The AI Watchdogs Are Multiplying — and None of Them Are Talking to Each Other

For most of the last three years, the story of frontier AI was a story of acceleration with almost no one at the controls. Models got bigger, capabilities jumped, and the question of who — if anyone — should be allowed to slow things down stayed politely unanswered. That is changing fast in the summer of 2026. The striking thing is not that oversight is finally arriving. It is that it is arriving from everywhere at once, in pieces that do not yet fit together.

Consider a single week in July. A frontier lab, a national government, a central banker and a US state each moved to put some kind of leash on AI — and none of them was coordinating with the others.

The lab that asked to be watched

Start with the most surprising voice: the industry itself. Demis Hassabis, the Nobel laureate who runs Google DeepMind, published a personal manifesto calling on the United States to stand up a new AI watchdog before year's end. His model is FINRA — the private, industry-funded body that polices Wall Street under government oversight. Frontier labs would share their most advanced models with the watchdog, initially on a voluntary basis and up to 30 days before release, so that independent experts could probe them for dangerous cyber, biological and deception capabilities. Once the regime proved "effective and robust," Hassabis argued, formalization "could quickly follow" — meaning models would eventually have to pass before they could ship. Crucially, he wants the rules to apply to every frontier-class model regardless of country of origin, and regardless of whether it is open or closed.

It is a remarkable position for the head of one of the labs doing the accelerating: build me a referee with the power to make me wait.

Nations reaching for a framework

Governments are moving too, though along their own lines. In Australia, Prime Minister Anthony Albanese announced an Office of AI to sit inside his own department, tasked with building a single national framework and Australian standards for the technology. The office pulls several threads into one place: safety standards, the fraught question of AI and copyright — where Canberra has refused to grant the free-use exemption that the US, EU and Japan have entertained — and the physical footprint of the boom, including new obligations on the data centres that AI demands. The pitch, in the government's own framing, is that Australia should be "more than a data warehouse."

The instinct is the same as Hassabis's — impose order on a runaway technology — but the vehicle could hardly be more different: a cabinet office rather than an industry self-regulator, national rather than global in reach.

The central banker's warning

Then there is the view from the financial system's guardians. Andrew Bailey, governor of the Bank of England, has named frontier AI as the single biggest issue on his desk, warning that AI-driven cyber tools have pulled risks the Bank thought were years away "right into the foreground." His prescription is explicitly internationalist. No country, he argues, can seal itself off from threats that cross borders by nature; even the United States "can't achieve what it sensibly wants to achieve… on its own." What Bailey wants is stronger, coordinated testing of models before they enter wide circulation — a financial-stability regulator, in other words, asking for something that sounds a lot like Hassabis's watchdog, from an entirely separate corner of the system.

The state that just said no

Finally, the bluntest intervention of all came not from a summit but from Albany. New York became the first US state to impose a statewide moratorium on new hyperscale data centres — a one-year pause on state environmental permits while regulators build a framework to protect ratepayers, the power grid and water supplies. This is not AI safety in the model-testing sense. It is AI governance by infrastructure: if you cannot regulate the software, you can at least regulate the enormous physical machine it runs on. Other states are already studying New York as a template.

Four watchdogs, no kennel

Line these up and the pattern is unmistakable. A lab wants a US-led global standards body. A national government is building a domestic office. A central banker wants international coordination on systemic risk. A governor is pulling permits for the buildings. Each is a legitimate response to a real hazard — and each is aimed at a different layer of the same problem, with no shared definition of what "frontier AI" even is, let alone who has authority over it.

That fragmentation is not a footnote to the governance story. It is the story. The comforting version of AI oversight imagined a single grand bargain — a Bretton Woods for algorithms. What is actually emerging is messier and more revealing: overlapping, uncoordinated jurisdictions, each grabbing the piece of the elephant it can reach. Whether these strands eventually braid into a coherent regime, or harden into a patchwork of conflicting rules that labs can arbitrage, is now one of the defining questions of the decade. For the moment, the watchdogs are multiplying faster than anyone is teaching them to work together.

Sources & further reading