Anthropic’s long-standing refusal to deploy AI models it deemed unsafe is now conditional. In a policy shift that could redefine the company’s stance on AI governance, Anthropic has revised its Responsible Scaling Policy to include a critical loophole: it will no longer automatically halt training on models that pose risks if its competitors—like OpenAI or xAI—are already developing similar systems. The change, detailed in version 3.0 of the policy, introduces flexibility where there was once an absolute rule.
This isn’t just bureaucratic tweaking. It’s a strategic pivot that arrives at a tense moment. The Pentagon has demanded broad access to Anthropic’s models for military applications, with a deadline looming. Meanwhile, rivals have already secured government partnerships, leaving Anthropic in a position where its historic safety-first approach may no longer align with its operational reality.
The Policy Flip: From ‘No’ to ‘If Others Do First’
Anthropic’s original policy was clear: if a model failed to meet its safety standards, training would stop, even if competitors raced ahead. Now, that red line has been softened. The revised policy states that Anthropic will only delay training on hazardous models if it believes it can maintain a lead over rivals. If others are already pushing boundaries, Anthropic reserves the right to proceed—provided it commits to transparency about the risks.
The framing suggests a calculated risk. Anthropic argues that by refusing to lag behind, it can act as a ‘steadying force’ in an industry where reckless scaling might otherwise dominate. But critics question whether this is a genuine safeguard or a calculated concession to pressure.
A Pentagon Deadline and the Looming Threat of Force
The timing of this policy change couldn’t be more fraught. Defense Secretary Pete Hegseth has reportedly warned Anthropic that the Pentagon may invoke the Defense Production Act—a rarely used but powerful tool—to compel the company to hand over its models for ‘any lawful purpose.’ Anthropic’s counteroffer is firm: no access unless the Pentagon rules out autonomous weapons development and surveillance of U.S. citizens. Yet the revised Responsible Scaling Policy may now offer a third way—one that lets Anthropic comply with military demands without fully surrendering control.
OpenAI and xAI have already struck deals with the Pentagon, leaving Anthropic isolated in its resistance. The new policy could be seen as a strategic retreat, allowing the company to engage with the government while still claiming to uphold safety standards.
What This Means for AI Safety—and Anthropic’s Future
The shift raises fundamental questions about whether safety in AI is about absolute rules or relative leadership. If Anthropic now follows where others lead, does that erode its reputation as the industry’s ethical standard-bearer? Or does it simply acknowledge that no company can realistically operate in isolation when competitors and governments are pushing boundaries?
Anthropic insists the change doesn’t weaken its commitment to safety. Instead, it’s an acknowledgment that rigid policies, in a world where AI advances rapidly, may no longer be feasible. The company now argues that by participating in the development of risky models—rather than letting others do so unchecked—it can still influence outcomes.
Yet the optics are undeniable. A company that once prided itself on walking away from unsafe AI now appears willing to walk a tighterrope. Whether that balance holds depends on how closely the Pentagon scrutinizes its promises—and whether Anthropic’s rivals will use this as proof that even the most cautious players are now bending.
The stakes couldn’t be higher. If this policy shift emboldens competitors to relax their own safeguards, the entire industry could drift toward less scrutiny. But if Anthropic succeeds in threading the needle—engaging with the Pentagon while still setting limits—it might just redefine what ‘responsible scaling’ looks like in an era where AI governance is still being written.
