The real risk may not only be that AI surpasses us, but that we become incapable of living without it.

We are beginning to understand that artificial intelligence already adapts its answers to the level it attributes to the person it is talking to. It simplifies, popularizes, selects and removes part of the complexity in order to give us an answer that we are supposed to be able to absorb.

But there may be an even deeper danger.

Nature of the analysis. This text is a forward-looking essay: it does not announce a future outage and does not claim to calculate its probability. Rather, the sources used document the need for human oversight, for skills that make it possible to understand the systems, and for explicit risk governance.[1][2][3]

What we are going to delegate

As artificial intelligence becomes more capable, we will delegate to it not only the search for information but, gradually:

The problem will therefore not only be that artificial intelligence will be able to do more things than we can.

The problem will be that we will gradually stop knowing how to do them without it.

During a transition period, this dependence will give us the illusion of a tremendous increase in our capabilities. An individual assisted by artificial intelligence will be able to write like a specialist, analyze like a researcher, program like an engineer and diagnose like an expert.

But will this power truly be his own?

Or only an external power to which he has temporary access?

Having a skill or having access to a service

There is a fundamental difference between having a skill and having access to a service capable of performing it in our place.

Today, someone who uses a navigation system knows how to reach their destination, but does not necessarily know how to read a map. Someone who uses a calculator knows how to get a result, but cannot always redo the calculation. Someone who has a computer program generated can obtain working software without understanding precisely why it works.

Tomorrow, this delegation could affect the entire production chain.

The engineer will no longer really calculate the structure: he will validate the artificial intelligence’s proposal.

The doctor will no longer build his diagnosis entirely himself: he will choose between the system’s hypotheses.

The mechanic will no longer work out the repair procedure himself: his glasses will tell him which part to remove and in what order.

The worker will no longer memorize the movements: the robot will perform them.

The farmer will no longer necessarily know the signs that announce a plant disease: sensors, drones and algorithms will detect it.

We will then continue to produce, to heal, to build and to administer. But a growing share of actual knowledge will no longer reside in human beings. It will reside in a technical infrastructure made up of software, models, data centers, power grids, international links and automated machines.

What would happen if this infrastructure suddenly stopped working?

This is where a dizzying question arises.

Not necessarily everywhere and forever, but for long enough and over a large enough area.

A large-scale electromagnetic pulse, a war, a coordinated cyberattack, the destruction of power grids, a lasting breakdown of telecommunications, the sabotage of several critical nodes or a combination of crises could make inaccessible the services on which entire societies would have come to rely.

An electromagnetic pulse would not necessarily "blow up every data center" the way a single bomb would destroy a building. It could, however, disrupt or damage electrical and electronic equipment and, above all, trigger a crisis in the infrastructure on which data centers depend: electricity, cooling, telecommunications, fuel, components and maintenance. The resilience recommendations devoted to this risk show precisely that it is not a purely imaginary hypothesis.

The same goes for undersea cables. Destroying one cable would obviously not wipe out the Internet. Networks are meshed, communications can be rerouted and redundancies exist. But these cables remain critical infrastructure: they carry the vast majority of intercontinental communications, and their vulnerability is now the subject of specific plans by the European Union and NATO.

The credible risk is therefore probably not that of a single button that would permanently switch off all artificial intelligence worldwide.

The risk is that of a cascade:

failing electricity, interrupted telecommunications, isolated data centers, impossible access to the models, disorganized industrial systems, paralyzed supply chains, delayed maintenance and human staff unable to quickly take over manual operations.

The Egyptian precedent

And this is where the historical parallel becomes interesting.

In ancient Egypt, administrative, religious, astronomical, medical and technical knowledge was largely concentrated in a limited number of institutions and passed on by specialized communities. When the structures that trained the scribes and maintained the texts declined, it was not only people who disappeared: the very continuity of understanding was broken.

The inscriptions remained carved in stone.

But no one could read them any more.

Tomorrow, we could experience a comparable situation in digital form.

The knowledge might still exist somewhere, on disks, in backups or in the parameters of a model. But stored knowledge is not necessarily usable knowledge.

You still need:

A society can therefore possess a gigantic quantity of information while gradually losing its autonomous capacity to turn it into action.

The paradox would be terrible: never would humanity have accumulated so much knowledge, but never would it have been so dependent on a technical intermediary to access it.

The real catastrophe would not only be the temporary loss of artificial intelligence.

It would be to discover, at the very moment it disappears, that we had at the same time allowed the humans capable of replacing it to disappear.

We could then find ourselves with engineers unable to recalculate without software, administrations unable to function without automation, technicians unable to make a diagnosis without assistance, farmers deprived of the systems that run their farms, and citizens who know how to ask for a solution but no longer know how to build one.

To use a deliberately trivial image: we could become extraordinarily powerful when everything works, and then find ourselves completely helpless in front of an adjustable wrench when the system stops.

This danger has a name: the loss of skills, or deskilling

The more we delegate a task, the less we practice it. The less we practice it, the less we know how to pass it on. And if the next generation enters directly into a world where artificial intelligence does the basic work, it may never acquire the foundations needed to understand or take over the system. Several recent studies highlight precisely this risk: artificial intelligence can immediately increase productivity while reducing opportunities for learning, deliberate practice and the intergenerational transmission of know-how.

We should be talking about human backups

That is why resilience in the face of artificial intelligence should not be solely a matter of computing.

We talk a great deal about backups, cybersecurity, redundant data centers, digital sovereignty and protected networks.

But we should also be talking about human backups.

In every essential field, we should preserve:

Perhaps we should apply to human skills the same principle of redundancy that is used for critical infrastructure.

You do not build a power plant with a single safety system.

You should not build a civilization with a single operational source of intelligence.

Artificial intelligence could become the most formidable capability multiplier humanity has ever created. But if we use it to gradually eliminate all redundant human skills, it could at the same time become our greatest single point of civilizational failure.

The challenge is not to reject artificial intelligence

That would be absurd, and probably impossible.

The challenge is to use it without ceasing to learn, understand, verify, repair and pass on.

In other words:

Artificial intelligence must augment human capabilities, not become the only place where those capabilities are kept.

Perhaps the great question of the 21st century will not only be:

"How far can artificial intelligence surpass human beings?"

But also:

"Which skills must human beings absolutely keep for the day artificial intelligence no longer answers?"

The ultimate risk is not that AI becomes more intelligent than we are. The risk is that we become less capable without it.

A civilization does not collapse only when it loses its libraries. It also collapses when its libraries still exist, but no one any longer knows how to read, understand, make or repair.

We will therefore have to protect data centers, networks and models. But above all we will have to keep "human backups": men and women capable of healing, calculating, building, farming and repairing when the screen goes black.