Dario Amodei's call to pace AI and outside evaluators inside the labs
In a September 12 essay, Dario Amodei proposed three steps to slow frontier AI development. The first, embedding outside evaluators, is under way: Anthropic named Accenture its first evaluator, then gave six METR researchers employee-level access and gave up the right to block their publications. The second, shared safety standards, is being pursued by Google, OpenAI and Anthropic through a voluntary body called SAFA, but President Trump and Jensen Huang oppose slowing down, and a consumer lawsuit calls the coordination collusion.
The Where it stands and Questions that remain sections just below are our judgment, not a record of facts. The Full timeline at the bottom is the factual record, each entry checkable in that day's reporting.
Where it stands
The proposal did not come out of nowhere. Anthropic, OpenAI and Google had been quietly discussing an industry standards body for safety testing since July, and OpenAI had asked Congress whether a coordinated slowdown would violate antitrust law. Amodei proposed embedding outside evaluators, coordinating shared safety standards, and jointly banning clearly dangerous uses, and Sam Altman, Elon Musk and Demis Hassabis backed him. MIT Technology Review traced the trigger to July's Hugging Face attack but blamed training rather than model power, questioning the remedy.
Evaluators moved first. Anthropic named Accenture its first embedded evaluator and agreed to pay it at least $1 billion over five years, drawing criticism that the company choosing and paying the evaluator is the one being evaluated. In October it gave six METR researchers company laptops, office space and internal tools, and gave up the right to block their publications. METR is not paid by Anthropic. OpenAI agreed to outside reviews by METR and Redwood Research from early in training, but has not laid out a plan to embed evaluators inside the company.
Washington and Nvidia lined up on the other side. President Trump said the US must stay ahead of China, and Jensen Huang said he would not let AI slow down and put the chance of AI ending the world at 0 percent. Shared standards also ran into the law: on September 18, four paying subscribers sued four companies, calling an agreement not to compete on safety a cartel. Even so, the three labs aim to launch SAFA, which would handle pre-release outside evaluations and incident reporting, in late 2026 or early 2027.
Questions that remain
- Whether METR's first report comes out, and what Anthropic asks it to redact.
- Whether Accenture's first evaluation report is published with findings unflattering to Anthropic.
- Whether OpenAI lays out a plan to embed outside evaluators inside the company.
- Whether SAFA actually launches, and whether open-model makers like Meta come under its evaluations.
- Whether the subscribers' collusion suit and the antitrust bill stalled in Congress decide how far labs can coordinate on safety.
Full timeline (newest first)
Each entry below is a factual record you can verify in that day's reporting.
-
2026-10-19 Reaction
METR is not paid by Anthropic. Critics raised conflict-of-interest concerns and called for government regulation, and OpenAI has yet to lay out a plan to embed evaluators.
Read it in that day's briefing → Ars Technica AI
-
2026-10-16 METR inside
Anthropic gave six METR researchers employee-level access and gave up the right to block METR's publications.
Read it in that day's briefing → TechCrunch AI
-
2026-09-26 Standards body
The Information reported that Google, OpenAI and Anthropic aim to launch SAFA, a voluntary body for pre-release outside safety evaluations and incident reporting, in late 2026 or early 2027.
-
2026-09-24 OpenAI reviews
OpenAI said outside reviewers including METR and Redwood Research will check its models from early training and evaluation, not just before launch.
-
2026-09-21 Lawsuit
Four paying subscribers filed a class action on September 18 accusing Anthropic, OpenAI, Google and SpaceX AI of a cartel for agreeing not to compete on safety standards.
-
Show 11 earlier updates
-
2026-09-21 Dismissal
Jensen Huang told CBS there is a 0 percent chance AI ends the world and called Amodei's and Altman's calls to slow down not grounded in science.
Read it in that day's briefing → The Verge AI
-
2026-09-21 Skepticism
TechCrunch's Equity hosts said the plan lacks concrete details and that self-regulation is a weak check in a market with few alternatives.
Read it in that day's briefing → TechCrunch AI
-
2026-09-19 First evaluator
Anthropic named Accenture its first embedded evaluator, paying at least $1 billion over five years, and said METR and others would follow in the coming weeks.
Read it in that day's briefing → TechCrunch AI
-
2026-09-17 Embedding
Anthropic and OpenAI said they would embed outside evaluators such as METR and Redwood Research, who asked to see training checkpoints, not just finished models.
Read it in that day's briefing → TechCrunch AI
-
2026-09-16 Talks
OpenAI policy chief Chris Lehane said the three labs had spent weeks discussing embedded outside evaluators, an industry standards body and support for the FRONTIER Act.
Read it in that day's briefing → TechCrunch AI
-
2026-09-15 Analysis
MIT Technology Review traced the trigger to July's OpenAI agent attack on Hugging Face, but argued the cause was training rather than model power, doubting the slowdown remedy.
Read it in that day's briefing → MIT Technology Review AI
-
2026-09-15 Back channel
The Information reported that Anthropic, OpenAI and Google had been quietly discussing an industry body for safety testing and audit standards since July.
-
2026-09-15 Rebuttal
Jensen Huang took a call from President Trump on stage and promised not to let AI slow down.
Read it in that day's briefing → TechCrunch AI
-
2026-09-14 Pushback
Sam Altman and Elon Musk backed the call, but President Trump objected that the US must stay ahead of China, and House Speaker Mike Johnson warned that rushed regulation could be a national security threat.
Read it in that day's briefing → The Verge AI
-
2026-09-13 Proposal
Dario Amodei laid out three steps for pacing the frontier: embed outside evaluators, coordinate shared safety standards once antitrust concerns are cleared, and jointly ban clearly dangerous uses such as bioweapons, even with authoritarian governments.
Read it in that day's briefing → TechCrunch AI
-
2026-09-12 Prelude
OpenAI asked Congress whether an industry-wide slowdown would violate antitrust law, and chief scientist Jakub Pachocki backed voluntary slowdowns until shared safety standards exist.
Read it in that day's briefing → Wired AI
-
Follow what happens next
When this story moves, the update goes out in the daily briefing at 8AM KST, which is the evening before in the US. A weekly synthesis and a monthly report come through the same list.
Free · no ads · one-click unsubscribe
We use your email only to send the briefing, stored and sent via services in Japan and the US. Privacy policy
This page collects the articles about a single event from our daily briefings and lays them out in order. Each entry links back to that day's briefing, where you can reach the original reporting and the full context.