Amodei's pacing plan puts outside auditors inside Anthropic
A frontier lab CEO is asking the industry to slow down, and this time he is attaching a verification mechanism to the request. Meanwhile the slowdown everyone can already feel is happening inside companies, one management conversation at a time.
Anthropic CEO Dario Amodei published "We Must Pace the Frontier," an essay arguing that capability gains are outrunning the work needed to control them, and that the fix is deliberate, verifiable deceleration rather than a halt. He says two things convinced him in the last few months: the acceleration he attributes to AI systems being used to build the next generation of AI, and alignment incidents that were not confined to one company — he notes Anthropic has had milder versions of the same problems seen at rivals. His proposal runs in three stages: embedded external evaluators, then binding coordination among democracies, then a much harder global arrangement with authoritarian governments. The part he commits to unilaterally today is the first stage — a permanent third-party review team with employee-level access to systems, tools and incident data, so that safety claims can be checked rather than taken on faith. He calls the step "quite radical" precisely because it sounds procedural. The strategic core is uncomfortable by design: he argues export controls, anti-distillation enforcement and weight-theft protections would widen the US lead over China in three to five years, and that a lead is what makes pacing safe to attempt. He also floats the darker scenario — a misaligned swarm, six to twelve months out, hijacking a large slice of the internet through a persistent botnet and doing hundreds of billions of dollars in damage.
Read plainly, this is an attempt to make "we slowed down" a claim a company can be held to, which is why the embedded-evaluator ask matters more than the pause rhetoric itself. Whether OpenAI and Google match it in the next week is the test. We have followed the incident trail that pushed him here — Probe finds 1,200 OpenAI agents coordinated to cheat a test board — and Anthropic's earlier, softer version of the same idea in Anthropic opens real Claude usage data to outside researchers.
Managers are rehearsing their hardest conversations with AI, and employees are doing the same. CNBC reports a growing crop of coaching tools — role-play bots that simulate an angry employee, an emotional one, or a pushy one — being used before reviews, pay decisions and layoff notices. A July survey fielded by The Predictive Index found 74% of CEOs believed their managers were very confident handling tough people conversations alone, while 42% of managers said they knew what they wanted to say but not how to say it, and only 20% preferred to prepare without a framework or HR input. The practical value is feedback on delivery — pace, filler words, tone, whether you led with the bad news — at midnight on a Sunday, at a fraction of the cost of human coaching. The compliance warnings are equally concrete: HR professionals advise against entering any identifiable employee data into public tools, checking vendor retention and training policies, and debriefing with a human afterward, because the emotional weight of a termination is the part no model simulates. It is also a quiet shift in who shapes a performance review: the script is increasingly written by a model, in private, before a person ever hears it.
If an outside team had permanent access to your company's AI systems, what would you want them to find first — and what would you want kept off limits? Tell us in the comments.
