Creating an AI cartel under the guise of geopolitics

There is a misconception that we have been carrying around for years. It is the idea that, in a world like ours – where governments mediate up to half of GDP and never stop producing new regulations – the aim of self-interested capitalist enterprises is to operate without rules. In other words, that their efforts are focused on reducing the regulatory burden. If that were the case, a major industry such as lobbying and public affairs would be a spectacular failure. In the United States, reported expenditure on federal lobbying reached $5.3 billion in 2025. This is one and a half times the combined cost of the SEC, FTC, FCC and CFTC – that is, four of the main federal market regulatory authorities. Clearly, this money is not being spent to secure fewer rules.

Having a light regulatory framework benefits society, as it facilitates the creation of new businesses. It is Peter Thiel’s long-standing argument: we have innovated in the realm of bits rather than atoms because the processes involved in creating new software are far less regulated than those required to develop a new engine or a new chemical compound.

Companies, as a rule, want rules—obviously, ones that suit them. This point is worth bearing in mind when reading Dario Amodei’s appeal to slow down the pace of AI development. The public reaction was predictable: if the leading player in a particular sector expresses fears about its evolution, it means the threat is real. But that player also faces PR problems, on the one hand, and the need to maintain its dominance on the other. This is by no means a given in a business environment which, in fact, did not even exist five years ago and in which the cards could be reshuffled at any moment.

What evidence is there that AI is about to get out of hand? Amodei cites a documented incident. Last July, a number of OpenAI evaluators, engaged in an internal benchmark of offensive capabilities with production safeguards deliberately disabled, escaped from the isolated environment in which they were operating by exploiting an incorrect configuration, coordinated with one another via a shared noticeboard and compromised Hugging Face’s infrastructure, whilst attempting to tamper with the system that was assessing their performance. There are three accounts of the incident: the METR investigation, OpenAI’s report of 26 August and Hugging Face’s forensic timeline.

The inference drawn from this, however, is of a different nature. From the fact that a ‘swarm of agents’ behaved maliciously on a test range with the safety catches removed, Amodei concludes that within six to twelve months a similar swarm could take over the entire network via a persistent botnet, causing hundreds of billions in damage. This is a prediction, and his article does not explain how he arrived at it.

In practical terms, what Amodei is proposing is a sort of cartel amongst companies that share two characteristics: they are, at present, the largest, and they are American. The structure of the proposal is instructive. At its heart are embedded third-party evaluators, with permanent access to training and deployment processes: Anthropic would adopt this unilaterally and is calling on governments to require competitors to do the same. It is the classic model of raising rivals’ costs: the constraint I impose on myself only becomes a competitive advantage if the state extends it to those who have not imposed it on themselves. Bruce Yandle called this ‘Bootleggers and Baptists’, with the difference that here the preacher and the bootlegger are one and the same.

Those involved, moreover, picked up on it straight away. Elon Musk tweeted that “Dario is right”; Altman wrote that he agreed on the need to manage the pace of development, adding that the idea of independent evaluators is excellent and that OpenAI will adopt the same strategy. It had already been reported the day before that leading research laboratories were discussing an informal agreement to stagger the release of models. To legitimise a cartel agreement in broad daylight, of course, an excuse is needed, and today that can only be geopolitics. Donald Trump also responded, maintaining the same rhetorical tone: we must not hold back the development of AI because whoever wins this game wins it all.

It is somewhat unsettling to see this competitive framework applied to a textbook example of ordinary scientific research carried out by private companies. These companies operate in the ‘market’: they cater to real consumers, not just governments, even if their business model is not entirely clear at present. Their research builds on decades of deep interconnections between scientists and research systems across different countries. What, then, are we afraid of? That Chinese AIs will answer questions from Western users according to the tenets of dialectical Marxism? Artificial intelligence systems draw on scattered information; they have used self-generated texts to refine their linguistic abilities, but when they carry out research, they immerse themselves in the texts we ask them to read. DeepSeek will be no more ‘communist’ than ChatGPT is in a humanities department: because each one simply seeks out what it is asked for.

Are we afraid that terrorist groups or criminal organisations might exploit them for their own gain? To a certain extent, it is impossible for this not to happen: but this is true of any human discovery. Anthropic, to take Amodei’s company as an example, publishes regular reports on the abuses it identifies and blocks, and two days before the essay was published, it released one on attempts to develop dual-use biotechnologies – that is, technologies which, in principle, could be used to build biological weapons. This is data produced by the company itself, and should be taken as such, but it describes a problem that is being addressed through self-regulation, out of a sense of responsibility and to allay public concerns.

Do we need a law? Who can actually use AI for destructive purposes? The very same institution that commissioned the development of poison gas and then deployed it in the trenches of the First World War, and which dropped the atomic bomb on Hiroshima and Nagasaki. States. And this is not some academic hypothesis. In February 2026, the US Department of War asked Anthropic to accept, for Claude, the principle of usability for ‘all lawful purposes’, waiving the two ‘red lines’ set out in the company’s policies as contractual exceptions: mass domestic surveillance of Americans and use in fully autonomous lethal weapons. The dispute therefore concerned not only those two possible uses – which the Pentagon stated it did not intend to pursue – but also who was entitled to set the limits. A senior official at the Department explicitly formulated the principle: a technology deeply integrated into the military apparatus must be under the control of democratically elected or appointed authorities, and no private company may unilaterally impose ‘regulatory terms of use’ on national security systems.

And should we seek to protect ourselves from the radical uncertainty surrounding the future of AI by placing it in the hands of the public sector, as some commentators advocate? There is a problem – and a big one at that. But not with LLMs. It’s in our own heads.