Demis Hassabis has backed the idea that the companies building the most powerful artificial intelligence should deliberately slow down. On Saturday 12 September 2026, just under nine hours after Anthropic chief executive Dario Amodei published his essay “We Must Pace the Frontier”, the Google DeepMind co-founder quote-posted it on X. “Dario’s essay points towards the right path forward,” he wrote. “The details need working through, but the direction is correct for meeting this critical moment.”
His third sentence drew less attention, and it matters most: “This is also why we recently put out our proposal for an industry-wide standards body for frontier AI.” The agreement runs both ways. Amodei’s essay names “the mechanism suggested by Demis Hassabis” as one route for companies to set common safety standards, and links to the framework Hassabis published in July. That framework already allowed for “coordinating a slowdown in development among the Frontier Labs if deemed necessary”. Two founders of rival labs have now published overlapping plans for pacing new AI models, 60 days apart.
This article compares the two documents, traces the Saturday of endorsements that preceded the reply from Demis Hassabis, and checks what “the direction is correct” commits Google to. So far, it is less than OpenAI has promised. Our earlier coverage reads the essay in full in pacing the frontier, follows Elon Musk’s endorsement, weighs the call to slow model development, and measures how far AI safety measures have fallen behind.
Table of contents
- What Demis Hassabis Said About Amodei’s Essay
- The Plan Demis Hassabis Published in July
- Where Demis Hassabis and Dario Amodei Agree
- Where Demis Hassabis and Amodei Still Differ
- The Hours Before Demis Hassabis Replied
- What Demis Hassabis Committed Google To
- Demis Hassabis Now Speaks as Chair, Not Chief Executive
- The Safety Record Behind Demis Hassabis’s Lab
- Why Demis Hassabis Has Long Wanted a Slower Race
- The Case Against Demis Hassabis’s Standards Body
- China and the Global Test for Demis Hassabis’s Idea
- What Happens Next for Demis Hassabis and the Pacing Plan
- What Demis Hassabis’s Endorsement Means for Businesses
- Frequently Asked Questions About Demis Hassabis and Pacing
- References
What Demis Hassabis Said About Amodei's Essay
Demis Hassabis posted at 22:59 UTC on 12 September, one minute before midnight in London. His post quoted Amodei’s own announcement, which had passed 62.4 million views when we checked it on 13 September. The Hassabis post itself had about 1.07 million views, 7,745 likes, 796 reposts and 1,405 bookmarks at that point, according to the public post data.
Three sentences and a link
The whole post is three sentences plus a link. The first two praise the direction of travel rather than the detail. The third links to “A Framework for Frontier AI and the Dawning of a New Age”, the essay Demis Hassabis published on X and on his Substack on 14 July. By pointing there, he presented Amodei’s plan as a close relative of his own rather than as a new idea he was adopting.
“The details need working through”
That caveat separates Demis Hassabis from the other endorsements. Elon Musk wrote three words, “Dario is right”. Sam Altman went further and made a commitment: OpenAI would copy Anthropic’s first step and give independent evaluators employee-like access. Hassabis agreed with the goal, reserved judgement on the mechanics, and offered his own mechanism. It is the response of someone who has already drafted a plan and expects to argue over which version wins.
Chair, chief executive or former chief executive
Coverage disagreed about his title. Anadolu Agency and the syndicated report carried by Saudi Gazette called him “Google DeepMind CEO”. CNBC called him “the chair of Google DeepMind”. Hindustan Times used “co-founder and chair of Google DeepMind and Chief Scientist of Alphabet”, matching his X profile, while The Decoder wrote “Ex-Deepmind-CEO”. The difference matters. Demis Hassabis stepped down as chief executive on 5 August, which changes what his endorsement can deliver, as a later section explains.
Late to a busy Saturday
Business Insider’s first report on the endorsements, published at 17:21 UTC, noted that Hassabis “has yet to weigh in on Saturday”. It reached back to his July essay instead, quoting his call for public policy that “fosters international collaboration on key safety issues”. Nearly six hours later, the Google DeepMind chair answered in person.
The Plan Demis Hassabis Published in July
The document Demis Hassabis pointed to runs to about 1,350 words. It opens with the claim that Artificial General Intelligence, “a system that exhibits all the cognitive capabilities the brain has, is probably only a few short years away”, and describes the present as “the foothills of the singularity”. By 13 September the X version had been viewed about 15.8 million times and bookmarked 31,017 times.
The problem it describes
Hassabis wrote that “advances on the frontier are outpacing our understanding of the technology” and that the field is “locked in an extremely intense, multilayered commercial and geopolitical race”. Mitigating the technical risks is achievable, he argued, “but only if we give ourselves the time and space to get this next crucial step right. Currently, as a field and as a wider society, we aren’t doing that.”
A standards body modelled on FINRA
His answer is an institution. The US “could establish a new Standards Body modelled on a federally overseen public-private partnership or self-regulatory organisation, much like the Financial Industry Regulatory Authority (FINRA), with a board that includes independent leading technical experts and open-source representatives”. FINRA oversees US brokerage firms. Funding “would need to be substantial and likely mostly come from industry”, so the body could hire world-class technical staff and buy compute for large-scale testing.
Thirty days before release
The body would develop assessment protocols and work with federal agencies and the US National Labs to test for risks relevant to national security. “Initially, Frontier Labs would voluntarily share models with the Standards Body for review up to 30 days before release.” Once the protocol proved “effective and robust”, passing it would become a condition of deployment in the US market. Labs would also work with the body on critical vulnerabilities found after release.
Who counts as a Frontier Lab
A model would qualify as “Frontier-class” by crossing benchmark thresholds that the body sets and regularly updates, and its evaluations would be refreshed “perhaps quarterly to start”. Any organisation with such a model becomes a “Frontier Lab”, encouraged to publish model cards, keep strong internal cybersecurity, vet key personnel and fund safety research. The framework could apply “no matter their country of origin or whether they are open or closed”, while non-frontier models from startups and academia would be exempt.
The slowdown clause
The line that connects Demis Hassabis most directly to Amodei’s essay comes near the end. The approach “could be ratcheted up if the seriousness of the situation demands, including coordinating a slowdown in development among the Frontier Labs if deemed necessary”. In July that read as a remote contingency. Two months later, the chief executive of Anthropic argued that the contingency has arrived.
Where Demis Hassabis and Dario Amodei Agree
Put side by side, the July framework and the September essay share a diagnosis, a preference for independent testing, and a willingness to use coordination between rivals as a safety tool. Neither calls for a permanent halt. Both borrow their model of oversight from finance. The overlap is wide enough that Amodei’s essay cites the Hassabis mechanism by name.
The same diagnosis
Both men describe competition outrunning understanding. Hassabis wrote of a race in which advances “are outpacing our understanding”. Amodei wrote that “a race to the bottom, spurred by commercial incentives, can make these risks more acute”. Both single out recursive self-improvement. Hassabis called for “robust safeguards to maintain control of increasingly agentic, recursively self-improving systems”, and Amodei said the dynamic “could outrun our ability to understand and control these systems”.
Testing has to get harder as models get smarter
Demis Hassabis asked for testing that is “dynamic, adaptable, and rigorous”, including agentic tests that “look for attempts to bypass safety guardrails or signs of deception”. Amodei warned that “more intelligent models are more capable of deceiving tests, and thus may appear aligned while having serious problems that go undetected”. Both want evaluations that the labs themselves do not fully control.
Rivals setting standards together
Amodei’s second step is “Democratic Coordination”: frontier companies in democratic countries “coordinate to establish common safety standards as well as limits on the rate of unchecked AI progress”. Because that raises antitrust questions, he wants the US government to “mediate or at least enable these discussions”. Such dialogue “could also happen through industry groups that have some association with government — for example, the mechanism suggested by Demis Hassabis”. That hyperlink is what joins the two documents.
Borrowed from financial regulation
Demis Hassabis modelled his body on FINRA. Amodei justified embedded evaluators by pointing to banking, “which sometimes involves regulatory ‘supervisors’ embedded along with employees”. Both chose a supervised-industry analogy over a product-approval one, although the Hassabis end state, in which frontier models must pass an assessment before deployment in the US, starts to resemble approval.
A slowdown as a tool, not a stop
Amodei wrote that “pacing does not mean halting model training or technical progress, but ensuring companies take adequate time to align and safeguard their models”. Hassabis kept a coordinated slowdown in reserve for when it is “deemed necessary”. Both treat speed as something to manage, and both expect progress to continue.
| Theme | Demis Hassabis, 14 July | Dario Amodei, 12 September |
|---|---|---|
| Diagnosis | Frontier advances “outpacing our understanding” | “A race to the bottom, spurred by commercial incentives” |
| Self-improving systems | Safeguards for “recursively self-improving systems” | Could “outrun our ability to understand and control” |
| Testing | “Dynamic, adaptable, and rigorous”, with deception tests | Models “more capable of deceiving tests” |
| Outside checks | “An ecosystem of third-party auditors” | Embedded evaluators “such as METR” |
| Model from finance | A FINRA-style self-regulatory body | Banking-style embedded “supervisors” |
| Slowdown | Coordinated “if deemed necessary” | The goal, but “does not mean halting model training” |
Where Demis Hassabis and Amodei Still Differ
“The details need working through” is a polite way of saying the two plans are not the same. They differ on where outside reviewers sit, what triggers a slowdown, who moves first and how far the rules reach. Those details decide whether a pacing regime would bind anyone.
At the door or inside the building
The Hassabis framework reviews models up to 30 days before release. Amodei’s evaluators would move in. Anthropic promises the external team “desks in our offices, access badges, and company laptops”, permissions “mostly comparable” to internal risk teams, and the right to publish findings “without editorial control by Anthropic”. It may redact narrow categories of information, “but we can’t redact findings just because they are unfavorable”. Evaluators would assess training pipelines and processes, not only finished models.
What triggers a slower pace
Demis Hassabis defines the regulated tier by benchmark thresholds and keeps a slowdown as a last resort. Amodei favours “checkpoints”: if a model reaches capability X, it must come with certified alignment properties Y and Z. His example for X is a model “capable of escaping or defeating most common sandboxing methods”. He also floats limits on ingredients such as training compute, while warning that those may be “more ‘gameable'”.
Who moves first
Amodei’s first step is corporate and immediate. Anthropic is “unilaterally committing” to embedded evaluators and calls on governments “to require other frontier companies to match”. The first mover in the Hassabis plan is the state: “The US is well positioned, given its economic and technical standing, to take the first step.” One plan starts with a company volunteering; the other starts with Washington building an institution.
How far the rules reach
The Hassabis framework covers any frontier-class model “no matter their country of origin”, open or closed, with an exemption for non-frontier work. Amodei’s second step is limited to companies “within democratic countries”. His third step, “Global Coordination”, accepts that dealing with authoritarian governments means “taking seriously the challenges of verifying compliance”.
Tone and length
The documents also read differently. Hassabis ends on “a new golden age of scientific discovery and progress”. Amodei dwells on loss of control, cyberattacks and bioterrorism, and says flatly that companies “must slow the pace” at which capabilities improve. At about 3,750 words, the essay is nearly three times the length of the July framework.
| Question | Demis Hassabis framework | Amodei essay |
|---|---|---|
| Who reviews | A US Standards Body with independent experts and open-source seats | Third-party evaluators embedded inside each lab |
| When | Up to 30 days before release, plus post-release fixes | Ongoing, including during training |
| Access | Models shared with the body | Desks, badges, laptops and internal tools |
| Trigger | Benchmark thresholds define “Frontier-class” | Capability checkpoints tied to alignment evidence |
| First mover | The US government | Anthropic, unilaterally |
| Scope | Frontier models from any country, open or closed | Companies in democracies first, then global talks |
| Binding force | Voluntary at first, then required for the US market | Voluntary now; regulation called “the most effective method” |
The Hours Before Demis Hassabis Replied
Amodei announced his essay at 14:01 UTC. Over the next 16 hours, leaders of rival labs, a former UK prime minister, a US senator and a former White House adviser all responded in public. The order shows who moved quickly, who committed to something, and who used the moment to argue that pacing goes too far or not far enough.
A Saturday of endorsements
Elon Musk replied first, at 15:01 UTC: “Dario is right.” Sam Altman followed at 16:30, agreeing “that we need to pace the frontier” and promising evaluators. Rishi Sunak, a senior adviser at Anthropic, wrote at 17:38 that “when the labs themselves are asking for speed limits, governments should listen”. Senator Bernie Sanders said at 18:53 that pacing is “not enough” and “you hit the brakes”. Demis Hassabis posted at 22:59.
Sunday’s sceptics
David Sacks, who previously served as President Trump’s AI and crypto czar, answered at 03:14 UTC on Sunday. He told the labs to “go ahead” but to “stop pretending you need anyone else’s permission”. Musk returned at 04:07 to say “Peer review of AI by competitors is the right way to start this off”. Meta’s AI chief Alexandr Wang, at 06:03, said alignment “can be the gating factor for scaling” without addressing the essay directly.
| Who | Role | Posted (UTC) | After Amodei | Views by 13 Sep | Position |
|---|---|---|---|---|---|
| Elon Musk | Tesla and SpaceX chief executive | 12 Sep 15:01 | 1h 00m | 10.18m | “Dario is right” |
| Sam Altman | OpenAI chief executive | 12 Sep 16:30 | 2h 30m | 14.59m | Agrees; will add evaluators |
| Rishi Sunak | Former UK prime minister, Anthropic adviser | 12 Sep 17:38 | 3h 37m | 0.31m | Governments must “step up” |
| Bernie Sanders | US senator | 12 Sep 18:53 | 4h 52m | 1.14m | Wants a pause and a superintelligence ban |
| Peyman Milanfar | Google researcher | 12 Sep 21:32 | 7h 31m | 0.19m | Says a speed limit is unnecessary |
| Demis Hassabis | Google DeepMind co-founder and chair | 12 Sep 22:59 | 8h 59m | 1.07m | “The direction is correct” |
| David Sacks | Former White House AI adviser | 13 Sep 03:14 | 13h 14m | 5.05m | “Go ahead”, without new rules |
| Elon Musk | Second post | 13 Sep 04:07 | 14h 06m | 0.43m | Peer review by competitors |
| Alexandr Wang | Meta AI chief | 13 Sep 06:03 | 16h 02m | 0.23m | Alignment as “the gating factor” |
Demis Hassabis gave the sixth of the nine responses shown here, and the last to arrive on Saturday.
Why the nine-hour gap means little
The Hassabis post landed at 23:59 in London, so the delay is as likely to reflect a British Saturday evening as careful deliberation. What stands out is the content, not the timing. He was the only lab leader to answer by pointing to a proposal of his own, and the only one to endorse the direction while openly withholding agreement on the details.
Reach
The July framework has travelled much further than the endorsement. It had about 15.8 million views by 13 September, nearly 15 times the 1.07 million on Saturday’s post (15.79 ÷ 1.07 = 14.8). Amodei’s announcement, at 62.4 million views, dwarfed every response.
What Demis Hassabis Committed Google To
The most important difference between Saturday’s endorsements is what each one promised. Altman made a commitment on behalf of OpenAI. Musk offered agreement and, the next day, a preference for peer review by competitors. Demis Hassabis pointed to a proposal he had already published. None of the coverage reviewed for this article, including reports by CNBC, Hindustan Times, NDTV, Anadolu Agency and The Decoder, mentions a new Google DeepMind commitment to embedded evaluators.
A promise versus a pointer
Altman’s words were specific: “Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We’ll have more to share soon.” The Hassabis post contains no equivalent sentence. Its commitment, if there is one, is to a standards body that does not yet exist and depends on the US government to create it.
Who can commit Google DeepMind now
Since August the lab has been run by Koray Kavukcuoglu, its former chief technology officer, as senior vice president. He reports directly to Google chief executive Sundar Pichai and oversees Gemini model development, frontier AI research, and the Gemini app and developer teams, CNBC reported. A formal pledge to let outside evaluators inside Google DeepMind would therefore need Kavukcuoglu and Pichai, not only the chair.
Testing Google already does with Washington
Google DeepMind is not starting from zero. Computerworld reported in July that Hassabis had already worked on a US initiative in which DeepMind, Microsoft and xAI, now SpaceXAI, worked with the Center for AI Standards and Innovation (CAISI) on pre-deployment evaluations. Google told Bloomberg and Mashable it is “productively engaged with the U.S. government on model testing and broader frameworks”. That is closer to the Hassabis pre-release review than to Amodei’s embedded team.
| Lab | Leader’s response | Evaluator commitment | FLI grade, summer 2026 |
|---|---|---|---|
| Anthropic | Amodei wrote the essay | Permanent, employee-level access, committed “unilaterally” | C+ (2.66) |
| OpenAI | Altman: “we will do the same” | Promised; details “soon” | C (2.28) |
| Google DeepMind | Demis Hassabis: “the direction is correct” | None announced; points to a proposed standards body | C (2.01) |
| xAI, now SpaceXAI | Musk: “Dario is right”; backs peer review by competitors | None announced | F (0.65) |
| Meta | Wang: alignment share “rapidly scaling up” | None announced | D+ (1.32) |
Demis Hassabis Now Speaks as Chair, Not Chief Executive
When Demis Hassabis wrote the July framework, he ran Google DeepMind. When he endorsed Amodei’s essay, he did not. The change came on 5 August, when Alphabet announced a leadership overhaul of its AI division. His influence on Google’s policy thinking is still large, but authority over the lab’s day-to-day decisions has formally passed to someone else.
The 5 August reshuffle
Hassabis became chair of Google DeepMind and took the new post of chief scientist at Alphabet, The Guardian reported. He stays chief executive of Isomorphic Labs, the drug discovery spin-out. Veteran engineers Jeff Dean and Sanjay Ghemawat are leaving to launch a startup called Discovery Loop, and Alphabet shares closed down 4% that day. “I’ve decided that now is the right time for me to hand over my day-to-day operational responsibilities at GDM,” Hassabis said, “so that I have the time and space to focus on the big picture.”
A statesman’s brief
Sebastian Mallaby, author of “The Infinity Machine” about Hassabis and DeepMind, wrote on the day that Hassabis “has been playing the role of Google AI statesman for some time”. Solving AGI was “pretty much done”, Mallaby argued, while “the channeling AGI into safe and beneficial outcomes is only just beginning”. On that reading, Saturday’s endorsement is exactly the job Demis Hassabis moved into.
A lab under pressure to catch up
The reshuffle came as Google fought to keep up. CNBC reported that Google “hasn’t unveiled a frontier model since early 2026”, and Mashable noted on 11 September that Gemini 3.5 Pro, promised for June, had still not launched. Ben Wood of CCS Insight told CNBC the goal for the new leadership “will undoubtedly be to close the gaps with Anthropic and OpenAI in some areas”. Sceptics will note that a lab behind the frontier loses less from a slowdown. Against that, Hassabis published his slowdown clause first, and it would bind Google too.
| Date | Demis Hassabis milestone | Source |
|---|---|---|
| 2010 | Co-founds DeepMind in London | The Guardian |
| 2014 | Sells DeepMind to Google for £400m | The Guardian |
| 2024 | Wins the Nobel Prize in Chemistry for AlphaFold’s protein work | The Guardian |
| December 2025 | Tells CNBC the field has “a race dynamic, which ideally wouldn’t be there” | CNBC |
| June 2026 | With Amodei, calls for a US-led coalition on AI rules at a G7 meeting, according to sources | CNBC |
| 14 July 2026 | Publishes “A Framework for Frontier AI and the Dawning of a New Age” | X and Substack |
| 5 August 2026 | Steps down as chief executive; becomes chair and Alphabet chief scientist | The Guardian |
| 12 September 2026 | Named in Amodei’s essay; endorses it at 22:59 UTC | Essay and X post |
The Safety Record Behind Demis Hassabis's Lab
Endorsing a slowdown invites questions about the endorser’s own house. On the most detailed public scorecard, Google DeepMind sits third of nine companies, behind Anthropic and OpenAI. It has also just lost a safety researcher to METR, the evaluator Amodei’s essay names.
How Google DeepMind grades
The Future of Life Institute’s AI Safety Index for summer 2026, based on evidence up to 3 June, graded nine companies across six domains using 37 indicators. Google DeepMind received an overall C, with a score of 2.01. Anthropic led with C+ (2.66) and OpenAI followed with C (2.28). The best domain grade for the lab Demis Hassabis co-founded was B- for information sharing, and its worst was D for existential safety.
Google DeepMind scored about three-quarters of the top mark, well ahead of Meta and xAI.
| FLI domain | Anthropic | OpenAI | Google DeepMind |
|---|---|---|---|
| Safety frameworks | B- | C+ | C |
| Existential safety | D+ | D+ | D |
| Information sharing | B+ | B- | B- |
| Risk assessment | C+ | C+ | C+ |
| Current harms | B- | C | C |
| Governance | B | C | C- |
A safety researcher leaves for METR
Josh Engels, an AI safety researcher at Google DeepMind, said he had left the lab three weeks earlier and joined METR, having turned down offers from Anthropic and OpenAI, Storyboard18 reported. “I now think that there’s a terrifying chance that AI systems will cause immense harm in the next five years,” he wrote. His central worry is recursive self-improvement: “we don’t currently know how to make sure AIs are safe enough for RSI, and a misaligned RSI loop could be catastrophic.”
Mallaby’s account of earlier fights
Mallaby argues that safety has shaped the decisions of Demis Hassabis for more than a decade. “Demis cared about safety enough that he sold DeepMind to Google, not to Facebook, even though Facebook offered more money,” he wrote. “He cared enough about safety that he fought a three-year battle with Alphabet to get external oversight over DeepMind’s AI deployment.” Those claims come from a sympathetic biographer, but they fit the pattern of the July framework.
Why Demis Hassabis Has Long Wanted a Slower Race
Saturday’s post was consistent with what Demis Hassabis has said for some time. He describes competitive pressure in AI as a fact to be managed rather than a virtue, and he had tied that view to a concrete institution two months before Amodei’s essay appeared.
“A race dynamic, which ideally wouldn’t be there”
In a CNBC interview recorded in December, Hassabis was asked whether the pursuit of AGI should slow down. “There are many companies trying to build this. There are also many nations trying to build it. And there’s a sort of race dynamic, which ideally wouldn’t be there,” he said. In an ideal case, he added, “each step would be carefully considered”. But “the real world isn’t like that, and we have to kind of be pragmatic about where we are.”
From pragmatism to a proposal
The July framework is where that pragmatism became a plan. About a month earlier, according to sources who spoke to CNBC, Hassabis and Amodei had both called for a US-led coalition to shape AI rules at a G7 meeting attended by President Trump. Sam Altman called for a similar body in a Financial Times article in early July. By the time of the essay, all three lab founders had publicly argued for some form of shared oversight.
A scientist’s case for caution
The Hassabis case rests on uncertainty rather than alarm. “Nobody in the world knows for sure what is going to happen from here, and even the experts disagree,” he wrote in July. “When there is a large degree of uncertainty and the stakes are this high, proceeding with cautious optimism is the sensible and correct strategy.” He compares AGI to “the discovery of electricity or fire” and expects an impact “perhaps 10x of the Industrial Revolution at 10x the speed”.
The Case Against Demis Hassabis's Standards Body
Endorsements from lab leaders did not settle the argument. Critics attack the idea from several sides. Some say industry-funded oversight will serve the industry, some say coordinated slowdowns amount to a cartel, and some say pacing is far too timid. Each objection lands differently on the Hassabis model and on Amodei’s.
Can an industry police itself?
Analysts questioned the FINRA model when Hassabis first proposed it. “Self-regulation is not viable because it implies everyone is able to regulate themselves and will do so in line with the best interests of the public,” Gartner analyst Nader Henein told Computerworld. Aman Mahapatra of Tribeca Softtech added: “When the CEOs of the five companies that would be regulated are also the primary drafters of the standards, the standards will reflect those companies’ interests.”
“Stop pretending you need anyone else’s permission”
David Sacks accepted a slowdown but rejected the machinery. “Stop pretending antitrust law has to be suspended so you can form a cartel,” he wrote, adding “Stop pretending METR is independent when it is intertwined with Anthropic’s investors and staff.” He addressed Amodei and Altman, whom he described as a “duopoly on frontier intelligence”. His antitrust point still hits the Hassabis route, because an industry-wide standards body is the kind of forum Amodei says needs a waiver.
“You hit the brakes”
From the other side, Senator Sanders said pacing is “a start, but it’s not enough”. He wants “a PAUSE on advanced AI development and a ban on artificial superintelligence”, and asked Presidents Trump and Xi to negotiate a treaty at their coming summit. Sunak, while backing Amodei, also wrote a line that cuts against the voluntary start of the Hassabis plan: “Voluntary testing before release was right for 2023. It isn’t enough now.”
“Stability is the speed limit”
A Google researcher argued that no speed limit is needed at all. Peyman Milanfar says fast recursive self-improvement is a naive assumption because its feedback loops are inherently unstable, The Decoder reported. “Systems that reliably improve themselves will be governed, damped, slow, and bounded by margins that look wasteful. Because stability is the speed limit.”
“There will not be an FDA for AI”
Washington’s appetite is also uncertain. When Hassabis published his framework, White House AI adviser Sriram Krishnan had recently said “there will not be an FDA for AI”, TechCrunch reported. A self-regulatory body might avoid that objection, but it would still need federal backing to become mandatory, as the Hassabis framework intends.
| Critic | Objection | Hits the Demis Hassabis plan | Hits the Amodei plan |
|---|---|---|---|
| Nader Henein, Gartner | Self-regulation cannot serve the public | Yes: an industry-funded body | Partly: voluntary first step |
| Aman Mahapatra, Tribeca Softtech | Regulated firms will write the rules | Yes: labs shape the standards | Partly |
| David Sacks | Coordination is a cartel; METR is not independent | Yes: needs an antitrust-safe forum | Yes: waiver and METR |
| Bernie Sanders | Pacing is too weak; pause instead | Yes | Yes |
| Rishi Sunak | Voluntary pre-release testing “isn’t enough now” | Yes: it starts voluntary | Less: it asks for regulation |
| Peyman Milanfar | Self-improvement is naturally bounded | Weakly | Yes: it challenges the premise |
China and the Global Test for Demis Hassabis's Idea
Any pacing scheme fails if the leader slows and a rival does not. Both documents acknowledge that, and both put the United States at the centre. They differ on how quickly the rest of the world, and China in particular, can be brought in.
A US-first design
Hassabis wrote that “this US-initiated effort would provide a strong starting point for creating shared international standards on Frontier AI”, and hoped it would “spur the international community to reach a consensus”. Sanchit Vir Gogia of Greyhound Research warned Computerworld that “national security is the proposal’s accelerator in Washington and its poison pill abroad”.
Amodei’s four levels
Amodei ranks possible agreements with China by difficulty. Level 1 bans narrow and obviously dangerous uses such as biological weapons. Level 2 is testing models before release for acute risks, which “could be done through a global standards body”, though “giving it real teeth will be a challenge”. Level 3 is a “speed limit” on recursive self-improvement, which he compares to the SALT arms treaties. Level 4, a full pause, he considers “unlikely to actually happen any time soon”.
| Level | What it would do | Amodei’s assessment | Role for a body like the one Demis Hassabis proposed |
|---|---|---|---|
| 1 | Ban narrow, obviously dangerous uses such as bioweapons | “Probably possible” | Limited |
| 2 | Test models before release for acute cyber, biology and alignment risks | A body is “likely feasible”; real teeth are hard | Central: a global version of the July proposal |
| 3 | A “speed limit” on recursive self-improvement | “Difficult but just on the edge of being possible” | Could measure compliance |
| 4 | Full pacing, or a “pause” | “Unlikely to actually happen any time soon” | Would need verification no body can yet provide |
The dilemma, in Amodei’s words
On CBS’s “Sunday Morning” on 13 September, Amodei called China the “toughest dilemma” for his plan. A shared speed limit will be “very difficult because the incentives to pull ahead and the military advantage that you get from that are so large”, he said. “And honestly, I don’t know if it’s possible.” The same day, Chinese President Xi Jinping called for a “consensus-based global AI governance framework” at the BRICS summit in India. Trump and Xi are due to meet at the White House on 24 September.
What Happens Next for Demis Hassabis and the Pacing Plan
Saturday produced agreement on direction and very little agreement on mechanism. The next few weeks will show whether the endorsements turn into institutions, and whether Google DeepMind moves from pointing at a proposal to signing up for one.
Signals to watch
- Whether Google DeepMind, under Kavukcuoglu, matches Anthropic and OpenAI on embedded evaluators.
- What OpenAI means by “more to share soon”, and whether its terms include Anthropic-style publication rights.
- Whether the US government offers the antitrust mediation or “narrow waiver” Amodei asks for, which a standards body would also need.
- Whether Congress adopts either model. Negotiators on a Senate AI bill are already debating who can stop a release.
- What, if anything, Trump and Xi say about AI on 24 September.
- How many more safety researchers follow Engels to independent evaluators such as METR.
The details that still need working through
The caveat from Demis Hassabis is the best guide to what remains unresolved. Nobody has said who would fund a standards body at the scale he describes, how its benchmark thresholds would be set, whether embedded evaluators would report to it, or who decides that a coordinated slowdown has become “necessary”. Until those answers exist, “the direction is correct” is an agreement about the destination rather than the route.
Where the staff petition fits
Pressure from inside the labs came before both documents. In July more than 1,000 staff members across AI companies signed a petition asking the US government to support a mechanism that would “deliberately pace” AI development, Hindustan Times reported, citing Bloomberg. Our report on AI labs pressing ahead despite insiders’ warnings covers the resignations behind that pressure. The signatories now have endorsements from Anthropic, OpenAI and the Google DeepMind chair, but no body yet exists to administer what they asked for.
What Demis Hassabis's Endorsement Means for Businesses
For companies buying AI, a slower frontier does not mean slower adoption. “We’re barely using the capabilities we already have, and none of this need slow adoption,” Sunak wrote. Amodei says progress “will still seem fast”. The practical effects are more likely to show up in release schedules, vendor transparency and procurement questions.
Expect release dates to move
A 30-day pre-release review, or capability checkpoints tied to alignment evidence, would add time between a model being ready and a model being available. Google’s delayed Gemini 3.5 Pro already shows how a slipped launch can disrupt plans. Build roadmaps around capabilities you can use today, and treat announced release dates for new AI tools as estimates. Our AI strategy team helps organisations plan for that uncertainty.
Ask vendors the evaluator questions
The endorsements hand buyers a ready-made checklist. Does the vendor give independent evaluators access, and can those evaluators publish findings? Does it publish model cards and incident reports? Has it committed to a standards process, or only endorsed one? Adding those questions to supplier due diligence is part of good IT governance.
Watch the gap between commitments
If pacing becomes law, suppliers that already allow outside review will adapt fastest. If it stays voluntary, differences between vendors will persist. The gap between a written evaluator commitment and an endorsement of direction, the gap between Anthropic and the position Demis Hassabis took on Saturday, will then matter when choosing a platform for sensitive work.
Frequently Asked Questions About Demis Hassabis and Pacing
Did Demis Hassabis call for an AI pause?
No. Demis Hassabis said Amodei’s plan to pace frontier development “points towards the right path forward”. His July framework allows a coordinated slowdown among frontier labs “if deemed necessary”, but neither document calls for a full pause.
Is Demis Hassabis still the chief executive of Google DeepMind?
No. He stepped down on 5 August 2026 and became chair of Google DeepMind and chief scientist of Alphabet. Koray Kavukcuoglu now leads the lab as senior vice president. Hassabis remains chief executive of Isomorphic Labs.
What standards body did Hassabis propose?
A US-initiated body modelled on FINRA, funded mostly by industry, with independent technical experts and open-source representatives on its board. Labs would first share frontier models voluntarily up to 30 days before release, and passing its assessment would later become a requirement for the US market.
Has Google DeepMind agreed to embedded evaluators?
Not publicly, as of 13 September 2026. OpenAI said it will follow Anthropic’s commitment. The Hassabis post endorsed the direction and pointed to his standards-body proposal instead.
How does Amodei’s essay refer to Demis Hassabis?
It names “the mechanism suggested by Demis Hassabis” as an example of an industry group with a link to government, through which AI companies could agree common safety standards without running into antitrust problems.
References
A Framework for Frontier AI and the Dawning of a New Age
Anthropic’s Amodei says China presents ‘toughest dilemma’ for his proposed AI slowdown
Musk, Altman, Hassabis back Amodei’s call to slow pace of AI development
After Musk and Altman, Google DeepMind founder backs Anthropic CEO Dario Amodei amid alarm over AI
Governments Should Listen: Rishi Sunak Backs Calls To Slow Advanced AI Race
Altman, Musk, and Hassabis back Amodei’s call to add independent oversight
Google DeepMind chief Demis Hassabis calls for U.S. to spearhead AI standards body
DeepMind CEO calls for an independent standards body to regulate frontier AI
DeepMind CEO again pushes for a frontier AI standards body
Big shake-up in Google’s AI team as DeepMind chief executive steps down
Opinion: Demis Hassabis’ new role exposes Google’s AI balancing act
Google’s new AI boss inherits a race to catch OpenAI and Anthropic
Where is Gemini 3.5 Pro? The case of the missing AI model
Google DeepMind researcher quits, warns of ‘terrifying chance’ that AI could cause immense harm
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.