What Anthropic published
On June 4, 2026, Anthropic published "When AI Builds Itself" — proposing a globally coordinated pause or slowdown on frontier artificial intelligence development. The report argues that AI systems are accelerating their own development at a pace that may outstrip existing safety and governance frameworks.
The report was published by the Anthropic Institute and authored by company co-founder Jack Clark and Marina Favaro, head of its research institute. The core argument: AI systems are increasingly contributing to the development of future AI models, accelerating progress in ways that could eventually culminate in what researchers call "recursive self-improvement" — a scenario in which an AI system becomes capable of designing and developing its own successor with limited human involvement.
Anthropic's central recommendation is that the world should have the "option" to slow or temporarily pause frontier AI development "to enable societal structures and alignment research to keep up." The company called a worldwide slowdown "likely a good thing" but stressed that if only one company stopped, competitors would race ahead. Critically, Anthropic did not commit to a unilateral halt.
The paper landed in the same week that Fable 5 — Anthropic's most powerful model ever released to the public — went live on every device on earth. The same company. The same week. One hand releasing its most powerful product to the world. The other raising a warning flare about where the next product might lead.
The data that made them write the paper
This is not a philosophical essay from a think tank. It is a warning backed by Anthropic's own internal production metrics — and those numbers are more striking than the proposal itself.
As of May 2026, more than 80% of code merged into Anthropic's own codebase was written by Claude. Not by human engineers. By the AI itself. Engineers at the company were merging 8x more code per day in Q2 2026 compared to 2024.
AI task-completion horizons — essentially how complex a task an AI can handle autonomously — have been doubling roughly every four months. In March 2024, models could handle tasks that took about 4 minutes. Apply four doublings per year to that baseline. By early 2026, models are handling tasks that take hours. By late 2026, if the trajectory holds, tasks that take days. By 2027, tasks that take weeks. The tasks that take weeks are the ones currently performed by research teams, engineering departments, and entire divisions of large organisations.
In internal tests, Anthropic's Mythos Preview model achieved approximately 52x performance improvements, far exceeding what skilled human engineers could accomplish. That number — 52x — is not a benchmark score. It is a measure of how much faster an AI can improve an AI system than a human engineer can. If that ratio is even approximately correct, the implication is direct: the entity most capable of improving AI is now AI itself. The human in the loop is becoming a bottleneck, not a safeguard.
Anthropic's typical engineer now merges roughly eight times as much code per day as in 2024, largely because AI systems now generate substantial portions of the software used throughout the company. The engineers are not redundant. But their role has shifted from writing code to reviewing it, directing it, and taking responsibility for what the AI has produced. That shift — from creation to governance — is the model for every knowledge profession in the coming decade. And it raises the question that the paper is really about: what happens when the AI being governed is the one doing the governing?
What a coordinated pause would actually mean
This isn't a call to unplug every GPU and go home. Anthropic is advocating for a coordinated, multilateral framework where major AI laboratories agree to temporarily halt the most advanced development work while safety research and verification systems catch up.
The proposed coordination would let advanced AI labs verify that global rivals have actually stopped or slowed their work, "and that a bad actor could not use the auspices of a coordinated slowdown to jump ahead in secret."
The authors wrote: "In the absence of a coordinated, global slowdown, we are left with the current situation: powerful technology being developed at breakneck speed by a variety of actors in a variety of countries, locked in a competition with one another where commercial and geopolitical rivalries are drowning out the larger existential-to-the-species aspects of the technology being built."
The analogy Anthropic used is international arms control — specifically the verification frameworks built around nuclear weapons, where the problem is not just building the weapon but knowing that your rival has stopped. AI presents the same challenge in a harder form: code can be trained anywhere, on any hardware, by any team, without the physical signatures that nuclear programmes leave. A pause without verification is a unilateral concession. A pause with verification requires a global governance architecture that does not yet exist.
Anthropic said its internal research institute plans to explore the issue in collaboration with others and "take actions" to help build the systems that a credible slowdown or pause would require, without being more specific. What those systems look like — technically, legally, geopolitically — is the question that governments, researchers, and institutions around the world now need to answer. The paper is the opening of a conversation, not the end of it.
The tension at the centre of this story
There is something worth sitting with before moving to implications. Anthropic is a company approaching a $1 trillion valuation. It has just released its most powerful model to the public. It is preparing for an IPO that could be one of the largest in technology history. And on June 4, twelve days after Fable 5 went live, it published a paper saying the industry should slow down.
Calling for a global development pause right before going public is either an act of extraordinary corporate conscience or an exceptionally savvy positioning move. A coordinated slowdown would freeze the competitive landscape at a moment when Anthropic is already one of the top players. New entrants get locked out. Existing leaders get to consolidate. Aiandnews
Both things can be true simultaneously. A company can genuinely believe that recursive self-improvement poses civilisational risk and also benefit commercially from a pause that freezes the competitive landscape in its favour. The value of the warning is not negated by the fact that the warning-giver has something to gain from being heard. What matters is whether the underlying claim is true — and the internal data Anthropic published suggests it is more than speculative.
AI task-completion horizons have been doubling roughly every four months. If that trajectory continues without a corresponding advance in alignment research, oversight tools, and governance infrastructure, the gap between what AI can do and what humans can verify it is doing will widen until verification becomes practically impossible. That is the concern. Not a rogue AI making deliberate choices against human values. A system improving itself faster than the humans responsible for it can understand what it has become.
What this means for the philosophy of education — and for AI Ready School
AI Ready School has covered this territory every week since April. DeepSeek made frontier AI free. GPT-5.5 made agentic AI real. Claude Mythos made it too dangerous to release. Snap fired 1,000 people and said AI wrote 65% of its code. Harvard published a study in Science showing AI outdiagnosing doctors. NVIDIA taught machines to think about the physical world. And Anthropic released Fable 5 to anyone with an internet connection.
Now the company behind the most powerful AI on earth is saying: we may be moving too fast for our own safety frameworks. And the question for every school reading this is not whether to have an opinion on global AI governance. It is what all of this means for the children in your classrooms today — the children who will inherit the world that Anthropic's paper is worried about.
The answer AI Ready School has built its entire philosophy around is this: the most important preparation is not technical. It is human. A child who understands how AI works, who has built with it, broken it, and learned where its reasoning fails, is not a child who will be surprised or overwhelmed when AI begins to improve itself. They are a child who will ask the right questions. Who will know what to look for. Who will understand, intuitively, why governance matters — because they have already experienced, in a NEO lab or through Cypher's guided thinking model, the difference between an AI that is directed well and one that is left to run unchecked.
Anthropic's central recommendation is that the world should have the "option" to slow or temporarily pause frontier AI development "to enable societal structures and alignment research to keep up." The societal structures that need to keep up are not just governments and laboratories. They are schools. The alignment research that matters most for the next generation is not the technical kind — it is the educational kind. The kind that produces citizens who can evaluate AI's outputs with genuine judgment, take responsibility for the decisions made with its assistance, and ask whether the system doing the work has been directed toward something worth doing.
That is the education AI Ready School was built to provide. Not because the future is frightening. But because it is consequential — and consequential futures require prepared humans, not just powerful machines.
The line from the paper worth reading in every school's next staff meeting
"In the absence of a coordinated, global slowdown, we are left with the current situation: powerful technology being developed at breakneck speed by a variety of actors in a variety of countries, locked in a competition with one another where commercial and geopolitical rivalries are drowning out the larger existential-to-the-species aspects of the technology being built."
The existential-to-the-species aspects. Written by the people building the technology. Published on June 4, 2026. Schools that read this sentence and treat it as someone else's problem are not being cautious. They are being absent from the most important conversation of the decade.
