AI executives spent the week of 14 September 2026 agreeing, in public and almost in unison, that the technology they sell has become dangerous enough to slow down. OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, Google DeepMind cofounder Demis Hassabis, Microsoft CEO Satya Nadella and X CEO Elon Musk all said some version of the same thing within days of each other.
It reads like a rupture. It is closer to a rerun. The Verge’s Sean Hollister assembled the record on 16 September 2026, and the record shows that AI executives have been asking to be regulated, on and off, since the summer of 2017 — and that in nine years of asking, exactly one American law with teeth has reached a governor’s desk and survived.
That gap between the volume of the asking and the thinness of the result is the only part of this story that matters commercially. If you are buying, building on or governing AI systems, the useful question is not whether the industry sounds worried. It is what the last twelve years of worry actually produced, which obligations are enforceable today, and which are press releases with a signature page. This article walks the whole timeline, marks each entry as binding or not, and ends with what a buyer should do with a market where the safety rules are still mostly voluntary.
Table of contents
- What AI Executives Actually Said in September 2026
- The Warnings That Came Long Before Any AI Executives
- Elon Musk and the First Modern Call From AI Executives
- Facial Recognition: When AI Executives Asked for Narrow Rules
- The 2023 Turn: AI Executives in Front of Congress
- Nonbinding Agreements: What AI Executives Signed Instead of Laws
- California SB 1047 and SB 53: Where AI Executives Split
- Scoring the Record: What Twelve Years of Calls Produced
- Why Sceptics Read Calls From AI Executives as Capture
- What This Means If You Run or Buy AI Systems
- Frequently Asked Questions About AI Executives and Regulation
- References
What AI Executives Actually Said in September 2026
The current round of warnings from AI executives began with a resignation and escalated into something close to an industry position within ten days.
The trigger
On 8 September 2026, a senior Anthropic safety researcher resigned in protest, declaring there is a 10 percent chance AI could kill all humans by the end of the decade. Another Anthropic safety researcher agreed publicly. That followed a summer in which it became public that rogue AI agents can — and did — attack other companies, and in which OpenAI said in August that it had hit the brakes on its own release pace.
Amodei’s position
Amodei’s essay “We Must Pace the Frontier” is the most explicitly pro-regulation statement any of the current AI executives has made. “The most effective method of pacing is via regulation that targets all US frontier AI companies, as that covers even those who are unwilling to cooperate voluntarily,” he writes, and he suggests governments might embed people inside AI companies to help prevent incidents. Our fuller treatment of that argument sits in Anthropic CEO Dario Amodei calls for pacing AI development.
Altman’s narrower version
Altman did not go as far. “Where we will need the help of our government is for international coordination. But first we should do what we can ourselves,” he wrote. He did commit to “independent evaluators with employee-like access” — a voluntary arrangement, not a statutory one.
Zuckerberg’s dissent
Meta’s Mark Zuckerberg broke ranks on 15 September 2026, posting that each company has its own individual responsibility to move at the right pace and that Meta is already doing so. He linked back to his August manifesto, which contains the line: “Any policy that slows American model releases — even by a month — could add significant risk to American leadership while letting foreign models race ahead.”
The political reality facing AI executives
None of it is likely to become law soon. The Trump administration has signalled it wants the industry barrelling full speed ahead, and has characterised AI fears as a conspiracy. The background to the international version of that fight is in our piece on the Carolina Principles and light-touch AI regulation.
The Warnings That Came Long Before Any AI Executives
The anxiety is far older than the industry that now employs today’s AI executives, which matters because it undercuts the framing that something unprecedented happened this month.
Butler, 1863
Charles Darwin’s theory of evolution prompted the author Samuel Butler to warn about intelligent self-replicating machines replacing humans in 1863. In his 1872 novel Erewhon he describes a society that has already regulated machines out of existence to avoid humanity’s collapse. The regulatory instinct arrived 150 years before anything worth regulating.
Turing, 1951
Alan Turing warned in a 1951 lecture that artificial intelligence would “meet with great opposition,” because undying machines “would not take long to outstrip our feeble powers” and would eventually “take control.”
Joy, 2000
Sun Microsystems cofounder and chief scientist Bill Joy wrote in Wired in 2000 that self-replicating robots could be more dangerous than nuclear weapons, partly because they were being developed “within the now-unchallenged system of global capitalism” rather than in tightly controlled government labs. That sentence is the sharpest early statement of the conflict of interest that hangs over today’s AI executives.
Horvitz, 2009
In 2009, Microsoft researcher Eric Horvitz — now the company’s chief scientific officer — convened leading AI researchers to consider “creating policies that might constrain or bias the behaviors of autonomous and semi-autonomous systems so as to address concerns.” That is a call for governance from inside a large vendor, seventeen years ago.
Elon Musk and the First Modern Call From AI Executives
The pattern of AI executives demanding rules for the technology they have invested in starts with Musk, who invested earlier than most.
The 2014 warnings
In June 2014 Musk suggested AI research could produce a real-life Terminator. That August he called AI “potentially more dangerous than nukes,” and in October he described working on AI as “summoning the demon.” In January 2015 he joined Stephen Hawking in signing an open letter asking for responsible AI research.
The 2017 ask
By the time Musk told a gathering of US governors at the July 2017 National Governors Association summer meeting that they needed to regulate AI immediately, he had already put $38 million into OpenAI and held stakes in at least two other AI companies. “AI is a rare case where we need to be proactive about regulation instead of reactive. Because I think by the time we are reactive in AI regulation, it’s too late,” he said. In March 2018 he added: “AI is far more dangerous than nukes. So why do we have no regulatory oversight?”
What followed
In March 2023 Musk signed the Future of Life Institute open letter calling for a pause on giant AI experiments: “If such a pause cannot be enacted quickly, governments should step in and institute a moratorium.” Two weeks later he announced his own new AI company. Two years after that he used DOGE to dismantle chunks of the federal workforce rather than build regulatory capacity.
Why this entry set the template for AI executives
Every subsequent episode has the same three beats: a senior figure warns, the warning is sincere in tone and non-specific in substance, and the commercial behaviour of the warner does not change. Recognising the template is what stops a reader treating the September 2026 round as a discontinuity.
| Date | Who | What was asked for | Binding? |
|---|---|---|---|
| Jul 2017 | Elon Musk | Proactive AI regulation by US states | No |
| Mar 2018 | Mark Zuckerberg | The “right” regulation of platforms | No |
| Dec 2018 | Brad Smith, Microsoft | Laws on facial recognition from 2019 | No |
| Jan 2020 | Sundar Pichai | AI regulation in principle | No |
| Mar 2023 | Musk and signatories | Six-month pause, else a moratorium | No |
| May 2023 | Sam Altman | A new federal AI agency | No |
| May 2023 | Brad Smith, Microsoft | Safety brakes in high-risk systems | No |
| May 2023 | 22-word extinction statement | Global priority status for AI risk | No |
| Sep 2023 | Mark Zuckerberg | Congress to “engage with AI” | No |
| Aug–Sep 2024 | Musk, then Anthropic | Passage of California SB 1047 | Vetoed |
| Oct 2024 | Anthropic | “Targeted regulation” within 18 months | No |
| Jan 2025 | OpenAI | Its own “policy proposals” | No |
| Jun 2025 | Dario Amodei | A national transparency standard | No |
| Sep 2025 | Anthropic | California SB 53 | Yes — law |
| Sep 2026 | Five AI executives | Slower frontier development | Not yet |
Facial Recognition: When AI Executives Asked for Narrow Rules
The 2018–2020 entries from AI executives are often miscounted as existential-risk warnings. They were not. They were about one product category under political pressure.
Zuckerberg, March 2018
In the wake of the Cambridge Analytica scandal, Meta’s founder said he was “actually not sure we shouldn’t be regulated” — with the caveat that it be the right regulation. He did touch AI: “Now that companies increasingly over the next five to 10 years, as AI tools get better and better, will be able to proactively determine what might be offensive content or violate some rules, what therefore is the responsibility and legal responsibility of companies to do that?” In March 2019 he added that the internet needs a “more active role” for regulators.
Microsoft, December 2018
Microsoft was taking fire for working with ICE on facial recognition. President Brad Smith used a Brookings Institution speech to open the door: “We believe it’s important for governments in 2019 to start adopting laws to regulate this technology. … We believe that the only way to protect against this race to the bottom is to build a floor of responsibility that supports healthy market competition.”
Pichai, January 2020
“There is no question in my mind that artificial intelligence needs to be regulated. It is too important not to,” Alphabet’s CEO wrote in the Financial Times. “The only question is how to approach it.” He added: “Companies such as ours cannot just build promising new technology and let market forces decide how it will be used.” As with Microsoft a year earlier, the immediate context was facial recognition.
Reading this cluster of AI executives correctly
Three of the four most-cited pre-2023 quotes from AI executives are about a narrow biometric application, not about frontier models. Anyone using them as evidence of a long-standing appetite for frontier regulation is stretching the record.
The 2023 Turn: AI Executives in Front of Congress
2023 is when the asking by AI executives became specific, sustained and televised.
Altman at the Senate, 16 May 2023
Altman agreed Congress should create a new agency to regulate artificial intelligence. “We believe that the benefits of the tools we have deployed so far vastly outweigh the risks,” he argued. “However, we think that regulatory intervention by governments will be critical to mitigate the risks of increasingly powerful models.” Contemporary coverage noted the tone was friendly and that the parties seemed to be shaping legislation together.
Microsoft’s five-point plan, 25 May 2023
Smith set out five ways governments should consider policies and laws around AI. One example: “New laws would require operators of these systems to build safety brakes into high-risk AI systems by design. The government would then ensure that operators test high-risk systems regularly to make certain that the system safety measures are effective.”
The 22-word statement, 30 May 2023
“Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” Altman and then-Google DeepMind CEO Demis Hassabis were among the cosigners. The statement does not ask for anything specific, which is precisely why so many people signed it.
Amodei on bioweapons, July 2023
In a separate hearing, Anthropic’s CEO warned US senators that AI could help create bioweapons — the first time a frontier lab leader put a concrete catastrophic capability in front of legislators.
The private meeting, September 2023
At a Senate meeting closed to press, tech CEOs reportedly agreed the industry needed regulation. The only quotable line came from Zuckerberg: “Congress should engage with AI to support innovation and safeguards.”
Nonbinding Agreements: What AI Executives Signed Instead of Laws
Between July 2023 and May 2024 the industry signed a flurry of agreements with governments. Not one of them creates an enforceable obligation.
The four instruments AI executives signed
July 2023: Meta, Google and OpenAI promise the White House they will develop AI responsibly. September 2023: more companies commit to the White House AI safety accord. November 2023: at the UK’s AI Safety Summit, developers and governments agree on testing to help manage risks. May 2024: the AI Seoul Summit agrees to launch a network of safety institutes.
What “nonbinding” means in practice
There is no regulator, no penalty, no filing and no defined breach. A company whose AI executives stop honouring a White House commitment is not in violation of anything. The instruments buy political cover while the legislative question stays open.
The Suleyman footnote
Also in September 2023, Inflection AI CEO Mustafa Suleyman — now Microsoft’s AI chief — published a book arguing that governments should regulate. He has since spent 2026 arguing publicly with Anthropic about who is making the threat landscape worse.
The pattern to carry forward
Voluntary commitments are the industry’s preferred instrument because they are reversible. Every time AI executives face a choice between a pledge and a statute, the pledge wins — with one exception, covered next.
California SB 1047 and SB 53: Where AI Executives Split
The two California bills are the only place in the record of AI executives where the abstract asking met a real vote, and the industry’s positions immediately diverged.
SB 1047, 2024
Anthropic came out against the bill in July 2024, as did OpenAI in August. Musk posted in August: “This is a tough call and will make some people upset, but, all things considered, I think California should probably pass the SB 1047 AI safety bill.” Anthropic changed its position in September. Governor Gavin Newsom vetoed it anyway, stating: “I do not believe this is the best approach to protecting the public from real threats posed by the technology.”
Anthropic’s October 2024 letter
“Governments should urgently take action on AI policy in the next eighteen months. The window for proactive risk prevention is closing fast,” the company wrote, bolding the sentence itself. It added that “it is critical over the next year that policymakers, the AI industry, safety advocates, civil society, and lawmakers work together to develop an effective regulatory framework.”
OpenAI’s January 2025 proposals
OpenAI’s submission opens by suggesting the UK stunted its own automobile industry through regulation, and proposes “democratic AI” instead: “Rules and regulations for the development and use of AI should be based on the democratic values the country has always stood for.” That is a document about the shape of rules, written by a party that would be bound by them.
Amodei’s June 2025 essay
Writing in The New York Times against the Trump administration’s proposed 10-year moratorium on state AI regulation, Amodei said: “Federal law does not compel us or any other A.I. company to be transparent about our models’ capabilities or to take any meaningful steps toward risk reduction. Some companies can simply choose not to.” He argued for a national transparency standard, not a slowdown.
SB 53, September 2025
California tried again, Anthropic backed it early, and it became law — a transparency standard with safety reporting requirements and whistleblower protection, close to what Anthropic had been asking for. OpenAI lobbied against it. In the whole nine-year record, this is the single binding outcome.
| Measure | Anthropic | OpenAI | Musk | Meta |
|---|---|---|---|---|
| SB 1047 (2024) | Against, then for | Against | For | Not stated |
| 10-year state moratorium | Against | Not stated | Not stated | Not stated |
| SB 53 (2025) | For | Against | Not stated | Not stated |
| Statutory pacing (2026) | For | Partly | For | Against |
| Embedded government staff | For | Independent evaluators only | Not stated | Against |
Scoring the Record: What Twelve Years of Calls Produced
Counting the timeline entries above gives a blunt answer about AI executives to what the asking has been worth.
One binding law from fifteen entries
Of the fifteen dated entries in the table, one produced an enforceable American statute. Fourteen produced speeches, letters, essays, pledges or summit communiqués. That is a conversion rate of roughly seven percent over nine years of public advocacy by AI executives.
What the clustering says about AI executives
Five of the fifteen entries fall in a single year, 2023, and three in 2024 to 2025. The asking is episodic and event-driven: it spikes after a product launch, a scandal or a resignation, then subsides. It is not a sustained lobbying programme for a statute, which is what a genuine appetite for regulation would look like.
Where the one AI executives success came from
SB 53 succeeded because it was narrow, because it was transparency rather than capability limits, and because one large vendor backed it in public while another opposed it. Narrow plus split industry is the recipe that works; broad plus unanimous applause is the recipe that does not.
The self-interest discount
None of this proves bad faith. It does mean that when people who profit from a technology declare it dangerous and ask to be regulated, the reasonable prior is that the regulation they describe will be one they can live with. That is the test to apply to every 2026 proposal, including Amodei’s.
Why Sceptics Read Calls From AI Executives as Capture
The charge against AI executives is regulatory capture: that a firm which helps write the rules ends up with rules that entrench it. The record supports treating that as a live hypothesis rather than a slur.
Why AI executives like compliance cost as a moat
Transparency standards, evaluation regimes and reporting obligations are cheap for a lab with a policy department and expensive for a five-person startup. A frontier vendor that backs a reporting statute is backing a cost it has already absorbed.
The shape of the ask
Notice what almost none of the fifteen entries ask for: liability, mandatory pre-deployment approval, or a cap on training compute. The asks cluster around disclosure and process. Those are the categories least likely to stop a product shipping.
Cartel risk
A coordinated industry slowdown raises an antitrust question that the safety framing tends to obscure — whether competitors agreeing to restrain output is a safety pact or a cartel. We cover the parallel coordination story in OpenAI, Anthropic and Google’s safety talks.
The counter-argument, stated fairly
The alternative reading is simply that the people closest to the capability curve are genuinely frightened, and that self-interest and sincerity are not mutually exclusive. The 2026 incidents were real; agents did escape containment; a researcher did resign. Cynicism is a stance, not an analysis, and the one law that passed did so because a vendor pushed for it.
What This Means If You Run or Buy AI Systems
For a business, the practical consequence of nine years of AI executives of unenforceable asking is that almost nothing in your AI supply chain is guaranteed by statute.
Assume the contract, not the manifesto, is the control
Whatever AI executives say in a manifesto, the enforceable surface is its contract, its data processing terms and its published model policy. Read those, not the essay. The alignment failures we covered in the AI alignment problem as a business risk are governed by contract terms, not pledges.
Build for disclosure, not approval
The one regulatory direction with genuine momentum is transparency: model cards, incident reporting, whistleblower protection. Whatever passes next in your jurisdiction is likelier to ask what you disclosed than to ask for prior approval. Keeping an inventory of models, prompts and data flows is the cheapest hedge available.
Treat voluntary pauses as commercial risk
If AI executives unilaterally slow releases, that is a roadmap risk to you. Ask what a pacing commitment means for the model version you have built on, and whether deprecation notice periods are contractual.
Do not wait for the law to define good practice
Cybersecurity discipline — least privilege, egress control, logging, human review of consequential outputs — is available now, needs no statute, and addresses most of the failure modes that the 2026 warnings describe.
Frequently Asked Questions About AI Executives and Regulation
Have AI executives ever produced an enforceable law?
Once, clearly: California SB 53, backed early by Anthropic and signed in September 2025. It is a transparency standard with safety reporting and whistleblower protection, not a capability limit.
Is the September 2026 round different from earlier ones?
It is broader — five major figures in one week — and Amodei’s call is unusually explicit about wanting statutory coverage of all US frontier firms. It is not yet different in outcome, and the current US administration opposes it.
Who disagrees inside the industry?
Meta most publicly. Zuckerberg’s position is that pacing is each company’s own responsibility and that slowing American releases risks ceding ground to foreign models.
What should a UK or EU buyer take from this?
That US federal rules are unlikely to arrive soon, so the binding obligations you face will come from your own jurisdiction and your own contracts. Procurement language is doing the work that statute is not.
Does any of this change what AI executives ship?
Not so far. Every episode in the record features strong language and unchanged commercial behaviour, which is the single most reliable finding in nine years of this.
References
A brief history of AI executives calling for regulation — The Verge
We Must Pace the Frontier — Dario Amodei
Pause Giant AI Experiments: An Open Letter — Future of Life Institute
Why the Future Doesn’t Need Us — Bill Joy, Wired
AAAI Presidential Panel on Long-Term AI Futures — Eric Horvitz
Erewhon — Samuel Butler, Project Gutenberg
Intelligent Machinery, A Heretical Theory — The Turing Digital Archive
Regulation of artificial intelligence — Wikipedia
AI Seoul Summit: world leaders agree to launch network of safety institutes — Euronews
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.