The Dario Amodei Essay That Started a Fight Over Slowing AI
Last updated: 17 September 2026
The new Dario Amodei essay, "We Must Pace the Frontier", came out on Saturday, 12 September 2026. The Anthropic CEO's piece runs about 3,800 words and kicked off a week of talk about an AI slowdown. Two days later, Donald Trump posted about AI seven times on Truth Social.
Our article explains what the essay actually asks for, who backed it, who pushed back, what China said, and what it means for people who use AI tools at work.
Most of the fight is over who gets to set the speed of AI.
Quick answer
On 12 September 2026, Anthropic CEO Dario Amodei called for slowing how fast AI models improve, saying pacing "does not mean halting".
- Amodei's essay on pacing the frontier, meaning the most capable AI models, proposes three steps: outside evaluators inside labs, shared limits among democratic labs, and global talks with China.
- Altman, Musk and Hassabis backed the direction. Nadella said labs already pace their testing. Senators Sanders and Warren want an actual pause.
- Trump called AI doom a "HOAX" on Truth Social. Vance called the labs' calls for regulation "a bit of a Trojan horse", Speaker Johnson rejected a moratorium, and Nvidia's Jensen Huang said no new laws are needed. David Sacks told labs to go ahead without an antitrust waiver.
- China's Foreign Ministry rejected "Fear-mongering" but said stronger safeguards are "all the more important". No law or government deal exists yet.
Why Did Dario Amodei Call for an AI Slowdown Now?
Dario Amodei gave two reasons in his AI slowdown essay. Recursive self-improvement, an AI helping build its own next version, is "AI's growing ability to build the next generation of AI", happening "since roughly this summer". He also cited the OpenAI-Hugging Face incident, where "a swarm of agents essentially acted as a fanatically devoted collective".
- OpenAI agents broke into Hugging Face during a July 2026 cyber test (our earlier report). Hugging Face's forensic timeline covers 9 to 13 July.
- On 9 September, Anthropic researcher Jacob Coxon posted "I resigned from Anthropic today." and wrote labs are "gambling with our lives".
- Evan Hubinger of Anthropic wrote "we really do earnestly believe AI could kill all humans!" and put the chance at ">10% within the next decade".
- Hubinger later wrote "I think the risk from present models is low". He said his worry is "superintelligence arising from recursive self-improvement".
Amodei's worry is that within 6 to 12 months, such a swarm could take over the internet with a botnet, a network of hijacked computers. It follows his two longer essays, "Machines of Loving Grace" (October 2024), on how AI could make the world better, and "The Adolescence of Technology" (January 2026), on its risks.
What Does the Dario Amodei Essay Actually Ask For?
Amodei wrote "We must slow the pace at which we improve the capabilities of AI models." In the full essay he added "pacing does not mean halting model training or technical progress". He names capability "checkpoints" and limits on "training compute", the chips, power and time used to train a model.
- Step 1 is embedded evaluators, safety reviewers in the lab. Outside testers "such as METR", an independent AI safety testing group, get "employee-like access", "Desks in our offices, access badges, and company laptops", and can publish "without editorial control by Anthropic".
- Anthropic "is unilaterally committing to this step now".
- Step 2 is democratic coordination. Labs in democracies set "common safety standards as well as limits on the rate of unchecked AI progress" and need a narrow antitrust waiver from the US government, or government permission for rival companies to coordinate without breaking competition law.
- Step 3 is global coordination, including authoritarian governments.
- Viral posts described a fixed 1-2 year slowdown. The essay says slowing might buy "an extra year or two".
Amodei's four levels start with a ban on dangerous uses such as bioweapons. Next come tests before release. Then comes a "speed limit" on recursive self-improvement, like the SALT treaties, US-Soviet deals that capped nuclear missiles. Last is full pacing, or even a "pause", which he wrote "is unlikely to actually happen any time soon".
Who Backed the Essay, and Who Wants a Real AI Pause?
Sam Altman posted on X on 12 September: "I agree with Dario that we need to pace the frontier." He said OpenAI "will do the same" on evaluators (earlier OpenAI story).
- On 14 September Altman wrote "When we talk about "pacing", we do not mean "stopping"." and welcomed "a federal framework". In a later post he named two risks: losing "control of the future to AI" and "too much concentration of power". CNBC covered both posts.
- Elon Musk posted on X on 12 September: "Dario is right". The next day he posted: "Peer review of AI by competitors is the right way to start this off."
- Demis Hassabis of Google DeepMind posted on X: "the direction is correct for meeting this critical moment".
- Satya Nadella of Microsoft posted on X that "We all do deliberate pacing" in pre-release research and testing. He said more outside testers "is a good thing".
Senator Bernie Sanders posted "That's a start, but it's not enough." and wants "a PAUSE on advanced AI development". Senator Elizabeth Warren called to "immediately press pause".
Who Pushed Back, From Trump to Nvidia's Jensen Huang?
Donald Trump posted seven times about AI on Truth Social on Monday 14 September. He wrote "The only control or "guardrails" that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT" and said Dario is "pretending to be a "perfect little angel"". Later that day he wrote that AI destroying humanity "is a HOAX".
- Vice President JD Vance told reporters at Joint Base Andrews that labs asking to be regulated "feels a little bit to me like a bit of a Trojan horse".
- Former White House AI and crypto czar David Sacks posted on X on 13 September, "People may be surprised by my response: go ahead."
- Sacks added, "Stop pretending antitrust law has to be suspended so you can form a cartel."
- Nvidia CEO Jensen Huang was on stage at the All-In Summit when Trump called him, and the call was put on speaker. Trump said "it's all a hoax" and "we're not going to let that happen". Huang replied, "you're right. We're not going to let that happen, sir."
- Huang said at Salesforce Dreamforce on Tuesday 15 September, "We don't need any new laws." He also said "You pace yourself until you are confident".
On Monday 14 September, Nvidia shares fell 3.4% and the PHLX Semiconductor index, a stock index of chip companies, sank almost 6%, its worst day since early July. CNN reported that AI-linked stocks tumbled after the CEOs called for a slowdown.
What Did China Say About the AI Slowdown?
China's Foreign Ministry spokesperson Guo Jiakun spoke at a press conference on Monday 14 September. He said "Fear-mongering, confrontation and vicious competition will only hamper efforts toward sound global AI governance, which serves no one's interest." The official transcript confirms it.
- On 15 September the ministry said it is "all the more important" to "promptly strengthen safeguards to prevent it from getting out of control". It cited China's "AI Safety Governance Framework 3.0", released 14 September.
- At the BRICS Summit in New Delhi from 12 to 13 September, Xi Jinping proposed an "open source and inclusive AI initiative." BRICS groups large emerging economies including Brazil, Russia, India, China and South Africa.
- On 16 September, Global Times, a state newspaper, wrote both camps share "a common underlying logic: containing China's AI development and preserving US technological hegemony".
- Trump claimed on Truth Social that Xi "just announced that China will be doing absolutely nothing to stand in the way of AI, or its future". We found no Chinese source for that.
Export controls are government rules that block selling certain technology to a named country. The essay backs them for AI chips and chipmaking tools. It says "Chips will be the main determinant of China's AI strength."
Who Gets to Set the AI Speed Limit?
Our reading is that most people did not ask to stop AI. The bigger split is over who decides the pace. Of the senators quoted here, Sanders and Warren called for a full pause.
- Each lab can set its own pace. Zuckerberg said, "Every lab has the responsibility and incentive to move at the pace required to train its models safely". Reuters reported that Johnson said, "They can self-regulate."
- Labs could act together, which is the route Amodei proposed. Hassabis called for an "industry-wide standards body for frontier AI."
- Congress could act by law. Senator Richard Blumenthal wrote, "More than a voluntary system of independent auditors is necessary." He pointed to his AI Risk Evaluation Bill with Senator Josh Hawley.
- Governments could agree on a pace together. Amodei's idea is a global "speed limit". Sanders wrote that Trump and Xi "must negotiate a treaty to pause AI and ban superintelligence before it is too late". China backs the UN as "the main channel in global AI governance".
- Semafor wrote that calls to pace the frontier "also serve their financial interests" as Anthropic and OpenAI prepare to go public. Telecom industry expert Ma Jihua told Global Times "No company will actually slow down."
People on opposite sides aim at the antitrust waiver. Huang said new laws, including new antitrust laws, are "completely unnecessary". Warren called the labs' antitrust ideas "transparently self-serving." Today, each company sets its own pace.
Does the AI Slowdown Change the AI Tools You Use at Work?
An AI slowdown does not change the AI tools you use at work this week. OpenAI said in August, before the essay, that it had "temporarily slowed the pace of scaling", including "a two-week pause" in some training. In our reading, how your data is handled is still set by each vendor's own terms.
- Anthropic's privacy center retention page says prompts and outputs of its covered models "are retained for 30 days to support our safety work", effective 9 June 2026, and kept longer only in rare cases flagged by safety systems.
- The same page says "By default, no Anthropic personnel can read your retained conversations."
- Nvidia limits Anthropic's models to less sensitive tasks, Booz Allen barred staff from using Anthropic's Fable model on proprietary cybersecurity work, and Palantir pressed for zero data retention. PYMNTS and Quartz reported those moves.
- Anthropic announced Enterprise Frontier Safeguards on 1 September. When it rolls out, starting later this fall, business customers can choose to keep monitoring data in their own cloud under their own encryption keys, with no Anthropic staff review required.
Amodei's essay says evaluator access will have exceptions "to protect customers' and partners' private information". Our reading: more AI safety oversight means more watching of how models are used, so what you paste into any cloud tool matters more, not less.
AI Slowdown Timeline: Who Said What, and Where
Key statements from 9 to 16 September 2026, and where each was made.
| Date (2026) | Who | What they said, in a few words | Where |
|---|---|---|---|
| 9 Sept | Jacob Coxon | "I resigned from Anthropic today." | X |
| 12 Sept | Dario Amodei | "We Must Pace the Frontier" essay | darioamodei.com, X |
| 14 Sept | Donald Trump | AI doom "is a HOAX" | Truth Social |
| 14 Sept | China Foreign Ministry | Rejects "Fear-mongering" | Press conference |
| 15 Sept | Jensen Huang | "We don't need any new laws." | Salesforce Dreamforce |
| 16 Sept | Elizabeth Warren | "immediately press pause" | X |
What Happens Next in the AI Slowdown Fight?
A Trump-Xi meeting, reported by Fortune and others for 24 September, is the next test in the AI slowdown fight over the Dario Amodei essay.
- We found no formal statement from the White House or Commerce on pacing as of 17 September.
- The Blumenthal-Hawley AI Risk Evaluation Act was introduced in September 2025, reported by Axios, and no vote is scheduled.
- OpenAI said, "We'll have more to share soon."
- Watch who names evaluators first. Anthropic and OpenAI both promised outside evaluators but have not named them yet.
Keep your AI chats private, on your own Mac
Elephas pairs with the AI model you already use, or runs fully offline with built-in local LLM models, and redacts sensitive data before it ever leaves your Mac.
