Anthropic built its brand on being the 'safe AI company' with strong commitments about when and how it would deploy models. Now they've quietly revised those commitments. The changes are subtle but significant for anyone who took their safety positioning seriously.
As someone who's watched companies pivot from principles to profits more times than I can count, I want to document exactly what changed and what it means.
The original commitment, articulated in Anthropic's founding documents and safety policies, included specific thresholds for model deployment. The company promised not to release models that exceeded certain capability levels without corresponding safety measures. They outlined Responsible Scaling Policies (RSP) that would govern how they handled increasingly powerful AI systems.
The new version? Let's just say it's more... flexible.
According to the updated policy documents, Anthropic has relaxed some of the specific deployment criteria and broadened the conditions under which they would release capable models. The exact language has shifted from hard lines to contextual evaluation.
Anthhropic's explanation is that they're adapting to reality. The original safety framework was developed when the company was smaller and the competitive landscape was different. Now they're facing pressure from OpenAI, Google, and a dozen well-funded competitors. The market moves fast. Customers want capabilities. Staying rigidly committed to safety thresholds that competitors ignore puts you at a disadvantage.
That's all true. It's also exactly the kind of reasoning that every company uses when they're moving the goalposts.
Critics, including some former Anthropic employees and AI safety researchers, argue that this is precisely the moment when commitments matter most. It's easy to have strong safety principles when you're a small research lab with nothing to lose. The test is whether you stick to them when there's actual pressure to compromise.
The changes aren't dramatic enough to trigger headlines in mainstream media. Anthropic isn't abandoning safety research or firing their safety team. But they are adjusting their policies to give themselves more room to release capable models without hitting the specific safety bars they previously set.
