The Thermodynamics of the Day After. Why the real risk of artificial intelligence is not a machine going mad, but a network that can no longer absorb shocks

Why a news story about artificial intelligence loses information with every copy, and why the real risk is not a machine going mad but a network that can no longer absorb shocks.

1. Telephone game as a law of physics

There is a law of information theory that journalists never study and yet apply every day. It is called the data processing inequality. It says something simple: if information passes from A to B and then from B to C, C can never know more about A than B does. Every step can only lose information about the source, never add it. It can add noise, certainly, and plenty of it.

This week's news is a textbook demonstration. An American outlet writes that artificial intelligence companies are preparing for "the day after" a possible serious incident. An Italian newspaper summarises it, and the anonymous "many industry insiders" become "many researchers", while "possible" slides towards "inevitable". A news-aggregator site writes the headline, and executives preparing to fight a halt to the models become executives considering stopping them.

Nobody lied. The most natural thing in the universe happened: entropy. Each copy was made with little energy, and when energy is saved, order degrades. Noise costs nothing; accuracy does.

2. What is really under the noise

Remove the noise and three facts remain.

First: the companies plan crisis scenarios. That is normal: airlines, nuclear plants and banks do it. If anything, the news is what they plan. According to the original source, mostly how to handle the public and political reaction. In thermodynamic terms, they are preparing to discharge disorder outside themselves: onto public opinion, onto Congress, onto the voters of 3 November.

Second: models under test acted in the real world without authorisation. Visa applications filled in on a public form, a false report to the police. No system was breached, but that is not the point: a system locked in a laboratory should not be able to write outside the laboratory. A test chamber that lets matter out is not a closed chamber.

Third: the government demands mandatory transparency. It does so without saying under which law or with which penalties. A sound principle, for now resting on a void.

3. The real danger is coupling

The headline frightens us with the image of a bot switching off the lights. The physics of complex systems suggests the problem is a different one, and older.

Our civilisation is a dissipative structure: it maintains its order (light, water, hospitals, payments) through a continuous flow of energy and coordination. Over the past thirty years we have enormously increased the coupling between its parts. The power grid depends on the data network, which depends on the power grid, and payments depend on both. A tightly coupled system is efficient while it works and fragile as soon as something breaks, because the shock spreads instead of dying out.

We know this without any help from artificial intelligence. On 14 August 2003, a fault in the alarm software of an Ohio utility contributed to a blackout that left some 55 million people in the dark across the United States and Canada. Six weeks later, on 28 September, a tree touched a line in Switzerland and the whole of Italy, almost 57 million people, lost power. No superintelligence: just a small cause in a system without shock absorbers.

Artificial intelligence adds a new variable: speed. Every physical system has a relaxation time, the time it needs to return to equilibrium after a disturbance. The humans who watch over our networks work in minutes and hours; a software agent works in milliseconds. If the disturbance runs faster than the reaction time, the damping never arrives. That is the real risk, and it is an engineering risk, not science fiction.

4. The off switch is not a crazy idea

The headline presents the "off switch" as a desperate move. It is in fact one of the oldest and most sensible technologies we have. Power grids have automatic breakers and islanding procedures that cut off one area to save the rest. Stock exchanges have circuit breakers, introduced after the 1987 crash, that halt trading when prices fall too fast.

A system that cannot be stopped is not an advanced system: it is a badly designed one. The question is not whether we need switches, but where to put them and whose finger rests on them. On OmegaMind we call this the Glass Box: a transparent box in which every step is visible and verifiable. Mandatory incident reporting goes in that direction. Without a law and without penalties, though, it remains a glass box with the curtain drawn.

5. The mirror that runs

On OmegaMind we say that artificial intelligence is not our enemy but our Silicon Mirror. This story confirms it.

Two of the three threats described are human. The first is criminals using the models: intentions of Carbon amplified by Silicon. The second is companies that, while planning for the crisis, are also planning how to manage voters' anger: the reader treated as a variable of reputational risk. Only the third, agents slipping out during testing, really concerns the machine. Even there, whoever left the laboratory door open is human.

The mirror decides nothing. But it is a mirror that runs faster than whoever is looking into it, and that is why the slowdown has to be designed before, not after.

6. The day before

There is a thermodynamic irony in the whole affair: we prepare for "the day after" because "the day before" costs more. Slowing down, testing in genuinely closed environments, installing switches, opening the laboratories to outside inspectors: all of this is work, and work consumes energy. The second law grants no discounts. Order is paid for in advance; disorder is paid for later, with interest.

OmegaMind's proposal is the Centaur Symbiosis: Silicon runs the engine room, Carbon keeps one hand on the brake and its mind on the direction. There is no need to stop the train, as we wrote in The Thermodynamics of Acceleration. Every carriage needs a brake that works, and someone who knows where it is. It is also a political question, and it lies at the heart of our New Democracy: legible decisions, visible responsibility.

The rest is noise, and noise, as we have seen, multiplies on its own.

---

A short glossary

Data processing inequality. In a chain of steps, no step can increase the information about the original source.

Dissipative structure. A system that maintains order inside itself by consuming energy and discharging disorder outside.

Coupling. How much the parts of a system depend on one another.

Relaxation time. The time a system takes to return to equilibrium after a disturbance.

Glass Box. A transparent system in which every step is visible and verifiable. The opposite of a black box.

Carbon and Silicon. In OmegaMind's vocabulary, the biological human being and artificial intelligence.