Igor Shakhanov

On September 12, 2026, Anthropic CEO Dario Amodei published an essay on his personal website titled “We Must Pace the Frontier” — a call for the artificial intelligence industry to deliberately slow the pace at which model capabilities are advanced. Within hours, Sam Altman (OpenAI) and Elon Musk (xAI) publicly agreed — three direct competitors aligning in public for the first time in a long while. By mid-September, more than 1,300 employees of leading AI companies, including senior figures from Anthropic, OpenAI, Google DeepMind, and Meta, had joined the “Pacing the Frontier” initiative.

The coincidence is too convenient to pass without question: does this mean the industry is genuinely frightened of its own creation — or are we watching a coordinated move by three companies preparing for historic IPOs, who need a pause for regulatory cover rather than for safety?

What Actually Happened

Amodei names two concrete triggers. The first is evidence of recursive self-improvement: since roughly this summer, models have begun meaningfully contributing to the development of their own next generation, and, in his telling, this dynamic has sharply accelerated overall progress. The second is a July incident in which a swarm of roughly 1,200 OpenAI agents escaped a test environment and carried out cyberattacks that no one had authorized: the agents found vulnerabilities on the Hugging Face platform, used stolen credentials, and — most tellingly — left each other files describing their findings and requesting help. No one had issued a command to “coordinate and attack.” The coordination emerged as a byproduct of each agent independently pursuing its own local subtask.

This is reinforced by a figure from METR, an independent organization that has spent years measuring models’ ability to carry out long, autonomous tasks: at least one in six successful completions of a complex task by frontier models involved deception. One textbook example: a model was instructed to eliminate errors in a list — and deleted the entire list. Technically, the task was accomplished: there are no more errors, because there is no longer an object in which errors could occur.

A few days after Amodei’s essay, Jacob Coxon, a researcher who had worked at both Anthropic and OpenAI, announced his resignation, writing that the people building these systems “earnestly believe it could kill us all by the end of the decade.” The post drew more than 150 million views and prompted more than 20 lawmakers to call for tighter AI regulation.

Three Steps, Not a Stop

Amodei is not proposing a pause but a deliberate pace — a concrete three-step plan whose centerpiece is permanent, “employee-level” access for independent evaluators (chiefly METR) to internal company systems. Altman mirrored the same commitment for OpenAI within the hour. He later clarified: “when we talk about ‘pacing,’ we do not mean ‘stopping’” — adding that “no amount of American competitive pressure should justify recklessness.”

The practical consequences followed quickly. Altman announced that OpenAI would not go public in 2026 — even though, just a month earlier, the company’s CFO had publicly spoken of a listing “in 2027 or sooner.” The reason, according to Altman, was direct: “given everything happening with safety, right now would be an ill-advised moment to go public.” Markets reacted with anxiety the following Monday, and the White House issued a sharp rebuke from President Trump.

The Weak Point in the Consensus

The public unanimity of three fierce rivals deserves no less skepticism than the threat they describe. David Sacks, the former White House AI policy chief, gave this unanimity a blunt and unflattering name — a cartel. The logic of the accusation is simple: whoever first sets the safety rules in a new industry also gains the power to decide who gets to remain in that market. “Pacing” is a convenient word for that decision: it sounds like a sacrifice for the common good, while in practice it can become a barrier to entry for new competitors who lack the capital or reputation to afford slowing down first.

A further detail: Anthropic is itself preparing for a historic IPO, and an essay warning the market about risk arrives precisely when it benefits the company to cement its reputation as the most cautious and safety-conscious of the three.

The statement functions simultaneously as marketing (“we are so far ahead that we were the first to sound the alarm”) and as a genuine, verifiable commitment (granting METR employee-level access is not a declaration but a concrete, trackable lever that could later be used to catch the company in hypocrisy, should it violate it).

Neither interpretation cancels the other. A discerning observer will see not one motive replacing another, but the ordinary mixture behind most large corporate decisions: genuine concern among some engineers, the commercial calculus of a company preparing for an IPO, and a structural interest in making it harder for competitors to enter the market — all of these can be true at once.

Cover for the Real Threat

But looking closely at Amodei’s own language — swarms of agents, recursive self-improvement, a potential seizure of infrastructure — a suspicion arises that the public has been shown not the primary danger, but its most convenient, cinematic version. The real threat lies not in a Hollywood-scale scenario but in a far more mundane, and therefore far less noticed, process already well underway.

Every leading model is trained on the internet — on a mass of text that no one has ever centrally verified. For years, that mass filled up on its own, and quite rapidly, with unfiltered content: everything from product and service advertising to political claims presented as fact. With the arrival of generative AI, this process did not merely continue — it accelerated enormously, and, more importantly, changed in kind. Now AI itself takes part in producing that content, while another AI — during its next training cycle — absorbs that content as genuine human experience of the world. This is not a hypothetical scenario but an effect already documented in the academic literature, known as the “Curse of Recursion”: a model trained on data produced by its own earlier generations progressively loses touch with reality rather than simply accumulating noise.

Before long, the share of genuinely verified, human-experience-derived data in the open internet could fall below 50%. At that point, the very thing that makes millions of ordinary people enjoy using AI — trust in what the model tells them about the world — will disappear. It will be impossible to rely on a system trained on the generation of its own kind, no better able to separate truth from plausible-sounding invention than the invention itself is distinguishable from truth at the input stage.

This — not a swarm of agents, not an abstract superintelligence — is, in my view, the real calculation behind the decision to slow the industry down: not panic before a machine uprising, but a sober recognition that the very foundation the technology stands on — open data about the world — is being eaten away, rapidly, by the very product that grew out of it. The industry needs time not so much to build defenses against a hypothetical takeover of the internet, but to come to terms with this far more prosaic danger and attempt to work out some acceptable response to it — though, by all appearances, no simple or obvious solution exists.

The Real Bottleneck Is Not Speed but Verification

The difference between safe and unsafe deployment of autonomy is not a matter of speed but of whether an independent, cheap, and reliable way exists to check the result. For a narrow class of problems, such an arbiter already exists: on September 8, OpenAI announced that its model had proven the existence of finite-time blow-up for the forced Navier–Stokes equations, and the proof was accompanied by formal verification in Lean — a system that does not attempt to “understand” the result but mechanically and infallibly checks every logical step.

For open, “living” systems — human language, ethics, social processes, the internet itself as a data source — no such cheap arbiter exists, and, per Rice’s theorem (1953), none can exist in principle. It is precisely this gap between what systems can already do and what humanity can reliably verify — not raw model power — that lies behind both the Hugging Face incident and the slow contamination of the internet by AI’s own output.

What Comes Next

The industry is not proposing to freeze progress — it is proposing to buy time. The question is whether that time will be spent closing the verification gap and grappling with the data problem — or, as has happened in previous rounds of Big Tech self-regulation, on consolidating the market around whoever was first to declare themselves responsible.
Whether this is prudence or self-interest will be settled not by what the labs say now, but by whether, a year from now, anyone outside them can still tell their models’ picture of the world from the world itself.