Anthropic was founded on a specific premise: that it was possible to build powerful AI and take safety seriously at the same time, and that if you could not do both, you should slow down. That premise has always been in tension with the commercial realities of competing against OpenAI and Google. Bloomberg now reports that the tension has produced a policy shift — one that is easy to characterize as a capitulation, harder to evaluate honestly, but worth understanding either way.
The change is to Anthropic's Responsible Scaling Policy, the framework it introduced in 2023 to govern how it develops and deploys increasingly capable models. The original policy included a commitment that Anthropic would not deploy AI it deemed potentially dangerous if competitors reached similar capability thresholds first — the logic being that competitive pressure was not a legitimate reason to take safety shortcuts. The updated version, per Bloomberg, adds a caveat: if competitors have already released equivalent capabilities, Anthropic's rationale for holding back is weakened, and it can proceed.
At first glance this reads as the safety-first lab quietly deciding that safety is negotiable when market share is at stake. That framing is not entirely wrong, but it is also incomplete. The original policy rested on a theory about how safety leadership works: if the careful lab leads, it sets norms; if it lags, it loses both influence and the ability to fund safety research. That theory has a structural problem — it assumes other labs will respond to Anthropic's restraint with restraint of their own, or that regulators will step in to enforce the norm. Neither has happened.
What Anthropic is implicitly acknowledging is that a unilateral pause in a race where competitors are not pausing is not safety — it is just losing. The updated policy still maintains thresholds and evaluation requirements; the change is in what happens when those thresholds are approached and a competitor is already past them. Whether that is a reasonable pragmatic update or an erosion of the core commitment depends heavily on what "potentially dangerous" actually means in practice, and Anthropic has not been fully public about those specifics.
This comes at a moment when Anthropic's positioning on safety is under pressure from multiple directions. It recently donated $20 million to a group pushing for AI regulations, signaling it still believes in external governance. It is simultaneously suing the Defense Department over a supply chain risk designation that it argues is politically motivated. And Dario Amodei has repeatedly warned publicly that AI could cause "unusually painful disruption" to labor markets — language that does not suggest the company thinks its technology is trivial or safe by default.
The uncomfortable reality is that no safety pledge survives contact with a competitive market intact. OpenAI's original non-profit structure and safety commitments look very different today than they did in 2015. The question for Anthropic is whether the update to its RSP represents a managed, transparent evolution of its framework — which it has argued it is — or the beginning of a slide toward the same logic its founders left OpenAI to avoid. Bloomberg's headline, notably, calls it a "race against rivals." Anthropic would probably prefer a different framing, but it is not wrong about what is driving the change.
For observers who have followed AI safety debates, the policy update is less surprising than the timing. Anthropic is raising capital, competing for enterprise contracts, and navigating a political environment that is actively hostile to the idea that AI should be slowed down for any reason. In that context, the question is not whether its commitment to safety has shifted — it clearly has — but whether enough of the underlying framework remains to matter when it really counts.