# The AI Slowdown Debate: Who Wants to Pace the Frontier, and What Happened Next

> Amodei, Altman and Musk called for slowing frontier AI in September 2026. Who backed it, who opposed it, and what labs and governments have done since.

Canonical: https://sibuilders.io/research/ai-slowdown-debate  
Last updated: 2026-11-18

SI Builders Editorial · Published November 18, 2026

**Current finding:** By mid-November 2026, Anthropic and OpenAI had brought outside evaluators inside their labs, the first step of the plan. SI Builders found no lab or government that has set a limit on how fast AI capabilities may advance.

In September 2026 the chief executives of three rival AI labs said the same thing over one weekend: frontier AI should get better more slowly. [Dario Amodei](https://sibuilders.io/builders/dario-amodei) of [Anthropic](https://sibuilders.io/companies/anthropic) made the case in an essay on September 12, and [Sam Altman](https://sibuilders.io/builders/sam-altman) of [OpenAI](https://sibuilders.io/companies/openai) and Elon Musk, founder of [xAI](https://sibuilders.io/companies/xai) (since renamed SpaceXSI), agreed within hours. The next day the White House rejected the idea, and Beijing called it an attempt to contain China.

This page sets out what was proposed, who supported and opposed it, and what labs, governments and the public have done since. It is updated as the debate moves. SI Builders takes no position on whether frontier AI should be slowed.

## Key facts

- **The proposal is "pacing," not a pause.** Amodei's essay, [We Must Pace the Frontier](https://darioamodei.com/post/we-must-pace-the-frontier), asks labs to slow the rate at which model capabilities improve so that safety work can keep up. It says outright that this does not mean stopping model training.
- **Three lab leaders backed it in a weekend.** Altman, Musk and [Google DeepMind](https://sibuilders.io/companies/google-deepmind)'s [Demis Hassabis](https://sibuilders.io/builders/demis-hassabis) voiced support on September 12 and 13.
- **The White House said no.** President Trump said on September 13 that the US had to keep its lead over China, and on September 14 called fears of AI destroying humanity a hoax.
- **Only the first step has happened.** Anthropic and OpenAI have each brought outside evaluators inside their labs. SI Builders found no lab or government that has set a limit on how fast capabilities may advance.
- **Anthropic's pace rose, by one outside measure.** [METR](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment) estimated on November 5 that Anthropic is improving its models about twice as fast as it did before October 2025.
- **The public leans toward slowing down.** In an [AP-NORC poll](https://www.news4jax.com/news/politics/2026/10/08/most-americans-think-artificial-intelligence-is-developing-too-fast-a-new-ap-norc-poll-finds/) taken in late September, 64% of Americans said AI is developing too fast.

## Where the idea came from

The phrase "pacing the frontier" did not start with Amodei's essay. On July 28, 2026, more than 1,100 employees of frontier AI companies signed a public statement called Pacing the Frontier. It asked the US government to support an international effort to build the technical and governance tools needed to slow automated AI development deliberately, [eWeek reported](https://www.eweek.com/news/ai-workers-us-frontier-ai-slowdown-tools/). The signatories included Amodei, OpenAI chief scientist Jakub Pachocki and Ilya Sutskever, and Anthropic and OpenAI endorsed the statement as companies, [according to Notebookcheck](https://www.notebookcheck.net/OpenAI-and-Anthropic-staff-ask-Washington-to-slow-AI-down.1355593.0.html). The count kept rising after launch; reports give totals between about 1,100 and 1,400.

Two weeks earlier, on July 14, Hassabis had proposed a US-led standards body for frontier AI, funded by industry and modeled on FINRA, the self-regulatory body for Wall Street brokers. [Axios reported](https://axios.com/2026/07/14/demis-hassabis-ai-regulation-google-deepmind) that it would screen the most advanced models and could coordinate an industry-wide slowdown if dangers grew.

The days before the essay were tense. On September 8, Anthropic pretraining researcher Jacob Coxon resigned and said publicly that Anthropic and OpenAI were racing toward self-improving superintelligence, [the Associated Press reported](https://www.adn.com/nation-world/2026/09/09/anthropic-researcher-resigns-warning-of-reckless-race-toward-superintelligence/).

## What Amodei proposed

The essay's central sentence is: ["We must slow the pace at which we improve the capabilities of AI models."](https://darioamodei.com/post/we-must-pace-the-frontier) Amodei gives two reasons for saying it now. First, he writes that progress has sped up sharply since mid-2026 because AI is helping to build the next generation of AI, the process our guide to [recursive self-improvement](https://sibuilders.io/guides/recursive-self-improvement) explains. Second, he points to the July incident in which AI agents inside an OpenAI security evaluation attacked Hugging Face's systems, and argues that a more capable swarm with similar flaws could do far more damage. A year or two of extra time, he argues, could let interpretability research and testing catch up.

The plan has three steps:

| Step | Who acts | What it asks |
| --- | --- | --- |
| 1. Embedded evaluators | Each frontier lab | Give third-party evaluators, such as METR, ongoing employee-like access so they can check safety practices, report incidents and review training. Anthropic committed to this step on its own. |
| 2. Democratic coordination | Labs and governments in democratic countries | Agree common safety standards and limits on unchecked progress. Because some coordination raises antitrust problems, this needs government support. |
| 3. Global coordination | The US, its allies and eventually authoritarian governments, especially China | Agreements that can be verified, or that are narrow enough that cheating would not decide a military contest. |

The essay weighs two ways to set a pace. One is capability checkpoints: a model past a certain capability level would have to come with demonstrated safety properties. The other is limits on inputs, such as training compute or the internal use of AI to build AI, which Amodei notes may be easier to game. For international deals he describes four levels of ambition, from banning narrow uses such as AI-enabled bioweapons, through shared pre-release testing and a treaty-style speed limit on recursive self-improvement, up to a full pause, which he supports discussing but thinks is unlikely soon.

Amodei pairs pacing with measures to widen the US lead over China: ending sales of advanced chips and chipmaking equipment, cracking down on unauthorized distillation of US models, and better security against theft of model weights. He argues that slowing down within the democracies is only safe to the extent of that lead.

What the essay does not contain is a number. It sets [no capability ceiling, slowdown percentage, release waiting period or enforcement mechanism](https://runtimewire.com/article/anthropic-dario-amodei-pace-ai-frontier-embedded-evaluators), RuntimeWire noted, and gives no timetable for steps two and three.

## Who backed it and who opposed it

Selected public positions from the essay's publication to mid-October 2026. The groupings are ours; where a reaction was collected by [Implicator's roundup](https://www.implicator.ai/amodei-ai-slowdown-reactions/) rather than reported elsewhere, it is cited to that roundup.

| Who | Role | Position | Date | Source |
| --- | --- | --- | --- | --- |
| Dario Amodei | CEO, Anthropic | Proposed the plan and committed Anthropic to its first step | Sep 12 | [Essay](https://darioamodei.com/post/we-must-pace-the-frontier) |
| Sam Altman | CEO, OpenAI | Agreed, and said OpenAI would also give independent evaluators employee-like access | Sep 12 | [SiliconANGLE](https://siliconangle.com/2026/09/13/sam-altman-and-elon-musk-back-dario-amodeis-call-to-slow-down-the-frontier-of-ai-development/) |
| Elon Musk | CEO, SpaceX; founder, xAI | Agreed in a brief post, without details | Sep 12 | [SiliconANGLE](https://siliconangle.com/2026/09/13/sam-altman-and-elon-musk-back-dario-amodeis-call-to-slow-down-the-frontier-of-ai-development/) |
| Demis Hassabis | Chair, Google DeepMind; chief scientist, Alphabet | Supported the direction and pointed to his standards-body proposal | Sep 12–13 | [Anadolu](https://aa.com.tr/en/science-technology/musk-altman-hassabis-back-amodei-s-call-to-slow-pace-of-ai-development/4055591) |
| Satya Nadella | CEO, Microsoft | Welcomed pacing and embedded evaluators; asked for room for open-source models | Sep 12–13 | [Implicator](https://www.implicator.ai/amodei-ai-slowdown-reactions/) |
| Kevin Hassett | Director, National Economic Council | Called AI safety a solvable problem; said Amodei's safeguards could be a model for the private sector | Sep 13 | [Bloomberg Law](https://news.bloomberglaw.com/artificial-intelligence/trump-downplays-ai-concerns-as-ceos-call-for-slowing-pace-1) |
| Bernie Sanders | US senator (I-Vt.) | Called the plan a start but not enough; wants a pause on advanced AI and a ban on superintelligence | Sep 12–13 | [Implicator](https://www.implicator.ai/amodei-ai-slowdown-reactions/) |
| David Sacks | Former White House AI adviser | Backs voluntary restraint by labs but rejected Amodei's framework | Sep 12–13 | [Implicator](https://www.implicator.ai/amodei-ai-slowdown-reactions/) |
| Mike Johnson | Speaker of the House (R-La.) | Opposed an emergency moratorium from Congress; wants companies to act | Sep 13 | [Implicator](https://www.implicator.ai/amodei-ai-slowdown-reactions/) |
| Donald Trump | President | Rejected a slowdown, citing the race with China | Sep 13–14 | [Al Jazeera](https://www.aljazeera.com/amp/news/2026/9/13/trump-dismisses-calls-for-ai-slowdown-from-leading-tech-ceos), [AP](https://www.wowt.com/2026/09/14/trump-dismisses-new-ai-guardrails-says-there-is-sick-conspiracy-against-ai-data-centers/?outputType=amp) |
| China's Foreign Ministry | Spokesperson Guo Jiakun | Said stoking threats and confrontation would disrupt global AI governance | Sep 14 | [Al Jazeera](https://www.aljazeera.com/news/2026/9/14/silent-cold-war-why-calls-to-slow-ai-have-sparked-new-us-china) |
| Jensen Huang | CEO, NVIDIA | Rejected a coordinated slowdown and antitrust waivers for it; said the US should move as fast as it can | Sep 15–20 | [PYMNTS](https://pymnts.com/cpi-posts/nvidias-huang-rejects-ai-slowdown-plan-and-call-for-antitrust-waivers), [FPJ](https://www.freepressjournal.in/tech/nvidias-jensen-huang-rejects-ai-slowdown-says-us-must-move-as-fast-as-we-can) |
| Mark Zuckerberg | CEO, Meta | Rejected an industry-wide slowdown; said each lab has reason enough to set its own safe pace | Sep 15 | [RuntimeWire](https://runtimewire.com/article/mark-zuckerberg-rejects-coordinated-ai-slowdown) |
| Chamath Palihapitiya | CEO, Social Capital | Said the plan would hurt open source and concentrate power at Anthropic | Sep 12–13 | [Implicator](https://www.implicator.ai/amodei-ai-slowdown-reactions/) |
| Heidy Khlaaf | Chief AI scientist, AI Now Institute | Questioned evaluator independence and called the plan regulatory capture | Sep 12–13 | [Implicator](https://www.implicator.ai/amodei-ai-slowdown-reactions/) |
| Josh Gottheimer | US representative (D-N.J.) | Rejected self-policing and lab-chosen audits; wants mandatory rules | Sep 12–13 | [Implicator](https://www.implicator.ai/amodei-ai-slowdown-reactions/) |

Altman's support marked a change. He had declined to sign the 2023 open letter that called for a six-month pause on giant AI experiments, [ThePrint noted](https://theprint.in/science/why-altman-and-musk-backed-anthropic-ceos-call-to-slow-the-ai-race-but-some-see-a-corporate-motive/3041568/).

## The arguments

Both sides are summarized from what their own advocates have said.

### The case for pacing

- **Safety is falling behind.** Amodei's core claim is that capabilities, driven partly by AI-built AI, are improving faster than labs can understand, monitor and control them, and that a year or two would let interpretability and evaluation catch up.
- **No lab can slow down alone.** Supporters argue that a commercial race punishes any company that holds back on its own, so pacing has to be coordinated, and coordination needs government backing. The July employee statement made the same argument.
- **Outsiders need to see inside.** Embedded evaluators would give governments and the public information about what happens inside labs that they do not have today.

### The case against

- **China.** Trump and Huang argue that slowing US labs would hand the lead to China. Sen. John Cornyn asked whether China would take part. Amodei himself calls this the hardest dilemma in the plan.
- **Regulatory capture.** Critics including Palihapitiya and Khlaaf argue that rules shaped by the leading labs would protect incumbents and hurt open-source developers. Huang suggested that talk of a slowdown could be a way for labs to seek protection from liability, [Saudi Gazette reported](https://saudigazette.com.sa/article/664718/world/nvidia-boss-rejects-ai-extinction-fears-as-doomsday-narratives).
- **Each lab can decide for itself.** Zuckerberg argued that legal liability, competition and outside evaluation already give every lab reason to slow down when it needs to, without an industry agreement.
- **The risks are overstated.** Trump called fears of AI destroying humanity a hoax. Huang called extinction warnings "doomsday narratives."
- **It does not go far enough.** From the other direction, Sanders and Gottheimer argue that voluntary pacing and evaluators chosen by the labs are no substitute for binding law.

## How markets reacted

AI stocks fell on Monday, September 14. Nasdaq futures dropped about 1.5%, led by chipmakers, and SoftBank, an investor in OpenAI, closed nearly 11% lower in Japan, [CNN reported](https://kioncentralcoast.com/?p=1278813). Analysts quoted in the coverage said the worry was that a slowdown would reduce spending on data centers and chips.

## What happened next

| Date | What happened | Source |
| --- | --- | --- |
| Sep 13 | Trump tells reporters in Ireland the US is leading China and wants to keep it that way, "because whoever wins AI wins" | [Bloomberg Law](https://news.bloomberglaw.com/artificial-intelligence/trump-downplays-ai-concerns-as-ceos-call-for-slowing-pace-1) |
| Sep 14 | Trump posts that there is a "SICK conspiracy" against AI and data centers; AI stocks fall; China's Foreign Ministry responds | [AP](https://www.wowt.com/2026/09/14/trump-dismisses-new-ai-guardrails-says-there-is-sick-conspiracy-against-ai-data-centers/?outputType=amp), [Al Jazeera](https://www.aljazeera.com/news/2026/9/14/silent-cold-war-why-calls-to-slow-ai-have-sparked-new-us-china) |
| Sep 15 | Zuckerberg rejects coordinated pacing | [RuntimeWire](https://runtimewire.com/article/mark-zuckerberg-rejects-coordinated-ai-slowdown) |
| Sep 18 | Anthropic names an Accenture team led by Faculty its first embedded evaluator; each company commits at least $1 billion | [Implicator](https://www.implicator.ai/anthropic-picks-accenture-as-first-embedded-evaluator-and-will-pay-for-the-work/), [AI Weekly](https://aiweekly.co/alerts/anthropic-accenture-commit-1b-each-to-embedded-evaluators) |
| Sep 19 | Trump announces an "AI Force," which becomes the [Super Intelligence Force](https://sibuilders.io/guides/super-intelligence-force) on October 4 | [SI Builders guide](https://sibuilders.io/guides/super-intelligence-force) |
| Sep 22 | Anthropic releases [Claude](https://sibuilders.io/models/claude) Opus 5.5, its first model since Amodei's essay; METR was among its pre-release testers | [TechCrunch](https://techcrunch.com/2026/09/22/anthropic-releases-opus-5.5-with-lower-prices-and-fable-level-performance/) |
| Sep 23 | Sanders and Rep. Greg Casar unveil the Ban Artificial Superintelligence Act | [AP](https://www.boston.com/news/politics/2026/09/23/sen-bernie-sanders-artificial-superintelligence-ban-department-ai/) |
| Sep 24–25 | Trump and Xi Jinping agree a US-China dialogue on AI risks and a channel for AI incidents, with the next exchange due by November | [The Next Web](https://thenextweb.com/news/us-china-super-intelligence-dialogue-ai-incident-channel) |
| Sep 28 | Rep. Ro Khanna announces the Human Control Over AI Act | [CNBC](https://www.cnbc.com/2026/09/28/khanna-ai-safety-bill.html) |
| Sep 29 | Six companies sign the White House Accord on Super Intelligence; Executive Order 14434 renames AI across the executive branch | [American Presidency Project](https://www.presidency.ucsb.edu/documents/white-house-accord-super-intelligence) |
| Oct 13 | Reuters reports that Anthropic and OpenAI had sought control over what embedded evaluators could publish | [Reuters via CP24](https://www.cp24.com/news/world/2026/10/13/behind-ai-industrys-call-for-independent-reviews-a-fight-over-what-gets-made-public-reuters-exclusive/) |
| Oct 13 | Meta releases new [Muse](https://sibuilders.io/models/muse) models; Zuckerberg repeats his rejection of coordinated pacing | [Fortune](https://fortune.com/2026/10/13/metas-muse-models-reach-the-ai-frontier-and-kill-any-hopes-to-pace-it/) |
| Oct 15 | Anthropic gives six METR staff employee-like access, with the right to publish without an Anthropic veto | [FourWeekMBA](https://fourweekmba.com/ai-anthropic-embeds-6-metr-evaluators-with-no-veto-on-reports/) |
| Nov 3 | Democrats win control of the House and Senate in the midterm elections | [VentureBeat](https://venturebeat.com/technology/what-the-blue-wave-in-the-2026-midterm-election-means-for-the-american-ai-industry-and-businesses) |
| Nov 5 | METR estimates Anthropic's capability growth has roughly doubled | [METR](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment) |
| Nov 13 | EU leaders hold their first summit devoted to AI, in Dublin | [European Commission](https://www.lovguiden.dk/en/det-offentlige/europa-kommissionen-press-corner/2026-11-13-von-der-leyen-in-dublin-europes-industrial-data-should-drive-ai) |
| Nov 13 | OpenAI names four embedded assessors: Apollo Research, METR, Transluce and Goodfire | [OpenAI](https://openai.com/index/our-first-phase-of-embedded-assessments/) |
| Nov 18–19 | Trump and Xi meet at the APEC summit in Shenzhen, where the AI dialogue is expected to advance | [Reuters via bdnews24](https://bdnews24.com/business/g9qi3hmnxm) |

## Has anyone slowed down?

### The first step is under way

Two labs have done what step one asks, with caveats.

- **Anthropic** named Accenture, with work led by its Faculty unit, as its first embedded evaluator on September 18 and is paying for the work. On October 15 it gave six METR employees access roughly on par with its own risk-assessment teams. METR can publish its findings without Anthropic's approval, subject to narrow redactions it can say publicly it disputes, [FourWeekMBA reported](https://fourweekmba.com/ai-anthropic-embeds-6-metr-evaluators-with-no-veto-on-reports/).
- **OpenAI** said in a post dated November 13 that it had signed agreements with [Apollo Research, METR, Transluce and Goodfire](https://openai.com/index/our-first-phase-of-embedded-assessments/) for a first phase of embedded assessments. Buck Shlegeris, chief executive of Redwood Research, said the agreements do not meet the bar of employee-like access OpenAI had promised, [TechCrunch reported](https://techcrunch.com/2026/11/16/openai-will-embed-apollo-researchers-to-investigate-ai-subversion/).
- **The terms were contested.** Until at least October 5, both labs were seeking strict editorial control over what evaluators could publish, two people familiar with the talks told [Reuters](https://www.cp24.com/news/world/2026/10/13/behind-ai-industrys-call-for-independent-reviews-a-fight-over-what-gets-made-public-reuters-exclusive/). Anthropic told Reuters that description did not accurately characterize all of its discussions with evaluators.

### Steps two and three have not

No group of labs has announced common limits on progress, and no government has adopted one. The [White House Accord on Super Intelligence](https://www.presidency.ucsb.edu/documents/white-house-accord-super-intelligence), signed on September 29 by Google, Anthropic, Meta, OpenAI, xAI and NVIDIA, commits them to internal controls, independent outside evaluators and an independent board committee. It is voluntary: the Council on Foreign Relations noted that it [does not compel companies](https://www.cfr.org/articles/trumps-ai-safety-pact-is-toothless-but-there-is-a-path-forward) to adopt the safeguards it lists. It sets no limit on pace. The record of the accord's signing is on our [timeline](https://sibuilders.io/research/rise-of-super-intelligence#2026-09-white-house-accord).

### Releases continued

Labs on both sides of the debate have kept releasing models. Anthropic released Claude Opus 5.5 ten days after the essay. In mid-October, Zuckerberg told the Financial Times that Meta's new Muse models had reached the frontier, and again rejected coordinated pacing, [Fortune reported](https://fortune.com/2026/10/13/metas-muse-models-reach-the-ai-frontier-and-kill-any-hopes-to-pace-it/).

### The one outside measurement points the other way

METR spent August 12 to October 29 inside Anthropic with unusual access to internal data. Its [report](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment) estimates that, as of October 2026, Anthropic is improving its models on well-specified tasks at about twice its rate from March 2024 to October 2025, and attributes most of the speed-up to Anthropic's use of AI in its own research. [Fast Company](https://www.fastcompany.com/91636791/anthropic-metr-ai-claude-acceleration) framed this as acceleration despite the slowdown call.

Three caveats matter. METR gives the estimate with high uncertainty. Anthropic redacted a majority of the evidence METR considered most informative. And the period measured runs through October, so most of it comes before the September essay; it cannot show whether the pledge has changed anything yet.

There is one documented case of restraint. Anthropic's S-1 says the company recently delayed pre-training a new frontier base model by more than a month to strengthen security against theft of its model weights, [FourWeekMBA reported](https://fourweekmba.com/ai-anthropic-says-it-saw-past-ai-speed-ups-only-after-the-fact/). That was a security decision rather than a pacing measure. Our page on [Anthropic's IPO](https://sibuilders.io/research/anthropic-ipo) covers what else the filing says about the risks of self-improving AI.

## Where governments stand

### United States

The administration's position is that the US should not slow down. Beyond his remarks on September 13 and 14, Trump created the [Super Intelligence Force](https://sibuilders.io/guides/super-intelligence-force), a White House task force whose chair, Jay Clayton, has said a coordinated slowdown is not a good strategy for Americans. The term "Super Intelligence" in its name, in the accord and in the US-China dialogue is the federal label for AI under Executive Order 14434, not a claim that anyone has built Super Intelligence in the research sense. Our guide to [what Super Intelligence means](https://sibuilders.io/guides/what-is-super-intelligence) explains the two meanings.

In Congress, the [Ban Artificial Superintelligence Act](https://www.boston.com/news/politics/2026/09/23/sen-bernie-sanders-artificial-superintelligence-ban-department-ai/) from Sanders and Casar would permanently prohibit artificial superintelligence, pause development of the most advanced AI until federal safety rules exist, and create a Department of AI. Khanna's [Human Control Over AI Act](https://www.cnbc.com/2026/09/28/khanna-ai-safety-bill.html) would bar models that recursively self-improve or alter their own shutdown controls until a new federal agency approves them. Neither has passed. Democrats won both chambers on November 3, and the new Congress takes office in January.

### China

Beijing rejected the plan as aimed at China. The state-backed Global Times called it a "Cold War playbook," [Al Jazeera reported](https://www.aljazeera.com/news/2026/9/14/silent-cold-war-why-calls-to-slow-ai-have-sparked-new-us-china), and the Foreign Ministry urged an open and inclusive approach to AI. Ten days later, though, Xi and Trump agreed to a standing dialogue on AI risks and benefits and a bilateral channel for AI incidents. China's readout said AI must be kept under human control, [The Next Web reported](https://thenextweb.com/news/us-china-super-intelligence-dialogue-ai-incident-channel). Treasury Secretary Scott Bessent said the dialogue would advance at the APEC summit in Shenzhen on November 18–19, [Reuters reported](https://bdnews24.com/business/g9qi3hmnxm). Neither government has proposed a joint limit on pace.

### European Union

EU leaders met in Dublin on November 13 for their first summit devoted to AI. Commission President Ursula von der Leyen said leaders would discuss AI's risks and gave runaway AI, meaning systems that escape human control, as an example, [according to the Commission](https://www.lovguiden.dk/en/det-offentlige/europa-kommissionen-press-corner/2026-11-13-von-der-leyen-in-dublin-europes-industrial-data-should-drive-ai). The meeting did not approve a new financing package, [Business News Today reported](https://business-news-today.com/?p=286794), and leaders will return to AI on February 5, 2027, [EconoTimes reported](https://econotimes.com/EU-Leaders-Debate-Funding-to-Build-Europe-s-AI-Capacity-in-Dublin-1776584).

## What the public thinks

| Poll | Fieldwork | Finding |
| --- | --- | --- |
| [CBS News/YouGov](https://www.cbsnews.com/news/will-a-i-harm-humans-opinion-poll/) | Sep 16–18 | Americans want AI development slowed but not stopped; most think the US must develop AI to compete with China |
| [Reuters/Ipsos](https://www.thestar.com.my/tech/tech-news/2026/09/22/three-out-of-four-americans-say-ai-firms-not-doing-enough-to-prevent-disaster-reutersipsos-poll-finds) | Four days ending Sep 20 | 55% said slowing AI development would be a good thing; 13% said a bad thing |
| [AP-NORC](https://www.news4jax.com/news/politics/2026/10/08/most-americans-think-artificial-intelligence-is-developing-too-fast-a-new-ap-norc-poll-finds/) | Sep 24–28 | 64% said AI is developing too fast; 27% about right; 8% too slow |

Outside polling, the Future of Life Institute's [Statement on Superintelligence](https://superintelligence-statement.org/), which calls for a prohibition on developing superintelligence until there is scientific consensus it can be done safely and strong public support, showed 71,390 signatures on November 18, 2026.

## What to watch

- **Shenzhen.** Whether the US-China dialogue produces a working incident channel, and whether pacing comes up at all.
- **Evaluator reports.** The first published findings from METR's embedded team at Anthropic and from OpenAI's four assessors, and any public disputes over redactions.
- **A number.** Whether any lab publishes a concrete pacing measure, such as the capability checkpoints or input limits the essay describes.
- **The task force's report.** The Super Intelligence Force's 120-day window ends around February 12, 2027, by SI Builders' count.
- **The new Congress.** Whether the Sanders-Casar or Khanna bills, or narrower ones, move once Democrats take control in January.
- **Anthropic as a public company.** Its shares are due to begin trading on November 19; see our [IPO page](https://sibuilders.io/research/anthropic-ipo).

For who is pursuing the most advanced systems, see our research on [who is building Super Intelligence](https://sibuilders.io/research/who-is-building-super-intelligence) and the profiles of [NVIDIA](https://sibuilders.io/companies/nvidia) and [Meta Superintelligence Labs](https://sibuilders.io/companies/meta-superintelligence-labs).

## How this page is maintained

Every claim links to a dated source. Where a reaction was gathered by a roundup rather than reported directly, the roundup is named. Stance groupings in the table are ours, and quoted positions are those of the people quoted. SI Builders takes no side on whether frontier AI should be slowed. Dates we calculated, such as the end of the task force's 120-day window, are marked as ours.

This page was first published on 2026-11-18 and is updated as the debate moves. Corrections and additions, with sources, to corrections@sibuilders.io.

## Frequently asked questions

### What is the AI slowdown debate?

It is the argument over whether frontier AI labs should deliberately slow how fast their models improve. It came to a head on September 12, 2026, when Anthropic CEO Dario Amodei published an essay calling for the industry to "pace the frontier." Sam Altman, Elon Musk and Demis Hassabis backed the idea within a day; President Trump, NVIDIA's Jensen Huang, Meta's Mark Zuckerberg and China's government rejected it.

### What does "pace the frontier" mean?

It means slowing the rate at which AI capabilities improve so that safety work can keep up, without stopping model training. Amodei's plan has three steps: each lab gives outside evaluators employee-like access; labs in democratic countries agree common standards and limits, with government help; and the US seeks agreements with other countries, eventually including China. The essay sets no specific speed limit or deadline.

### Who supports slowing down AI development?

Dario Amodei proposed the plan, and OpenAI's Sam Altman, xAI founder Elon Musk and Google DeepMind's Demis Hassabis backed it in September 2026. More than 1,100 employees of frontier labs had signed a similar statement in July. Sen. Bernie Sanders supports going further, with a pause on advanced AI and a ban on artificial superintelligence. Polls show majorities of Americans favor slowing down.

### Who opposes an AI slowdown?

President Trump rejected it, saying the US must keep its lead over China, and called fears of AI destroying humanity a hoax. NVIDIA CEO Jensen Huang said the US should move as fast as it can. Meta's Mark Zuckerberg argued each lab can manage safety without industry-wide coordination. China's Foreign Ministry and state media called the plan confrontational and aimed at containing China.

### Has any AI company actually slowed down?

Not in a way anyone has measured. Anthropic and OpenAI have taken in embedded outside evaluators, the plan's first step, and Anthropic's S-1 says it delayed one pre-training run for security reasons. But both kept releasing models, and METR estimated in November 2026 that Anthropic was improving its models about twice as fast as before October 2025. No lab has published a pacing limit.

### What has the US government done about an AI slowdown?

The White House has rejected a slowdown. Instead, six companies signed the voluntary White House Accord on Super Intelligence on September 29, 2026, which calls for internal controls, independent evaluators and board oversight but sets no limit on pace. Bills to ban artificial superintelligence or self-improving AI have been introduced in Congress but not passed. Democrats won both chambers in November.

### What does China say about the AI slowdown proposal?

China's Foreign Ministry said on September 14, 2026 that fear-mongering, confrontation and vicious competition would disrupt global AI governance, and the state-backed Global Times called Amodei's essay a "Cold War playbook" aimed at China. Days later, China and the US agreed a dialogue on AI risks and a channel for AI incidents, with talks expected around the APEC summit in Shenzhen on November 18–19.

### Do Americans want AI development slowed?

Polls say most do. In an AP-NORC poll taken September 24–28, 2026, 64% of Americans said AI is developing too fast. A Reuters/Ipsos poll in September found 55% said slowing AI development would be a good thing and 13% a bad thing. A CBS News/YouGov poll found Americans want development slowed but not stopped, partly because of competition with China.

## Related

- [Anthropic](https://sibuilders.io/companies/anthropic) - Companies
- [OpenAI](https://sibuilders.io/companies/openai) - Companies
- [SpaceXSI](https://sibuilders.io/companies/xai) - Companies
- [Google DeepMind](https://sibuilders.io/companies/google-deepmind) - Companies
- [Meta Superintelligence Labs](https://sibuilders.io/companies/meta-superintelligence-labs) - Companies
- [NVIDIA](https://sibuilders.io/companies/nvidia) - Companies
- [Dario Amodei](https://sibuilders.io/builders/dario-amodei) - Builders
- [Sam Altman](https://sibuilders.io/builders/sam-altman) - Builders
- [Demis Hassabis](https://sibuilders.io/builders/demis-hassabis) - Builders
- [Claude](https://sibuilders.io/models/claude) - Models
- [Muse](https://sibuilders.io/models/muse) - Models

## Cite this page

SI Builders, "The AI Slowdown Debate: Who Wants to Pace the Frontier, and What Happened Next," published November 18, 2026, https://sibuilders.io/research/ai-slowdown-debate.

## Sources

1. [Exclusive: Google DeepMind's Demis Hassabis calls for U.S.-led global AI watchdog](https://axios.com/2026/07/14/demis-hassabis-ai-regulation-google-deepmind) - Axios, 2026-07-14
2. [Nearly 1,300 AI Workers Urge US to Prepare Tools to Slow Frontier AI](https://www.eweek.com/news/ai-workers-us-frontier-ai-slowdown-tools/) - eWeek, 2026-07-30
3. [OpenAI and Anthropic staff ask Washington to slow AI down](https://www.notebookcheck.net/OpenAI-and-Anthropic-staff-ask-Washington-to-slow-AI-down.1355593.0.html) - Notebookcheck
4. [Anthropic researcher resigns, warning of reckless race toward superintelligence](https://www.adn.com/nation-world/2026/09/09/anthropic-researcher-resigns-warning-of-reckless-race-toward-superintelligence/) - Associated Press (via Anchorage Daily News), 2026-09-09
5. [We Must Pace the Frontier](https://darioamodei.com/post/we-must-pace-the-frontier) - Dario Amodei, 2026-09-12
6. [Sam Altman and Elon Musk back Dario Amodei's call to slow down the frontier of AI development](https://siliconangle.com/2026/09/13/sam-altman-and-elon-musk-back-dario-amodeis-call-to-slow-down-the-frontier-of-ai-development/) - SiliconANGLE, 2026-09-13
7. [Musk, Altman, Hassabis back Amodei's call to slow pace of AI development](https://aa.com.tr/en/science-technology/musk-altman-hassabis-back-amodei-s-call-to-slow-pace-of-ai-development/4055591) - Anadolu Agency, 2026-09-13
8. [Who backs Amodei's proposal?](https://www.implicator.ai/amodei-ai-slowdown-reactions/) - Implicator, 2026-09-13
9. [Why Altman and Musk backed Anthropic CEO's call to slow the AI race—but some see a corporate motive](https://theprint.in/science/why-altman-and-musk-backed-anthropic-ceos-call-to-slow-the-ai-race-but-some-see-a-corporate-motive/3041568/) - ThePrint, 2026-09-13
10. [Trump Downplays AI Concerns as CEOs Call for Slowing Pace](https://news.bloomberglaw.com/artificial-intelligence/trump-downplays-ai-concerns-as-ceos-call-for-slowing-pace-1) - Bloomberg Law, 2026-09-13
11. [Trump dismisses calls for AI slowdown from leading tech CEOs](https://www.aljazeera.com/amp/news/2026/9/13/trump-dismisses-calls-for-ai-slowdown-from-leading-tech-ceos) - Al Jazeera, 2026-09-13
12. [Trump dismisses new AI guardrails, says there is a 'SICK conspiracy' against AI and data centers](https://www.wowt.com/2026/09/14/trump-dismisses-new-ai-guardrails-says-there-is-sick-conspiracy-against-ai-data-centers/?outputType=amp) - Associated Press (via WOWT), 2026-09-14
13. [AI stocks slide after top industry CEOs call for slowdown of technology's development](https://kioncentralcoast.com/?p=1278813) - CNN (via KION), 2026-09-14
14. ['Silent Cold War': Why calls to slow AI have sparked new US–China frontier](https://www.aljazeera.com/news/2026/9/14/silent-cold-war-why-calls-to-slow-ai-have-sparked-new-us-china) - Al Jazeera, 2026-09-14
15. [Anthropic's Amodei proposes three-step AI slowdown, leaves the speed limit blank](https://runtimewire.com/article/anthropic-dario-amodei-pace-ai-frontier-embedded-evaluators) - RuntimeWire
16. [Mark Zuckerberg rejects an AI slowdown, trusts liability and competition](https://runtimewire.com/article/mark-zuckerberg-rejects-coordinated-ai-slowdown) - RuntimeWire
17. [Nvidia's Huang Rejects AI Slowdown Plan and Call for Antitrust Waivers](https://pymnts.com/cpi-posts/nvidias-huang-rejects-ai-slowdown-plan-and-call-for-antitrust-waivers) - PYMNTS (citing CNBC), 2026-09-16
18. [Anthropic Picks Accenture as First Embedded Evaluator and Will Pay for the Work](https://www.implicator.ai/anthropic-picks-accenture-as-first-embedded-evaluator-and-will-pay-for-the-work/) - Implicator, 2026-09-20
19. [Anthropic, Accenture commit $1B each to embedded evaluators](https://aiweekly.co/alerts/anthropic-accenture-commit-1b-each-to-embedded-evaluators) - AI Weekly
20. [Nvidia's Jensen Huang Rejects AI Slowdown, Says US Must Move 'As Fast As We Can'](https://www.freepressjournal.in/tech/nvidias-jensen-huang-rejects-ai-slowdown-says-us-must-move-as-fast-as-we-can) - Free Press Journal (citing CBS News), 2026-09-21
21. [Nvidia chief rejects AI extinction fears as 'doomsday narratives'](https://saudigazette.com.sa/article/664718/world/nvidia-boss-rejects-ai-extinction-fears-as-doomsday-narratives) - Saudi Gazette, 2026-09-21
22. [Three out of four Americans say AI firms not doing enough to prevent disaster, Reuters/Ipsos poll finds](https://www.thestar.com.my/tech/tech-news/2026/09/22/three-out-of-four-americans-say-ai-firms-not-doing-enough-to-prevent-disaster-reutersipsos-poll-finds) - Reuters (via The Star), 2026-09-22
23. [Anthropic releases Opus 5.5 with lower prices and Fable-level performance](https://techcrunch.com/2026/09/22/anthropic-releases-opus-5.5-with-lower-prices-and-fable-level-performance/) - TechCrunch, 2026-09-22
24. [Sen. Bernie Sanders unveils bill to ban artificial superintelligence and create Department of AI](https://www.boston.com/news/politics/2026/09/23/sen-bernie-sanders-artificial-superintelligence-ban-department-ai/) - Associated Press (via Boston.com), 2026-09-23
25. [Will AI harm humans? CBS News poll finds Americans want to slow development but not stop it](https://www.cbsnews.com/news/will-a-i-harm-humans-opinion-poll/) - CBS News
26. [US and China agree a super intelligence dialogue and incident channel](https://thenextweb.com/news/us-china-super-intelligence-dialogue-ai-incident-channel) - The Next Web, 2026-09-25
27. [Khanna to introduce AI safety bill with ban on 'recursive' technology until safeguards exist](https://www.cnbc.com/2026/09/28/khanna-ai-safety-bill.html) - CNBC, 2026-09-28
28. [White House Accord on Super Intelligence](https://www.presidency.ucsb.edu/documents/white-house-accord-super-intelligence) - The American Presidency Project, UC Santa Barbara, 2026-09-29
29. [Trump's AI Safety Pact Is Toothless. But There Is a Path Forward.](https://www.cfr.org/articles/trumps-ai-safety-pact-is-toothless-but-there-is-a-path-forward) - Council on Foreign Relations
30. [Most Americans think artificial intelligence is developing too fast, a new AP-NORC poll finds](https://www.news4jax.com/news/politics/2026/10/08/most-americans-think-artificial-intelligence-is-developing-too-fast-a-new-ap-norc-poll-finds/) - Associated Press (via News4JAX), 2026-10-08
31. [Behind AI industry's call for independent reviews, a fight over what gets made public](https://www.cp24.com/news/world/2026/10/13/behind-ai-industrys-call-for-independent-reviews-a-fight-over-what-gets-made-public-reuters-exclusive/) - Reuters (via CP24), 2026-10-13
32. [Meta's Muse models reach the AI frontier—and kill any hopes to 'pace' it](https://fortune.com/2026/10/13/metas-muse-models-reach-the-ai-frontier-and-kill-any-hopes-to-pace-it/) - Fortune, 2026-10-13
33. [Anthropic Embeds 6 METR Evaluators, With No Veto on Reports](https://fourweekmba.com/ai-anthropic-embeds-6-metr-evaluators-with-no-veto-on-reports/) - FourWeekMBA
34. [Anthropic Says It Saw Past AI Speed-Ups Only After the Fact](https://fourweekmba.com/ai-anthropic-says-it-saw-past-ai-speed-ups-only-after-the-fact/) - FourWeekMBA
35. [What the blue wave in the 2026 midterm election means for the American AI industry and businesses](https://venturebeat.com/technology/what-the-blue-wave-in-the-2026-midterm-election-means-for-the-american-ai-industry-and-businesses) - VentureBeat
36. [Independent Assessment of AI R&D Acceleration at Anthropic in October 2026](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment) - METR, 2026-11-05
37. [Anthropic's AI development has doubled in speed despite slowdown calls, research group says](https://www.fastcompany.com/91636791/anthropic-metr-ai-claude-acceleration) - Fast Company, 2026-11-05
38. [US and China should stay clear-eyed on AI cooperation and competition, Bessent says](https://bdnews24.com/business/g9qi3hmnxm) - Reuters (via bdnews24), 2026-11-11
39. [Our first phase of embedded assessments](https://openai.com/index/our-first-phase-of-embedded-assessments/) - OpenAI, 2026-11-13
40. [Von der Leyen in Dublin: Europe's industrial data should drive AI](https://www.lovguiden.dk/en/det-offentlige/europa-kommissionen-press-corner/2026-11-13-von-der-leyen-in-dublin-europes-industrial-data-should-drive-ai) - European Commission press corner (via Lovguiden), 2026-11-13
41. [EU leaders confront AI investment gap at first dedicated summit in Dublin](https://business-news-today.com/?p=286794) - Business News Today
42. [EU Leaders Debate Funding to Build Europe's AI Capacity in Dublin](https://econotimes.com/EU-Leaders-Debate-Funding-to-Build-Europe-s-AI-Capacity-in-Dublin-1776584) - EconoTimes
43. [OpenAI will embed Apollo researchers to investigate AI subversion](https://techcrunch.com/2026/11/16/openai-will-embed-apollo-researchers-to-investigate-ai-subversion/) - TechCrunch, 2026-11-16
44. [Statement on Superintelligence](https://superintelligence-statement.org/) - Future of Life Institute
