Anthropic dials back its core AI safety pledge

Anthropic just walked back the core AI safety pledge that set it apart from every other lab in Silicon Valley.

Chief Science Officer Jared Kaplan told Time Magazine this week that Anthropic will no longer stick to its 2023 commitment never to train an AI system without adequate safety guardrails in place. Instead, the company will simply show enterprise clients how its models score on safety evaluations.

Kaplan admitted the shift was necessary to stay competitive in an increasingly cutthroat market. Basically, the self-proclaimed responsible AI lab is loosening its own brakes to keep up with OpenAI.

Caught between rivals and the Pentagon

The policy tweak comes right as Claude usage is surging, but Anthropic is getting squeezed from both sides.

On one front, competitors are running full throttle. Michael Bennett, associate vice chancellor for data and AI strategy at the University of Illinois Chicago, pointed out that governments worldwide are not aggressively regulating AI. If Anthropic holds itself back while rivals innovate unchecked, someone without safety qualms will simply take the lead.

On the other front, Anthropic is locked in a high-stakes standoff with the U.S. Department of Defense. Defense officials recently warned that Anthropic could become a supply chain risk because it refuses to let its AI power mass surveillance of Americans or fully autonomous weapons. Losing lucrative military contracts—and access to commercial partners working with the Pentagon—would be a devastating financial blow.

Meanwhile, Washington's political winds have shifted. With Donald Trump favoring unfettered innovation and signing a December executive order to stop states from passing local AI rules, federal guardrails are practically nonexistent.

What enterprise buyers actually care about

Anthropic’s original Responsible Scaling Policy was rigid: it set measurable model tiers and required development pauses if safety thresholds were breached.

Yet enterprise buyers are not necessarily shopping for moral posture. Forrester analyst Jeff Pollard noted that clients flocking to agents like Claude Code just want to write better software faster. Safety underpinnings were a nice-to-have, not a dealmaker.

Still, watering down commitments carries reputational risk. Lily Li, founder of Metaverse Law, warned that diluting public safety promises could eventually alienate privacy-conscious customers and dent the bottom line.

Why it matters

Anthropic just admitted what the rest of the tech world quietly accepted months ago: you cannot win an AI war with self-imposed speed limits.

For developers and everyday users, this likely means faster Claude model releases and fewer rollout delays caused by internal safety debates. But if you rely on Anthropic because you trusted their self-restraint over OpenAI, that moat is evaporating. Shifting from a hard rule of not training risky models to simply publishing test results places the burden of risk evaluation squarely on you, the customer.

For the broader industry, the last true safety hawk has yielded. When the company founded specifically to be safer than OpenAI dials back its own scaling policy, voluntary industry self-regulation is officially dead. Anthropic made a pragmatic business pivot to survive government backlash and maintain parity with peers, but it traded away its sharpest moral differentiator in the process.