Topic / Subject

TIME reports Anthropic dropped the central pledge in its Responsible Scaling Policy, backing away from the “we won’t train unless safety is guaranteed in advance” stance.

TL;DR

Anthropic is shifting from a hard “stop sign” promise to a transparency-and-reporting approach: more public roadmaps, recurring risk reports, and fewer automatic pauses.

Key Details

TIME reports Anthropic removed the core commitment from its 2023 policy that barred training unless safety measures were assured in advance. Anthropic’s Jared Kaplan told TIME the company doesn’t think unilateral commitments make sense if competitors keep racing. TIME says the revised policy emphasizes publishing Frontier Safety Roadmaps and Risk Reports every three to six months. The new approach narrows when the company would “delay” development to specific high-risk conditions rather than a categorical stop.

Breakdown

This is a brand shift. Anthropic has spent years positioning itself as the safety-forward lab, so dropping the headline pledge is going to land as “the race changed them.”

Anthropic’s argument (per TIME and Business Insider) is basically: safety doesn’t work as a solo vow if everyone else keeps sprinting. So instead of promising “we’ll stop,” they’re promising “we’ll show our work” and only pause under narrower conditions.

The real test won’t be the policy PDF. It’ll be what happens when the next “this feels risky” capability shows up: do the reports create pressure and accountability, or do they become paperwork that everyone ignores?

What to Watch Next

How detailed and frequent the Risk Reports actually are (and whether they include meaningful evaluations). Whether other frontier labs adopt similar “less binary, more reporting” policies. Any concrete examples where Anthropic delays or changes a release because of the new framework.

Sources

TIME — Exclusive: Anthropic Drops Flagship Safety Pledge

Business Insider — Anthropic is dropping its signature safety pledge amid a heated AI race

Comment

Do you trust transparency reports to keep AI labs honest, or does safety need hard “stop” rules to mean anything?