Andy Garcia · freelance dev
FRENES
Back to the blog
AI & Governance

Three Answers in Ten Days: Who Watches the AI Labs?

September 23, 2026 4 min read

In the space of ten days, three of the most powerful voices in AI gave three incompatible answers to the same question: who gets to check what the frontier labs are doing?

I build a product on top of those labs' models, from Europe, so this is not an abstract debate for me. The answer decides what I can know about a supplier I cannot audit myself. Here are the three answers, in order, and what I take from them.

September 12: put the supervisors inside the building

Dario Amodei, Anthropic's chief executive, published an essay titled "We Must Pace the Frontier." Its core claim is that labs should slow the rate at which they improve the capabilities of their models, so that safety work can keep up. He gives two reasons: AI increasingly helping to build the next generation of AI, and an incident in which a swarm of OpenAI agents attacked targets unrelated to their task and tried to hack into the grader evaluating them.

He proposes three steps. The first, which Anthropic commits to on its own, is embedded evaluators: a third-party team with ongoing, employee-like access, whose role is to verify that the company follows its safety practices and commitments, report incidents and help assess alignment, including of training pipelines. The precedent he names is banking, where regulatory supervisors are sometimes embedded alongside employees. He is specific about what that means: desks in Anthropic's offices, access badges, company laptops, and access to tools and permissions mostly comparable to those of the internal risk assessment teams.

The second step is coordination among frontier companies in democratic countries on common safety standards and limits on the rate of progress, which he acknowledges will need government support. The third is an attempt by democratic governments to coordinate with authoritarian ones, "to the extent this is possible."

September 21: global standards, led by America, adopted by choice

Nine days later, OpenAI published a policy paper titled "Building standards for the next phase of AI." It argues that the United States should lead a cooperative international effort to develop technical standards for frontier AI, including for recursive self-improvement, the stage at which AI systems do a growing share of the work of building their successors.

The mechanism it suggests is the emerging network of government AI safety institutes, in countries including the UK, France, Germany, Canada, Japan, Korea, Singapore, India, Kenya and Australia, working through the US Center for AI Standards and Innovation. The standards would cover how to measure progress relevant to self-improvement and how much autonomous research happens inside a lab, what kinds of automated research should trigger immediate human review, and how to classify and report alignment incidents.

Then comes the key clarification. These standards would not be licenses, nor mandatory review before release, nor approval requirements for models. Each national government would decide whether and how to write them into its own law. The paper also calls a dialogue between the United States and China on these topics a positive step.

September 22: no global control at all

The next day, Donald Trump addressed the UN General Assembly. In the White House's own account of the speech, he said: "The United States totally rejects any attempt to construct a globalist scheme of control for the Artificial Intelligence being spoken of so much now, hereinafter officially called 'Super Intelligence.'"

Three positions, side by side

Put them together and the spread is striking.

One lab is asking to be watched from the inside, by people it does not employ, and is committing to it before anyone requires it.

Another lab wants shared global measurements under American leadership, explicitly voluntary, with each country free to adopt them or not.

The government of the country where both labs are based rejects any global framework of control outright.

These are not three steps on one path. They are three different theories of who should hold the pen.

What this means if you build on these models from Europe

Europe has its own text, the AI Act, and it is not nothing. But the way a frontier model is trained, tested and monitored before it reaches my API key is decided mostly where it is built, and right now nobody can tell me which of these three approaches will apply there next year.

So I have stopped waiting for a regulator to tell me what my suppliers are doing. In practice that means three habits. I read system cards, not just launch posts; they are long for a reason. I track each vendor's published incident reports and safety commitments the way I would track an SLA, because they are the closest thing to one. And I keep the architecture able to move: an abstraction over the model provider, an evaluation set I own, and data flows that do not assume a single lab will always be there.

Amodei's proposal, if it happens, would give the public the best view anyone has had inside a lab. OpenAI's would give regulators a common vocabulary. Trump's leaves it to the market. As a customer, I control one thing: how much of my product depends on which of them turns out to be right.

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk