A few months ago, an idea as solemn as it was absurd reappeared: copying the Cold War treaties to control the development of artificial intelligence (AI). The proposal, promoted by some American businessmen and academics, says that just as the United States and the Soviet Union limited the expansion of nuclear weapons through international agreements, now the world should do the same with AI. The comparison sounds serious, but it does not stand up to the first analysis. It is a nostalgic attempt to apply recipes from the last century to a phenomenon that does not even obey the laws of the physical world.
During the Cold War, the problem was clear: two powers with atomic arsenals capable of destroying the planet needed to avoid a direct confrontation. Each bomb had material, weight and location; it could be counted, measured and monitored with satellites. From that materiality arose the possibility of mutual control. If one made one more warhead, the other could detect it. That’s why treaties worked: there was symmetry of power, shared technical knowledge, and a tangible object that could be verified.
AI has none of that. There are no bombs, no uranium, no silos hidden in deserts. There are data centers: they are huge warehouses of servers where electrical impulses and mathematical calculations circulate that no one fully understands. Most modern systems operate like black boxes: they produce results that not even their designers can explain. When you train a language model or a deep neural network, what you obtain is not a set of rules but a network of statistical correlations so complex that it is no longer interpretable. That is why the idea of “monitoring” or “verifying” AI development is a fantasy. It would be like signing a treaty to control dreams: brain waves can be recorded, but not the content.
The problem is not technical, but epistemological. There is no way to open the black box and really understand what is going on inside. Machines learn in ways that we cannot follow step by step; its logic is emergent, not programmed. In the nuclear case, inspectors measured radiation or identified prohibited materials. In AI, the only visible evidence is power consumption or the number of chips used, which says nothing about what the machine learns. Aspiring to regulate something that no one understands is equivalent to trying to write a penal code for chance.
And even if that impossibility were magically resolved, another major obstacle would remain: the absolute imbalance of power. In the Cold War there were two rival blocs; now there is only one. The United States controls the entire material chain of AI: the most advanced chips, the machines that manufacture them, the training software, the data centers, the platforms and the investment funds. No other country has access to that entire ecosystem. China can make backward versions, but it depends on Western components. Europe lacks mass training companies and Russia does not appear directly. There is no balance, and without balance there is no possible treaty. A pact requires pairs; However, today there is a monopoly.
The rest of the world watches as a spectator to a race that has already ended. Pretending to sign an international agreement between “technological powers” is ignoring the fact that these powers do not exist. There are not two AI armies competing: there is a single bloc with full control over infrastructure and knowledge. It is no longer even about the United States as a State, but about private conglomerates with more resources than most countries. To think that they will voluntarily submit to a global inspection system is an act of political naivety or deliberate hypocrisy.
The idea of a “verification treaty,” furthermore, is based on a false assumption: that AI is a risk that can be contained with regulations. In reality, the problem is not that someone “makes” dangerous superintelligence, but that we are all increasingly dependent on systems that no one controls. The danger is not in an explosion, but in the slow and total absorption of human decision-making capacity by automatic processes. There is no missile to deter, but algorithms that replace essential functions without anyone understanding or being able to audit them.
Behind the rhetoric of the treaties lies an attempt to calm consciences: to pretend that humanity is in control while true authority moves to structures that do not respond to governments or laws. It is a form of diplomatic self-deception. Talking about “international verification” in this context is like proposing a microscope to measure faith.
The new regulatory discourse of the Cold War applied to AI is not a rational project: it is a staging to disguise the void. It seeks to organize a universe that no longer has a physical border or political balance. Nuclear weapons could destroy the Earth; AI, on the other hand, silently dissolves sovereignty over the real. And you cannot put a seal or a treaty on that: you can only understand it, if there is still time to do so.
Things as they are.
Mookie Tenembaum tackles tech topics like this every week with Claudio Zuchovicki on his podcast Artificial Intelligence, Financial Perspectivesavailable on Spotify, Apple, YouTube and all platforms.
Image gallery
