The increase in information is not accompanied by a corresponding increase in the time required to interpret it. This compression is the source of the strategic problem. Almost all of modern military history could be reinterpreted through the lens of this tension. The telegraph accelerated command. Radio further shortened the time between observation and order. Radar provided a few extra minutes in the face of an aerial threat. Satellites extended the range of observation. Digital networks then virtually eliminated certain distances. Artificial intelligence now adds a different kind of transformation:

It no longer simply accelerates signal transmission.

It accelerates interpretation. That, to my mind, is where the real change begins. When a machine can detect, classify, compare, prioritize, and propose an action before a human operator has reviewed all the data, the strategic problem shifts toward decision time itself.

To show something.

To prevent seeing anything else.

Saturate the system with contradictory information.

Cognitive warfare also affects time.

To create a false sense of urgency.

To delay a decision.

To speed up another one.

Cognitive warfare, therefore, attacks the adversary's model of the world less directly than it does the body itself. It is a particularly profound form of warfare because all military action begins with a representation.

Before firing, a target must be identified.

Before mobilizing, one must have interpreted an intention.

Before retaliating, you must have decided that an attack is real.

If the representation is false, all available power can be used against an illusion. This idea is ancient. Strategists have always used deception, camouflage, and diversion. What's different today is the scale. Artificial intelligence can produce, detect, and analyze signals at a speed no human team could sustain.

This turns deception into an industry. Thousands of fake accounts, millions of pieces of content, synthetic images, cloned voices, and credible documents can be produced almost instantly. But it would be simplistic to limit cognitive warfare to deepfakes. The real problem is much broader:

How does a system know what it should believe? That question applies to an artificial intelligence as much as to a society. The difference is that AI can process a gigantic quantity of information and therefore absorb a gigantic quantity of manipulation as well. It becomes at once a detection tool and a new attack surface.

It is possible to try to deceive a human operator. Now, it is also possible to try to deceive the model that assists that operator. This is where a new depth of conflict emerges. The adversary is no longer simply trying to hide a vehicle. They may be trying to produce a signature that leads the algorithm to classify it in the wrong category.

It is no longer just trying to jam radar. It can try to disrupt the data chain that feeds artificial intelligence. This war against algorithmic perception seems crucial to me. We are entering a world where systems will have to learn to ask themselves not only:

What do I see?

But also :

Is someone trying to make me see this?

The Blind Spot—the blind spot—extends here into the strategic field. Every intelligence has limits of perception. But when an adversary knows them, it can try to exploit them. A bias then ceases to be merely a defect. It becomes a strategic vulnerability. That is why algorithmic biases must also be viewed through a security lens.

A model that overemphasizes certain signals, misclassifies certain situations, or generalizes too quickly can be manipulated by an actor who understands its logic. This produces an almost biological dynamic: adversarial competition. In living organisms, predator and prey co-evolve. One develops camouflage, the other improved perception. One accelerates, the other becomes faster. Military AI systems will likely enter into this ongoing co-evolution. Every improvement in detection will lead to an improvement in camouflage. Every new identification system will spawn a method to deceive it.

The race never stops. That is precisely the mutant's logic. Only mutants survive because the environment itself learns. Technology frozen in adaptive conflict quickly becomes obsolete.

This dynamic is particularly troubling when it reaches the decision loop. Military doctrine often refers to the observation-orientation-decision-action cycle. Whoever completes that loop faster can impose their tempo on the adversary. Artificial intelligence accelerates each of these stages. It can observe more sensors, orient by comparing more scenarios, assist decision-making, and sometimes trigger an automated action.

The temptation is then immense:

Gradually eliminating human delays to gain a few seconds. And that is precisely where the danger begins. A second saved can constitute a tactical advantage.

But eliminating too many seconds can also eliminate the time needed to recognize a mistake. Speed is therefore both a weapon and a risk. This is one of the fundamental paradoxes of the era of military AI. For a long time, the strategic advantage belonged to whoever could decide fastest.

Now, it could also belong to those who know when to slow down. This idea seems essential to me. Intelligence is not always about acceleration. Sometimes, it consists of deliberately introducing latency.

A critical system must be able to signal: this signal is ambiguous enough to require a second verification. This slowdown may seem ineffective from a purely competitive standpoint. But it can prevent an irreversible escalation. This is where de-escalation engineering becomes a true technical architecture.

We have too often imagined de-escalation as a diplomatic problem separate from technology. This is no longer possible. The system itself must contain brake mechanisms, just as an industrial machine has emergency stops.

No one considers the stop button a weakness of the machine. On the contrary, it is a condition of its safety. We should think the same way about military systems augmented by AI.

What is the stop button?

Who can take back control?

In what situations should the system refuse to act automatically?

What signals require two independent sources?

What degree of uncertainty does human intervention impose?

These are questions of architecture, not just ethics. A security doctrine that does not become code, procedures, interfaces, and chains of command remains an intention. This aligns perfectly with everything I have learned in the industrial world.

Safety is not something that is proclaimed. It is something that is conceived. A dangerous device must be designed in terms of possible defects.

What happens if this sensor fails?

What if this data is incorrect?

What if this operator misinterprets the alert?

What if two systems give contradictory answers?

Military AI must be subject to the same rigorous standards. A model cannot be considered reliable simply because it achieves excellent average performance. In a strategic system, averages are insufficient. The nature of the extreme error also matters. A model can be correct in 99.9 percent of cases and catastrophic in the remaining thousandth if that thousandth triggers an irreversible response.

This is precisely the difference between commercial optimization and critical safety. A flawed advertising recommendation costs little. An incorrect classification in a defense chain can lead to deaths or even war. Therefore, we need a shift in our metric culture.

Engineers will need to learn to measure not only accuracy, but also the severity of false positives and false negatives, uncertainty, and robustness under attack. It is a science of error before it is a science of performance. This expression seems apt to me. Security consists first and foremost in understanding how the system can make mistakes.

The same applies to human beings. The great strength of an experienced decision-maker is not in never making mistakes. It is in recognizing certain forms of their own errors. Therein lies the blind spot.

If artificial intelligence is to become truly strategic, it will need to be able to map its areas of uncertainty. This would represent a significant shift from the fantasy of the infallible computer. A truly useful AI should not simply provide answers; it should produce a topography of trust.

Here is what I consider highly probable.

Here's what's less so.

Here is the missing data.

Here are the competing hypotheses.

Here is what, if true, would invalidate my scenario.

This last ability seems essential to me:

Knowing what could invalidate one's own interpretation. A machine capable of organizing contradiction would be far more valuable than one that merely accumulates confirmations. And this relates to a very concrete military issue. In a crisis situation, an organization naturally seeks signals consistent with the threat it fears.

This is a well-known human bias. AI can either correct this bias or amplify it. If it is trained to search only for what the general staff asks it to, it can produce an enormous amount of confirmation power.

Every new signal is then interpreted in line with the dominant scenario. The system becomes closed-in.

It is almost algorithmic paranoia. Everything becomes evidence. That is a major danger. A sound architecture should force the system to produce a contradictory case.

If the primary scenario is an attack, what data supports the scenario of a drill, a mistake, or a diversion?

If the probability of hostile action increases, what factors reduce it?

This counter-analysis could be automated. One could almost imagine two opposing models: one tasked with constructing the main hypothesis, the other with attempting to refute it. A kind of algorithmic adversarial procedure. This aligns perfectly with my conception of law. Adversarial debate is not an obstacle to truth.

It is a method for reducing error. An important decision must be able to withstand the onslaught of opposing arguments. Applying this philosophy to military AI seems particularly fruitful to me. This would prevent a single model from becoming an absolute cognitive authority.

This pluralism of models also matters in critical systems. Depending on a single intelligence architecture creates a single point of failure. Two systems trained differently may see different things. Their disagreement becomes information.

This is an extremely important principle. Disagreement is often seen as a problem to be solved. It can become a sensor. If several truly independent models converge, confidence increases.

They may diverge sharply, in which case the system should slow down. Divergence becomes a signal of uncertainty. We return exactly to the logic of medical or scientific diagnosis. A single measurement may be wrong.

Multiple independent measures reduce risk. But independence is essential. Three models trained on the same data and built according to the same logic can reproduce the same bias. Redundancy, therefore, is not simply duplication.

It requires diversity. This is where a biological lesson returns. Monocultures are fragile. A parasite adapted to one can destroy the whole system.

Diversity breeds resilience. The same principle could become central to military cognitive systems: diversity of sensors, models, doctrines, and human teams. The most homogeneous system can be the fastest. It can also be the most vulnerable to a common error.

This tension between speed and diversity will likely be one of the major design choices. Political regimes themselves could be affected differently. A centralized system can rapidly deploy a common architecture across all its forces. This produces exceptional coordination. But if this architecture has a deep vulnerability, it will spread everywhere.

A more distributed system is more difficult to coordinate, but it can better withstand a local failure. This is precisely the logic of networks. The military command of the future will therefore have to find a balance between centralization and local autonomy. Artificial intelligence could enable remote units to operate even when communications are degraded.

This is tactically very powerful. A drone, a vehicle, or a robotic unit could continue certain missions without constant communication with headquarters. But this autonomy also increases the difficulty of control. How far should a machine be allowed to act without validation?

This is where the debate over autonomous weapons comes to the fore. We must first be precise. The word “autonomous” covers very different realities. A system may navigate by itself, select a route, avoid an obstacle, identify an object.

It may also be authorized to select a target and initiate lethal action. These levels should never be confused. Autonomy in navigation is not autonomy in lethal decision-making. Autonomy in identification is not autonomy in engagement.

Granularity of vocabulary is a moral and strategic necessity here. Saying “autonomous weapon” without specifying which function is autonomous destroys the analysis. The critical question arises when the chain of human responsibility becomes too remote. An officer may define a mission, an area, and criteria, and then a system may operate for several hours without communication.

Who is then responsible for an error?

The person who designed the algorithm?

The one who provided the parameters?

Who authorized the deployment?

Responsibility becomes distributed. However, excessively distributed responsibility can ultimately lead to no one being clearly accountable. This poses a major legal risk. The system cannot become a black hole of accountability. Therefore, precise traceability of decisions must be maintained.

What parameters were used?

Which version of the model?

What data was available?

What is the confidence level?

Which human approved what?

These records must be preserved like aircraft black boxes. The analogy with aviation seems very strong to me. Aviation has built a safety culture by analyzing every accident, every incident, every causal chain. It has understood that error is rarely the product of a single mistake. It often results from a series of small anomalies that align.

Military AI will need to develop the same culture. It should not just look for someone to blame, but understand the architecture that made the error possible. This approach is much more politically challenging because organizations often prefer to personalize the blame.

But a mature safety culture seeks the systemic cause. This is precisely what artificial intelligence should be able to do: not only identify what happened, but reconstruct the chain of decisions. This capacity for retrospective causality will become essential. We have already entered an era where many complex systems are partially incomprehensible to a single individual.

An AI-enhanced military network can combine satellites, drones, ground sensors, cyber intelligence, and human data. A decision can result from the interaction of dozens of systems. Who, then, truly understands the whole picture? Perhaps no one.

This is where systemic risk increases. We can build architectures that no human can fully represent. This already exists in finance and the internet. But war adds irreversibility.

A market can be closed.

A transaction can sometimes be cancelled.

A strike cannot be recalled after impact.

Irreversibility must therefore become a central variable in human control. The more irreversible and dangerous an action is, the more demanding the threshold for autonomy should be, in my view. Allowing an AI to adjust the temperature of a building is not the same as allowing it to identify a target. Autonomy should be conceived in relation to reversibility.

This formulation seems fundamental to me. The law could draw inspiration from it. The more a decision is likely to infringe upon a life, a liberty, or a strategic balance, the higher the levels of control, traceability, and challenge it should require. This does not eliminate AI.

This defines its role. But the military problem is precisely that the adversary may not accept this caution. If one side builds ultra-fast systems and accepts greater autonomy, the other may fear being outpaced. Caution then becomes perceived as a handicap.

This is the classic mechanism of the arms race. Each country adopts a technology not because it considers it ideal, but because it fears that the other will adopt it. This is why international governance becomes necessary. Without minimum common rules, the competitive logic drives one toward maximum speed.

But a race for cognitive speed can be more dangerous than a race for the number of weapons. If each power believes it must respond before fully verifying the situation, the overall system becomes extremely unstable. The parallel with high-frequency trading is illuminating: algorithms there already react to algorithms. We have already seen how automated interactions can produce extremely rapid and difficult-to-understand movements.

Applying this logic to a military confrontation would be dizzying. Two systems could mutually interpret each other's actions and accelerate escalation without any human decision-maker consciously intending to cross certain thresholds. It is a kind of geopolitical flash crash. This image seems important to me.

A flash financial crash can be corrected in minutes. A flash nuclear crash does not exist. There is no return to the previous situation. This is why certain categories of weapons must remain subject to particularly stringent time limits and human controls.

The reduction in decision-making time is therefore perhaps the central risk of strategic AI. We often talk about superintelligence as an abstract future. But the immediate danger may be much simpler:

An intelligence fast enough to push humans to act too quickly.

The perceived speed of the adversary influences behavior. If I believe they can detect and strike in seconds, I might be tempted to delegate more to my own systems. We then enter a loop where each automation justifies the adversary's automation. This is precisely the classic security dilemma.

The defense of one becomes a threat to the other. A faster decision system may be presented as defensive, but the adversary may perceive it as a first-strike capability. Ambiguity increases. We must therefore think not only about what the technology does, but about how it is interpreted by the other side.

Strategic stability depends heavily on perception. A capability designed to be defensive can become destabilizing if it alters the balance of deterrence. This is where transparency and declaratory doctrines become important. States will likely need to explain some of the limitations they impose on their own systems.

Not to reveal their operational secrets, but to build a minimum level of trust. For example, by stating that a certain type of decision will remain under human control. However, this statement must be credible.

Credibility requires procedures, perhaps verification mechanisms. We are back to the old problem of arms control, but applied to a software object. How do we verify that an AI system does not possess certain capabilities? It is infinitely more difficult than counting missiles.

A missile is visible. Software can be modified quickly. Controlling algorithmic weapons will therefore likely be one of the most complex diplomatic problems of the century. We will not simply be able to inspect a warehouse.

It may be necessary to certify architectures, examine logs, and control certain uses. And this obviously raises issues of sovereignty and secrecy. We are dealing with almost uncharted territory here. But that does not mean it is impossible to create standards.

States can begin with rules governing their use rather than a total ban on certain technologies. They can agree that some decisions will remain human. They can maintain communication channels regarding incidents involving autonomous systems. They can exchange information after certain anomalies to avoid misinterpretations.

De-escalation often begins with modest procedures. This is precisely the mistake of those who expect a grand, perfect treaty. Trust architecture can be built layer by layer.

A notification protocol.

A hotline.

A joint investigation mechanism.

A shared terminology.

It all seems bureaucratic until the day it prevents a war. Sometimes, details save the world. I like this idea because it ties into my obsession with invisible infrastructure. The grand historical narrative celebrates leaders and battles. But stability often depends on civil servants, engineers, procedures, and charts that will never be photographed.

Apparent chaos sometimes rests on an extremely rigorous order. This is exactly what Nicolas Detry demonstrated about the Abode of Chaos: the appearance of disorder does not mean the absence of structure. Geopolitics sometimes works the same way. The world seems chaotic because we look at visible events.

Beneath these layers lie thousands of protocols, conventions, and procedures that make interactions possible. When these structures disappear, true chaos ensues. This is why destroying diplomatic channels is always dangerous. Even between enemies, a minimal framework must be maintained.

With AI, this need increases because time is shortening. An incident that once required several hours of analysis may now only require a few minutes. Communication must therefore become faster and more direct. We can imagine that in the future, some hotlines themselves will be assisted by AI capable of translating and comparing signals, and identifying misunderstandings.

But extreme caution is necessary. The machine must not become a new filter that distorts the message. Here again, humans must be able to trace it back to the source. Linguistic traceability will be important.

Machine translation can subtly alter intent. In a crisis, a word counts. This shows how cognitive warfare begins far from the battlefield. It begins in user interfaces.

What alert does the president see?

What words does the system use?

Flashing red or cautious yellow?

A probability displayed as 80 percent or the phrase "probable attack" do not produce the same psychological effect.

Interface design therefore takes on a strategic dimension. This is a vast subject and yet often underestimated. The interface is not neutral. It structures attention. An operator surrounded by dozens of red alerts may cease to distinguish what is urgent.

A system that over-masks uncertainty creates an illusion of certainty. The way data is presented can therefore save or ruin a decision. This brings us back to the ergonomics of critical systems. Airline pilots, air traffic controllers, and power plant operators are already familiar with this problem. AI-enhanced warfare will have to learn from these fields.

How to avoid cognitive overload?

How to prioritize without hiding things?

How can uncertainty be represented?

How can we prevent an operator from mechanically following a recommendation?

These are questions that are almost as much artistic as technical, because they concern perception. Form organizes meaning. Here again, I find myself reflecting my work as a visual artist. An image does not simply convey information.

It organizes the gaze. Placement, scale, and contrast alter perception. A military interface does exactly the same thing. It is a staging of reality.

And every staged event carries a responsibility. This is a particularly interesting dimension of AI: it brings together disciplines that we had artificially separated. Computer science, psychology, design, law, and strategy become a single architecture. Cognitive warfare forces us to think of the human being in the loop not as a mere final signature, but as an organism with its own limitations.

Stress.

Fatigue.

Fear.

Bias.

A procedure that works perfectly in a calm environment can fail in a crisis. Therefore, training must include situations where the AI makes mistakes. If operators never see the machine fail during their training, they risk trusting it too much. They must learn to disobey the algorithm when circumstances demand it. This is a new skill.

We have taught soldiers how to use weapons. We will have to teach them how to challenge artificial intelligence. Human sovereignty depends on the ability to say no to the system. But to be able to say no, one must sufficiently understand how it works. This implies a higher level of technical expertise within institutions.

AI cannot be a black box entrusted to a few specialists while decision-makers simply use its findings. The command must possess algorithmic literacy. It does not need to code the model. But it must understand what a probability, a data bias, a hallucination, or an adversary attack means.

Otherwise, he cannot truly exercise his authority. Cognitive sovereignty therefore requires education for the decision-maker. This applies to the military, ministers, judges, and business leaders. We are entering a period where algorithmic illiteracy could become a form of vulnerability for those in power.

A leader who does not understand the technology on which their decision-making depends is like a ruler who cannot read their own maps. They become dependent on those who interpret them for them. That is precisely the story of power:

He who controls the information often controls part of the decision.

AI amplifies this phenomenon. Engineers, data scientists, and model providers are potentially becoming strategic players. This raises a new civil-military question: To what extent will the armed forces depend on private companies for their cognitive capabilities?

A private platform can become indispensable to national defense. This creates a very particular relationship between capital and sovereignty.

States have always worked with defense contractors. But general AI can be used simultaneously in both civilian and military sectors.

The boundary is becoming more porous. Companies that see themselves as technology-driven may find themselves at the heart of major strategic issues. They will then have to define their own ethical guidelines, while states will seek to secure access. Tension is inevitable.

The cloud, satellites, AI often belong to private actors. Public power no longer necessarily controls all the critical infrastructures it uses. This is another characteristic of the twenty-first century. The State remains legally sovereign, but part of its operational capacity rests on private infrastructure. Cognitive warfare makes this dependence even more visible. A conflict could target not only military bases, but commercial data centers, communication networks, satellite constellations. The boundary between civilian and military spheres is extraordinarily difficult to draw. . This raises questions of international law, but also of resilience.

Should a country be able to continue functioning if a private provider stops its service? Again: redundancy. Sovereignty, for me, is also the art of not having a single switch that someone else can flip. And AI is creating new switches.

Access to the model.

To the cloud.

To the processors.

Regarding updates.

Every dependency is potentially a lever. This explains why China, the United States, and Europe are each seeking to secure their own layers. But this fragmentation itself creates another risk: if adversarial systems become completely opaque to one another, misinterpretations increase. We therefore have an extraordinary contradiction.

Sovereignty encourages separation. Stability encourages a minimum of mutual understanding. Bridges must be maintained across these separate structures. Perhaps this is where international institutions need to evolve.

They were built for physical weapons, borders, states. They will have to learn to talk about models, data, autonomous systems. This vocabulary is still too new. But time is running out.

Because technology advances faster than institutions. This is almost always the case. Laws come after invention. Military doctrine too.

However, some decisions made today will create architectures that are difficult to modify later. Once an army has deeply integrated autonomous systems into its chain of command, reversing course will be costly. The lock-in effect also exists in defense. That is why initial choices matter so much.

They become norms through accumulation. The infrastructure is still there. The first protocol seems technical. Ten years later, it has become doctrine.

Security and human control must therefore be integrated from the design stage, not as a later fix. This is precisely the principle of security by design. We should also talk about de-escalation by design: designing systems to prevent ambiguity from automatically triggering escalation.

For example: requiring independent confirmation before certain actions, introducing variable timeframes depending on the severity, maintaining spaces for human intervention, recording the reasons for the recommendation, allowing for reversal. These are not obstacles to innovation. They are the conditions for its responsible use. I firmly believe that the next major military innovation may not simply be a faster weapon.

It could be a system capable of being fast when needed and slow when needed. An intelligence capable of changing its temporality. It sounds almost philosophical. But it is a matter of engineering.

AI can calculate in microseconds. It is not obligated to act immediately. Action can be conditional upon thresholds. The speed of calculation must be separated from the speed of decision-making.

This is where external memory becomes cognitive architecture. We often make the mistake of believing that a system capable of fast calculations must act fast. I do not believe that. It can use its speed precisely to examine more hypotheses before acting.

Computing power can be transformed into prudence. This is a crucial idea. A more powerful system could slow down human decision-making while simultaneously providing more perspectives. If humans once had two minutes to analyze three scenarios, AI could allow them to examine fifty in those two minutes.

Cognitive acceleration can therefore lead to strategic deceleration. This reversal strikes me as particularly interesting. It shows that technology does not have a predetermined destiny. Everything depends on the political architecture we build around it.

AI can be a machine for hastening or a machine for delaying.

Confirmation machine or contradiction machine.

Targeting machine or machine to prevent a dubious target.

Technology does not automatically dictate its use. We design it. That is why ethics should not come after the engineer. It must be part of the specifications. The lawyer, the philosopher, the military, the psychologist, the interface specialist must work with the engineer from the very beginning.

This aligns with my deeply transdisciplinary approach. Big questions do not respect administrative boundaries. The Abode of Chaos itself never separated art, technology, law, history, or matter. AI now confirms that these separations were often artificial.

Cognitive warfare is data warfare.

But also of storytelling.

Of perception.

From psychology.

It permeates the social body.

A state can be militarily intact yet cognitively disorganized. A population saturated with conflicting narratives can lose faith in all sources. This is perhaps one of the most effective forms of attack:

Not to make people believe a specific lie, but to make them believe that nothing can be known.

This strategy of eroding trust is formidable. If every image can be fake, every voice cloned, every document synthetic, the reaction may be to no longer trust anyone. This is an epistemic crisis. Generative AI significantly increases this risk.

But it can also provide verification tools. We are therefore entering a race between generation and authentication. Production becomes easy. Proving origin becomes valuable.

One can imagine that digital provenance, cryptographic signatures, and chains of authenticity will take on enormous importance. This immediately brings me back to the Art Market. Art history has always been obsessed with provenance: who owned the work? Where was it exhibited? What documentation accompanies it? In a synthetic world, all information may have to acquire its own provenance.

What camera produced this image?

When ?

Has it been modified?

Who signed it?

Verification will become an infrastructure. Artprice's experience with historical data, signatures, catalogs, and provenance suddenly seems extremely relevant. We have always known that information without provenance context loses some of its value. Generative AI generalizes this problem to the entire digital world.

Every document can become suspect. Trust will therefore have to be rebuilt technically. And this will have direct military consequences. A head of state receives a video showing a major event.

How does he know it is authentic? The verification system must be faster than the decision-making cycle. Otherwise, forgery wins. This creates a new strategic infrastructure:

Real-time proof. We have never had a greater need to authenticate information quickly. It is almost a paradox: the more content we can produce, the more scarcity shifts toward proof. The cognitive future may well be a trust economy.

Whoever can prove origin has an advantage. This battle will extend far beyond nation-states. The media, universities, businesses, and citizens will all be involved. A democracy cannot function sustainably without a minimum of shared reality.

Political disagreement is normal. Disagreement about the very existence of facts becomes far more dangerous. AI could therefore threaten democracy through cognitive overload rather than direct control.

Too much content.

Too fast.

Too personalized.

Human attention span is becoming a scarce resource. That is precisely why I use the term cognitive warfare. The brain is the battleground. Not just the soldier's.

The citizen's perspective is also affected. Adversaries will seek to influence elections, mobilizations, markets, and trust in institutions. The line between war and peace is blurring. An influence campaign can be waged continuously without a declaration of war.

It is a gray area. States must learn to respond without turning every disagreement into a hostile operation. Here again, classification is essential. Not all foreign criticism constitutes cognitive warfare.

Not all false information is a coordinated operation. If the state sees the enemy everywhere, it risks becoming paranoid itself. The cure could be worse than the threat. Therefore, standards of evidence are necessary.

Attribution.

Intention.

Coordination.

Before discussing cognitive attacks, a sufficiently robust chain of events must be established. This is precisely the discipline I apply to information. Correlation and causation should not be confused. Thousands of accounts can spontaneously spread the same narrative. It is crucial to distinguish between organic and coordinated actions.

AI can help. It can also overinterpret. Yet another blind spot. A democracy must resist manipulation without sacrificing its pluralism.

It is an extremely delicate balance. The temptation to censor widely in the name of security can destroy precisely what we wanted to protect. Cognitive resilience, therefore, is not about removing all dangerous information. It is about making society more capable of contextualizing it.

Education.

Transparency.

Solid media.

Verification tools.

It is less spectacular than a large-scale surveillance system. But probably more robust. A population trained to recognize certain manipulations becomes harder to destabilize. It is exactly like cybersecurity. You cannot prevent all malicious emails.

Users are also taught to recognize phishing. Cognitive defenses must function in the same way. Humans remain part of the architecture. But we must accept that our brains have structural vulnerabilities. We like simple stories, information that confirms our beliefs, and emotional content.

AI systems can exploit this with exceptional precision. Personalization makes propaganda individual. Propaganda used to be a message for millions of people. Tomorrow, it can be different for every individual.

This is a massive shift. Persuasion is becoming algorithmic. A system can test thousands of formulations, learn what triggers a reaction in a person, and adapt the message. The line between this and marketing is already barely visible.

The difference lies in the purpose. This is why personal data is becoming a matter of cognitive security. The more an actor knows about my fears, my habits, my relationships, the more effective their message can be. Data protection is therefore not just an individual right.

It is becoming a component of national resilience. This is a consequence that we may not have sufficiently anticipated.

The Headquarter had outsourced the memory.

The network had outsourced its presence.

The drone had externalized the point of view.

AI will begin to outsource some of the processing. A data leak could enable extremely precise political targeting.

Once again, the boundaries between sectors are disappearing. The entire world is becoming cognitive infrastructure. This explains why I ultimately prefer to speak of a cross-cutting cognitive force rather than a “fifth force.” Land, sea, air, space, and cyberspace designate environments or domains of operation; AI permeates each of them.

AI is not an environment in the same sense. It is a force that permeates all environments and acts on the perception that coordinates them. It becomes the nervous system of the armed forces. And a nervous system is both power and vulnerability.

Severing a nerve can disable an intact muscle. Similarly, disrupting the cognitive layer of an army can render otherwise operational equipment useless. This is why cyberattacks, electronic warfare, disinformation, and AI are converging. They all seek, at different levels, to disrupt the flow between perception and action.

The conflict of the future can therefore, in certain situations, be won before the first shot is fired if the opponent is no longer aware of what is happening. The formula may seem dramatic. The logic it describes is ancient. Sun Tzu would probably have understood it immediately: the best fight is the one where the opponent is already acting according to our understanding.

AI makes this logic industrial and permanent. It could also transform deterrence. The stability of nuclear deterrence has notably relied on the credibility of a second-strike capability: each side had to believe that the other would retain the ability to retaliate after a first attack.

Uncertainty existed, but the architecture remained relatively understandable. AI can introduce new uncertainties. If a system greatly improves the detection of submarines or missiles, it can alter the balance. If a state believes it can neutralize adversary capabilities more quickly, stability may decrease.

A defensive technology can then have an offensive effect. This is still the logic of second-order consequences. The question is not only what the system does directly, but also what it changes in the adversary's beliefs.

Deterrence is a form of armed psychology. It depends on what each individual believes is possible. AI acts precisely on these beliefs. It could make some actors overconfident in their predictive abilities.

This could be a huge risk:

The illusion of being able to know the adversary. A model trained on years of data can produce behavioral probabilities. But a leader remains capable of an unexpected decision.

A revolution, an emotion, an accident can change the course of events. War always contains an element of contingency. A highly effective AI can create the feeling that this contingency has disappeared. That is dangerous.

We could mistake better prediction for certainty. History often punishes this arrogance. Many wars began because each side believed it could quickly predict victory. AI could reinforce this illusion if it produces overly convincing simulations.

A simulation is never reality. It is based on assumptions. Decision-makers must therefore be trained to treat models as tools, not as oracles. This is a fundamental philosophical distinction.

The oracle possesses the authority of mystery. The model must possess the limited authority of its data. This entire book is essentially a struggle against the oracle. Even the most advanced artificial intelligence must never become a new algorithmic clergy.

It must remain open to question. That is precisely where I find my phrase again: The work should question rather than impose. AI, too, should, as much as possible, question before imposing.

Intelligence that cannot be questioned becomes opaque power. In the military sphere, this opacity is intolerable. The decision-maker must be able to ask:

For what?

But we must be careful with this requirement of explainability. Some complex systems cannot always provide simple causality. They may offer factors, probabilities, or examples. Therefore, we must construct an explainability tailored to the decision.

The goal is not to transform every neural network into a universally understandable formula. The goal is to provide enough information to allow for responsible judgment. Again, proportionality. An immediate tactical recommendation does not require a treatise.

It needs essential elements:

Trust level, sources, alternatives, risks. The interface must condense without obscuring. It is almost an art of synthesis.

And synthesis is precisely what AI can do extremely well. We can use its power to mitigate its own risks. It can produce recommendations and simultaneously critique those recommendations. It can flag conflicts between sensors.

It can detect when a situation falls outside its training domain. This last ability will likely be crucial: Recognizing the unknown. A system is particularly dangerous when it encounters a radically new situation and continues to answer with the same confidence.

The military world is precisely a world where the adversary seeks to create new situations. The model must therefore possess a kind of epistemic alarm:

This does not look quite like what I know. It is a form of computational humility.

It should become a strategic virtue. We encounter this paradoxical idea again:

The best intelligence is not that which always responds. It is also that which knows when not to respond.

An experienced soldier sometimes knows something is unusual without being able to immediately explain it. This intuition for anomalies is invaluable. AI can become extremely effective at detecting anomalies. But it must then avoid automatically attributing hostile intent to them.

An anomaly is not an attack. Here's another point of contention. The algorithm detects a difference. The decision-maker adds a story.

It is in this space that cognitive warfare can take hold. The adversary can generate anomalies to overwhelm the system. Too many alerts can exhaust attention. It is an attack by fatigue.

Automated systems must therefore manage not only detection, but priority. And priority itself can be manipulated. The conflict enters a recursive logic. Every defense becomes a new attack surface.

We are entering an almost organic level of complexity. This makes the idea of total control illusory. No state will perfectly master all systems. We must therefore design for degradation.

What happens when AI stops working?

Can the army revert to a simpler mode of operation?

Do the operators retain their manual skills?

This question is fundamental. A technology that greatly increases efficiency can also create dependency. If humans lose the ability to operate without it, a failure becomes catastrophic. We have already seen this phenomenon in navigation. Many people no longer know how to read a map as they once did.

In civilian life, it is rarely a disaster. In cyber warfare, losing GPS can be. AI raises the same issue on a cognitive level. We will need to preserve degraded capabilities.

This is a classic military concept:

to function in a degraded mode. Cognitive sovereignty therefore requires an ability to return to a human level. Not because humans are always better.

Because resilience requires an alternative. A system entirely dependent on a single layer is fragile. That is exactly what I have always thought about Server Group infrastructures:

No critical function should rely on a single channel.

Redundancy may seem costly until the day it saves the entire system. Future armies will likely need to maintain slower, more analog procedures as insurance against digital failure. It is almost a historical irony: the more technologically advanced the world becomes, the more certain older technologies can become strategic because they are difficult to disrupt.

A cable.

A piece of paper.

A simple radio.

They can sometimes survive where the sophisticated system is disrupted. One should never despise the old. It can become a redundancy of the future. This is precisely the alchemical spirit:

No matter is ever truly dead.

It can regain functionality in a new configuration. Cognitive warfare is therefore also a war of resilience. Those who continue to understand even when their systems are degraded have an advantage. This applies to an army as well as to a society.

Cognitive resilience thus establishes itself as a component of national defense. This idea should probably enter security doctrines at the same level as cybersecurity. But its excessive militarization must be avoided.

An open society must accept disorder, criticism, and error. Resilience is not about producing unanimity. It is about preventing disagreement from destroying the very possibility of a shared reality. It is this difference that must be preserved. Apparent chaos can be a form of vitality.

This is precisely what the Abode of Chaos has always maintained. Visible disorder is not necessarily an absence of order. A democracy often resembles chaos because voices clash. But its order lies in the procedures that allow these voices to coexist.

A dictatorship may appear much more visually orderly, but its cognitive architecture can be fragile if no one dares to transmit bad news. Again:

Appearance and structure should not be confused. This lesson goes beyond art.

It becomes strategic. A highly disciplined army can be blind if information is poorly relayed. A more vocal organization can be more intelligent if it allows for dissent. AI amplifies this institutional choice.

It can be used to filter out dissenting voices or to amplify them. Therefore, it is less the technology than the doctrine that will determine the outcome. This is why I reject any deterministic view of artificial intelligence. It does not automatically condemn freedom.

It does not automatically guarantee progress. It amplifies existing human architectures. It is a multiplier. And the multiplier applies its power to what is given to it.

If the system is robust, it can increase its robustness. If it is paranoid, it can industrialize its paranoia. This statement seems central to understanding cognitive warfare. States must therefore work on themselves as much as on their models.

The best AI will not compensate for a flawed decision-making culture. This may be the real power struggle:

not just possessing the greatest model, but building the institution capable of using it without becoming dependent, blind, or hasty. This institution must know how to alternate between speed and slowness, automation and judgment, secrecy and transparency.

It is a complex architecture. It requires a level of maturity that technology alone cannot provide. And behind this military issue, another transformation is already emerging. If systems gradually become capable not only of analyzing and recommending, but also of operating machines in the real world, the line between artificial intelligence and robotics will become blurred.

The model will no longer exist only in the data center. It will have a body.

Drone.

Vehicle.

Ground robot.

Maritime system.

A software error can then become a physical movement. The problem will change in nature when these machines are no longer isolated but capable of coordinating their behavior.

What happens to command when the number of machines exceeds what a human can supervise individually?