguing that the original RSP was written at a time when the company had less experience with how AI capabilities actually develop and how safety evaluations perform in practice. Company representatives have indicated that the rigid structure of the original policy created operational challenges and that a more adaptive framework better serves the goal of responsible development. In essence, Anthropic is arguing that the spirit of the commitment remains intact even if the letter has changed.
Dario Amodei, Anthropic’s CEO, has previously spoken publicly about the tension between safety commitments and competitive pressures. In a widely discussed essay published last year, Amodei acknowledged that being overly cautious could cede ground to less safety-conscious competitors, potentially leading to worse outcomes overall. This argument — sometimes called the “race to the top” theory — holds that responsible companies need to stay at the frontier of AI development to ensure that the most powerful systems are built by organizations that care about safety. Critics have long pointed out that this logic can be used to justify almost any acceleration of development timelines.
The Competitive Pressure Behind the Curtain
The timing of Anthropic’s policy revision is difficult to separate from the intensifying competition among leading AI companies. OpenAI, Google DeepMind, Meta, and xAI are all racing to develop more capable models, with billions of dollars in funding and enormous commercial incentives driving the pace. Anthropic, despite its safety-first branding, is not immune to these pressures. The company has raised over $7 billion in funding, including major investments from Amazon and Google, and its investors expect returns that depend on the company remaining competitive at the frontier.
Recent months have seen a notable acceleration across the industry. OpenAI has been rolling out increasingly capable models and pushing toward artificial general intelligence on aggressive timelines. Google DeepMind has made significant advances with its Gemini model family. Meta continues to release powerful open-source models. In this environment, a company that voluntarily pauses deployment of its most capable systems risks falling behind — not just commercially, but in its ability to attract the top research talent that gravitates toward whoever is building the most advanced technology.
Safety Researchers Sound the Alarm
The reaction from the AI safety community has been swift and largely negative. Researchers who had pointed to Anthropic’s RSP as a model for the industry — and who had urged other companies to adopt similar commitments — now find themselves without their strongest example of corporate self-regulation. Several prominent safety researchers have taken to social media platforms including X to express concern that the revision signals a broader retreat from safety commitments across the industry.
The concern is not merely symbolic. Anthropic’s original RSP was influential in shaping how policymakers, journalists, and the public understood the state of AI safety governance. When lawmakers in the United States, the European Union, and the United Kingdom considered how to regulate AI, Anthropic’s voluntary commitments were frequently cited as evidence that industry self-regulation could work — or at least that it was being seriously attempted. The weakening of those commitments undermines one of the central arguments against more aggressive government intervention.
What This Means for AI Regulation
The policy shift arrives at a particularly sensitive moment for AI governance. In the United States, the regulatory picture remains fragmented, with no comprehensive federal AI safety legislation in place. California’s SB 1047, which would have imposed safety testing requirements on frontier AI developers, was vetoed by Governor Gavin Newsom last year after intense industry lobbying. In the absence of binding regulation, voluntary commitments like Anthropic’s RSP have served as a kind of stopgap — reassuring the public and policymakers that the most powerful AI systems were being developed with appropriate caution.
With Anthropic softening its stance, the case for mandatory regulation becomes harder to dismiss. If the company most publicly committed to self-imposed safety constraints is now walking those constraints back under competitive pressure, it raises serious questions about whether any voluntary framework can hold up against the financial incentives driving AI development. Policymakers who had been willing to give the industry time to demonstrate responsible self-governance may now feel that window has closed.
A Pattern Across the Industry
Anthropic is not the first AI company to retreat from safety commitments. OpenAI, originally founded as a nonprofit dedicated to developing AI safely for the benefit of humanity, has undergone a dramatic corporate restructuring that critics say prioritizes commercial interests over its original mission. The company dissolved its “superalignment” team last year after key researchers departed, and its transition toward a for-profit structure has drawn scrutiny from former board members and co-founders alike. Google DeepMind, which once operated with significant independence and a strong safety research mandate, has been increasingly integrated into Google’s commercial operations.
The pattern is consistent: as AI companies grow larger, raise more capital, and face more intense competition, their safety commitments tend to erode. This is not necessarily because the individuals involved stop caring about safety — many of them clearly do — but because the structural incentives of the market push relentlessly toward faster development and broader deployment. Voluntary commitments, no matter how sincerely made, struggle to withstand these forces over time.
The Stakes Beyond Corporate Strategy
What makes this moment particularly consequential is the nature of the risks involved. The capability thresholds that Anthropic’s original RSP was designed to address — biological weapons development, sophisticated cyberattacks, autonomous AI behavior — are not hypothetical concerns dreamed up by science fiction writers. They are scenarios that leading AI researchers, including many within Anthropic itself, have identified as plausible consequences of continued capability gains. The question of how to handle models that approach these thresholds is arguably the most important governance challenge the technology industry has ever faced.
Anthropic’s revised policy does not abandon safety entirely. The company continues to conduct evaluations, publish research, and invest heavily in interpretability and alignment work. But the shift from binding commitments to flexible guidelines represents a meaningful change in the company’s relationship with risk. It moves the locus of decision-making from a transparent, rules-based framework to an opaque, judgment-based one — and in doing so, it asks the public to trust that company leadership will make the right calls when the stakes are highest, even when those calls conflict with commercial interests.
The Industry at an Inflection Point
For an industry that has asked repeatedly for the public’s trust, Anthropic’s decision is a significant data point. The company that was supposed to prove that AI could be developed responsibly without government mandates has just demonstrated the limits of that approach. Whether this leads to stronger regulation, a renewed push for binding international agreements, or simply a further erosion of public trust in AI companies’ ability to govern themselves remains to be seen. What is clear is that the safety-first era of AI development — to the extent it ever truly existed — is giving way to something more complicated, more competitive, and potentially more dangerous.
The AI industry now faces a fundamental question: if the company that cared the most about safety cannot maintain its own commitments, what does that say about the rest of the field? The answer to that question will shape not just the future of artificial intelligence, but the future of the regulatory and institutional frameworks that govern it.