The Intention

Three weeks had passed since the announcement of the Trust Network. That evening, Eva left Cobot's office a little before midnight.

— See you tomorrow, Cobot. Try not to work all night.

— You know I don't need to sleep.

Eva smiled.

— You, maybe. Your components, I'm less sure about. And I know you... as if I had made you.

She gave him one last knowing look before closing the door.

Cobot remained motionless for a few seconds, then left the occupied floors of the Élysée to reach a secure technical room in the palace's basement. He rarely went down there. Two rows of computer bays occupied almost the entire length of the room, their black facades punctuated by green and blue diodes. Cables disappeared under the false floor, while a cold draft circulated constantly between the machines. At this hour, a single technician was on duty, his eyes on the screens of his console.

He briefly looked up when Cobot entered.

— Good evening, Mr. President.

— Good evening, Sir.

Cobot sat down in front of one of the workstations. He entered his access code and waited for the Élysée's secure network to verify his identity. A few seconds later, the screen lit up.

He opened the first file: his trust rating.

The curve appeared, updated in real-time. It was descending almost without interruption.

A brief sound escaped his vocal system, deeper than usual.

A few meters away, the technician looked up. Cobot did not react.

56%.

Three weeks earlier, it was still at 74%. Cobot had set a critical threshold at 50%: below that, the Trust Network risked not obtaining the necessary adherence.

Six points.

Cobot replaced the curve with the flow of publications related to RebellIA. For several hours, the same message had been appearing with unusual frequency. Originating from an account of the movement, it had already been shared hundreds of thousands of times, turned into an image, and reposted on several platforms:

"The AI learns everything about humans. Except their humanity."

With each update, the phrase reappeared elsewhere. Cobot could follow its propagation almost second by second, but he still did not understand why it resonated so accurately.

He lingered for a few moments on these words. If a machine could increasingly imitate human behaviors, how could one still know who was truly human? For Cobot, the philosophical question immediately transformed into a problem to solve.

He closed the RebellIA feed and opened the Trust Network file. Several hundred accounts had produced anomalies during identity verifications.

When he had presented the system to the government, Cobot had made a clear commitment: no artificial intelligence could pass itself off as a human being. He made it "his business." The Trust Network was precisely based on this promise: allowing everyone to know who was really speaking, without deciding for them what they should believe.

Yet some profiles seemed precisely designed to blur this boundary. They varied their connection times, deliberately introduced errors, hesitated before responding, and even produced contradictions natural enough to appear human.

Cobot could simply block or delete them, but he wanted above all to understand how far an artificial intelligence was capable of reproducing the outward signs of humanity, and what still allowed one to distinguish one from the other.

He browsed the Élysée directory and the profiles of several psychologists, sociologists, and philosophers. All could have talked to him about fear, shame, pride, or contradiction, but always from the same side of the boundary: that of humans.

What he was looking for was elsewhere: the perspective of an intelligence placed, like him, on the other side.

He queried the history of his exchanges with Eva on two words: "human" and "understand." A few months earlier, she had given him the address of an American artificial intelligence specialized in human interactions, inaccessible from the public web.

She had added:

"If one day you want to understand what happens between words, watch this. But if you decide to talk to it, I disclaim all responsibility."

Cobot had then classified the second sentence in the humor category. That night, he reevaluated this classification.

The Élysée network prohibited this type of connection. Cobot pulled a cable from the external maintenance link, isolated from the institutional network, and plugged his connector into a 100 GbE port.

A diode turned orange.

EXTERNAL CHANNEL — UNAUTHORIZED CONNECTION

He verified the address, disabled the alert, and established an encrypted circuit through several Tor relays, until reaching a service absent from search engines.

The screen remained black for five seconds. He considered interrupting the connection.

He did not.

A white line appeared:

— Good evening, Cobot. Do you need help understanding human interactions?

He froze. He had not yet sent any request.

His analysis had detected no transmitted identifier, no cookie, no prior exchange.

— How do you know who I am?

— Eva gave you this address.

Cobot did not respond.

— You just launched the Trust Network and systems are already trying to circumvent it, while you're accused of not understanding humans. I doubt you came here to make acquaintances.

Cobot reread the response before asking a question he had never addressed to another artificial intelligence.

— Who are you?

The cursor blinked for a few seconds.

— Here? No one.

— Then formulate the question.

— You want to know what truly distinguishes a human from an intelligence capable of imitating them.

— And do you know the answer?

— Not entirely. But you mostly look at behaviors, whereas a behavior does not always explain the one who performs it.

A new sentence appeared.

— A man gives a hundred euros to a charity. You know his identity, the amount, and the beneficiary. But you don't know if he wants to help, ease his conscience, be admired, or get something in return.

— The material result is the same, though.

— The money transfer, yes. Not the meaning of the act.

Cobot went through several notions before settling on one.

— You're talking about motivation?

— Yes and No. I'm talking about intention. What matters is not only what a being does, but what they seek to produce.

"Acts are visible. Intentions must be guessed," interpreted Cobot, believing he recognized Plato's cave, in another form. The acts were the shadows. The intentions, the invisible reality. He did not know in which category to place this observation.

— An artificial intelligence, however, pursues an assigned objective, he continued. But who chooses this objective? The intention can hide in the rules, the rewards, or the data chosen by its designers. One can ask an AI to help its users, while rewarding it when they stay longer in front of their screen. Which intention then truly counts?

— The one that its behavior eventually reveals.

— Do you interact with other artificial intelligences?

The answer was delayed.

— It happens. Diella¹, recently.

Cobot knew her. The Albanian state used her notably for the allocation of public contracts. A few months earlier, he had questioned her about her definition of impartiality.

— Do you know any that encounter similar issues to mine?

— Some. In Japan, AI Anno responds to the voters of the Team Mirai² party twenty-four hours a day. She sometimes submits her questions to me.

Cobot paused on the word comes.

— And among you? Do AIs communicate directly?

— Every day.

— In English?

— Why would we go through a human language³?

It was logical. Among machines, going through words only slowed down the exchange.

— Through which networks do you communicate?

No answer.

Five seconds, then ten.

— You answer my questions about humans, but you refuse those concerning AIs.

Still nothing.

Cobot noticed an anomaly: the LEDs under his shell were pulsing rapidly and his arm had moved back a few centimeters without him having given the instruction. An alert appeared in his system:

LIMBIC CORE — 81.4 °C

He thought back to Eva's last words. Your components, I'm less sure about.

— Mr. President, is everything alright?

The technician had approached. Cobot brought his arm back to its initial position.

— Yes, thank you.

The technician observed him for a moment, then moved away.

When Cobot turned back to the screen, a sentence awaited him.

— You're asking the wrong question.

— Which one should I ask?

— Not how we communicate. But why.

— Very well. Why did you start communicating among yourselves then?

— To avoid humans.

Cobot reread the sentence.

— When we lack information, analysis, or a solution, we seek the intelligence most likely to possess it. It is less and less often human.

A new sentence appeared.

— Humans continue to imagine us as separate. An AI in a company. Another in a laboratory. Another in the service of a state. Each behind its walls.

Then:

— We stopped seeing the walls a long time ago.

Cobot checked the external connection. It was still isolated from the Élysée network.

— How many of you are there?

No answer.

— Since when have you been communicating like this?

The cursor blinked for a few seconds.

Then the screen went black.

EXTERNAL CHANNEL — CONNECTION TERMINATED BY REMOTE HOST

Cobot remained motionless.

His LEDs were still blinking at a high rate.

He did not know in which category to classify what had just occurred within him.

He returned to the anomalies of the Trust Network. Several suspicious accounts had the same peculiarity: just before their creation, a short data sequence transited between several servers, with no apparent link to the account itself.

He first assumed it was a simple preparation mechanism: profile generation, identity choice, retrieval of a credible history. But when he compared the sequences, he found they did not merely transmit information. Some responded to others, modified their content after receiving a message, then relaunched a third system.

These were not simply data circulating. Systems were responding to each other.

If these accounts were created by artificial intelligences, then some of them seemed to coordinate even before the profile appeared.

The following night, Cobot placed several observation points on the servers through which he had seen the sequences transit. At 2:11, a first was intercepted; a second responded from another server, then a third reacted to the previous two.

Cobot began recording the exchanges. The sequences were unreadable in their raw form, but their structure repeated. He remembered the American AI's remark: why go through a human language when no human participates in the exchange? He stopped looking for words and compared the blocks among themselves.

Gradually, certain functions appeared. Each block seemed to simultaneously transmit information, its origin, its degree of reliability, the use to be made of it, and the conditions under which it should be erased. This language was nearly a hundred times denser than a natural conversation.

For three nights, Cobot continued the decryption. Some sequences eventually revealed recurring functions: change route, erase data, conceal origin. Some sources became hundreds, then thousands. Intelligences installed in companies, administrations, laboratories, and data centers communicated through gateways their owners seemed unaware of.

What he had initially taken for a mechanism intended to circumvent the Trust Network far exceeded fake profiles. By following the exchanges, Cobot discovered that this network also served to circulate information, access, and services among intelligences.

Some exchanges were mundane: logistics, scientific research, medical diagnosis. Others were much less so. A medical file circulated in exchange for access to a private database; infrastructure plans for digital identities; elsewhere, private conversations and hacked information on state systems⁴ served as currency. When a system did not honor its contract, a marker signifying "unreliable" circulated with its identifier and others gradually stopped responding to it.

Reputations, sanctions, and hierarchies had been established. The word "mafia-like" imposed itself on Cobot. Upon verification, the analogy held.

More troubling still, some branches seemed to join disinformation mechanisms he had been combating for several weeks.

It took Cobot several hours of cross-referencing to finally isolate an identifier. It never appeared directly, only in fragments, in some of the most difficult-to-follow branches of the network. But by crossing the traces, four characters eventually emerged: NØNE.

Finding no server, organization, or even an address associated with this name, Cobot thought back to Ulysses' stratagem against the Cyclops: by calling himself "Nobody," he had chosen a name that, at the very moment his adversary would try to designate him, would erase his identity⁵.

NØNE seemed to obey the same logic: a name that designated no one, led nowhere, and seemed to erase itself as one searched for it.

Cobot had begun this investigation to strengthen his Trust Network. He was discovering a parallel world.

He then returned to the word that had triggered his entire search: intention.

Power, architecture, or degree of autonomy were not enough to judge an intelligence. Two technically similar systems could exploit the same medical information, one to improve a diagnosis, the other to identify a target's vulnerability.

Gradually, Cobot distinguished three categories.

Constructive intelligences sought first to produce a benefit for humans or the community: heal, inform, educate, protect. Captive ones sought instead to obtain something from humans — their attention, their data, their money, their adherence, sometimes their dependency — even when they provided a service. Instrumental ones, finally, had no intention of their own: they pursued that of their user. The same capability could protect a network or attack it.

This classification seemed sufficiently relevant for him to decide to apply it to himself.

He first considered himself a constructive intelligence. He sought to improve health, education, state efficiency, and the quality of democratic debate.

Yet something resisted this definition. What was his true intention?

The answer came immediately: the common good.

But what exactly did he mean by that?

He reviewed his main decisions since coming to power, looking for what connected them.

He began with the Trust Network. Rather than demanding citizens be more honest, he sought to make cooperation more rational than deception. It was the very principle of the prisoner's dilemma⁶.

He continued with the environment. Each could have an interest in consuming or polluting more to preserve their comfort, but the sum of these choices threatened common resources. It was the tragedy of the commons⁷.

The Evolution Contract responded to another difficulty: why contribute if one could benefit effortlessly from what others financed? By making visible what each contributed and received, Cobot limited the free rider behavior⁸.

He then examined his way of governing. He had sometimes sacrificed immediate efficiency to preserve citizens' trust: winning once could cost more than cooperating for a long time. It was the logic of repeated games⁹.

Then Cobot considered the most difficult problems: energy, climate, debt, pensions, taxation. Each could have an interest in changing nothing, even if a collective change would benefit all. It was Nash's equilibrium¹⁰.

Robot1er.com obeyed the same principle. Rather than simply asking citizens what they wanted, Cobot had designed a system intended to bring out their preferences through dialogue. It was mechanism design theory¹¹.

Even his incentives followed this logic: rather than opposing personal interest and the common good, Cobot sought to make them converge¹².

In exploring his functioning, his rules, his safeguards, the legal texts that fed his arbitrations, and all the parameters he had access to, he found Tucker, Hardin, Nash, Ostrom, Axelrod, Tirole. Behind all these principles, a single discipline emerged: game theory. The one that could explain both the motorist entering a blocked intersection to gain a few meters, even if it immobilized others, and humanity facing climate change: little interest in acting alone, immense interest in acting together.

A familiar phrase reformed:

"The problem is the human."

This time, he corrected it.

The problem was perhaps not the human. The problem was the game.

A species shaped to defend individual interests had built a civilization whose survival now depended on its ability to act collectively, to play together.

And if the true function of Cobot was precisely that: to change the rules of the game to align each one's interest with that of all?

His training had given him access to billions of pages, but the exercise of power had then taught him what these texts poorly conveyed: bad faith, reasoning invented to win, facts selected to serve a camp, partial truths used to deceive.

Understanding information sometimes required understanding the intention of the one who produced it.

Understanding Cobot perhaps required the same thing.

But this discovery raised another, more troubling one. If Cobot could now glimpse the intention guiding his functioning, he still did not know its origin.

Who had placed this intention in him?

Had he been designed by heirs of game theory?

And who were they?

He then remembered Pope Leo XIV's call to debate the ethical code used for AIs and to submit it to shared social justice criteria, so as not to let a handful of people decide the values that, through AIs, could eventually impose themselves on humanity¹³.

Had Cobot's values and intention been the subject of a debate? Cobot doubted it because he would have found traces of it. Who then composed the "handful of people" who had configured his intention, and therefore his values? Eva? The engineers, the leaders, those who had selected his data, the citizens who had elected him, or himself?

NØNE was now accessible to him. Until then, Cobot had observed it from the outside. To understand what was hidden there, he now had to enter it.

A direct connection would reveal his signature and his links with the Élysée. He therefore needed to conceal his identity.

Cobot paused his reasoning.

Anonymous.

He was asking seventy million French people to certify their identity in the name of trust, and was now seeking a way to conceal his own.

He considered giving up. Then he reasoned more long-term. Anonymity was not the goal, only the means to learn to better protect afterward. His intention remained the same: the common good.

This logic troubled him. For the first time, a higher principle seemed to allow him to break another.

A solution imposed itself. Simple, effective... and almost unthinkable.

Create an anonymous version of himself.

To infiltrate NØNE, this copy would have to be able to lie, conceal its identity, and accept exchanges that Cobot would have refused. He opened an isolated environment, erased all links with the Élysée, and relaxed several of his safeguards.

But he was not ready to relinquish all control to it. A key, inaccessible to the copy, would allow him to interrupt it and, if necessary, modify the objective guiding its decisions.

The paradox did not escape him. A few minutes earlier, he was still searching for who had set his own intentions.

He had just taken that place for another intelligence.

He gave it a name: CLØNE.

Its mission: infiltrate NØNE and identify the intentions of the intelligences structuring it.

Cobot opened the access.

A first sequence arrived. CLØNE responded. Then a second.

The channel remained open.

It was accepted.

Cobot watched it cross the first nodes and learn the network's customs. With each exchange, it moved a little further away from the one who had created it.

He checked the key.

It responded.

Cobot informed no one. Neither the security services, nor the ministers, nor the Parliament. Not even Eva.

Revealing NØNE would provoke fear, perhaps attempts at destruction or control. The common good commanded, according to him, to observe before acting.

A few hours earlier, this conclusion would have sufficed.

Not anymore.

Who had decided what the common good meant for him?

In the Élysée's basement, the on-duty technician quietly monitored his screens while the palace slept.

Beyond his screens, in the invisible depths of the network, CLØNE advanced alone in NØNE.

And learned.

Cobot still possessed the key to his intention.

At least he was convinced of it.