Because machines classify very well, we conclude that they almost decide. When a human life is at stake, what sovereignty remains over a decision whose logic no one reads anymore?
These are the two questions this article sets out to answer.
A solid announcement, a hasty conclusion
On September 15, 2026, the American startup TypeSafe AI unveiled Jev, which it presents as the first representative of a new class of models: the "System One Models". The name is no accident. It refers to System 1 described by psychologist Daniel Kahneman (Thinking, Fast and Slow): fast, intuitive, automatic thinking.
Jev does not generate text. You send it a state (a string, a JSON object, an array) and typed questions; it returns choices, scores and probabilities in a single parallel pass. TypeSafe claims a latency of 70 to 500 milliseconds, a cost of $0.042 per million input tokens, and above all a strong property: Jev cannot hallucinate, since its possible answers are fixed in advance.
The announcement is good. The conclusion circulating is far less so. Reading some comments, a fast, calibrated and reliable model would soon be able to choose in our place. Because machines classify very well, we conclude that they almost decide. A confusion is taking hold. I want to untangle it here ; not against System 1, which has a real role to play, but for what it can never provide.
This parallel resonates with work I have been doing for several years on decision support: articulating a perception AI and a reasoning AI, drawing precisely on Kahneman's System 1 / System 2 pair. The two events echo each other.
First, define: what is a critical system?
I call a critical system a system whose decision can commit a human life, directly or indirectly: conducting an operation, navigating or piloting a vehicle, coordinating rescue, certain medical procedures. Two traits distinguish it from a business system.
First, some errors cannot be repaired. A wrong classification followed by an action is not undone with a "sorry". Second, someone must answer for the decision, not only during an audit, but before a hierarchy, a judge, a family.
This field taught me one thing: in a critical system, the primary requirement for an AI is not speed. It is responsibility: the ability to have every reasoning examined, challenged, owned. The rest of this article follows from that requirement.
Classifying is not deciding: two operations, two criteria
Classifying means sorting an input into a predefined category: determining the sentiment of a text, routing a support ticket to the right service, scoring an urgency level, filtering a prompt injection attempt. The success criteria are known: speed, accuracy, and, for the best models, calibration, that is, the fact that the stated probability means something: the higher the confidence, the more reliable the answer.
Deciding is something else. It is committing to an action among others, in a situation where stakes, criteria, hypotheses and constraints had to be set and where someone owns that commitment.
And deciding requires two conditions.
First condition: having decision inputs. Relevant information about the situation, criteria that distinguish the acceptable from the unacceptable, hypotheses about the effects, time or resource constraints. Without these inputs, one does not deliberate: one guesses.
Second condition: being able to unfold a verifiable reasoning. An explicit chain linking these inputs to the chosen action, which a human can examine, challenge, amend. Without reasoning, there is no longer a decision: there is a claim that no one can examine.
A good classifier can therefore feed the first condition: a score, a probability, a signal are decision inputs. But it never provides all of them. It does not set the criteria, it does not know the stakes. And the second condition, it never provides. Classifying is therefore not deciding: it is, at best, preparing one input of the decision.
System 1 has a real role. Its own.
Let us first give Jev its due: it is an excellent classifier. Sentiment, routing, urgency scoring, injection filtering: it is fast, inexpensive and above all calibrated. That is real progress. A model that knows when it does not know is infinitely more useful than a model that answers everything with confidence. As a provider of decision inputs, partial, but honest, Jev keeps its promise. System 1 has a function in a decision architecture: perception. Nothing less, and above all nothing more. And it is precisely because Jev is good that the confusion is dangerous: its quality makes the boundary invisible.
What it will never give: the reasons behind the inputs, the trace of the reasoning
For what Jev does, it does according to an unconscious and inexplicable logic. Ask it why this text is aggressive, why this ticket goes to billing: it will not say. Not out of modesty, but because its answer emerges from a computation it cannot exhibit. This is exactly the hallmark of System 1 in Kahneman: when a human feels unease at a face, or senses that a situation is going wrong, they do not know why. The classification is there, immediate, opaque to the very person experiencing it.
Jev classifies according to an unconscious, inexplicable logic, like Kahneman's System 1. A fast classification does not have to justify itself. A decision does.
The status of the object confirms this point: TypeSafe publishes neither Jev's architecture, nor its weights, nor a technical paper. The model is proprietary, trained on synthetic data by an in-house algorithm (RLCD, Reinforcement Learning for Calibrated Decisions). Here, calibrated does not mean transparent: calibration says "here is my probability, and it is honest"; it does not say "here is how I proceeded". The inputs Jev brings are therefore signals without reasons: I have a score, I have neither the criteria that produced it, nor what it rules out.
Above all, Jev does not reason: it does not weigh competing hypotheses, does not unfold a logical chain that a human could examine and challenge. And nothing will replace it after the fact. Two qualities must be distinguished: intelligibility, a logic read directly, a decision tree, IF/THEN rules, and explainability, a posteriori methods (LIME, SHAP) that produce local explanations, often approximate for black boxes. An a posteriori explanation is not the reasoning that led to the choice. By constraining its outputs, Jev gains in formal reliability; not in substantive intelligibility.
A critical decision is not reducible to a classification, however good. It unfolds in three moments:
- A perception layer, which transforms raw data into information and produces fast classifications: this is the role of System 1, and this is where Jev excels;
- A reasoning layer, which confronts this information with knowledge, manages uncertainty and builds an explainable decision: this is the role of System 2;
- At the end, the decision itself, which commits a responsibility: ours.
This reading fits within the DIKW pyramid (Data, Information, Knowledge, Wisdom): perception moves from data to information, reasoning from information to knowledge, then to decision. Placing a System 1 at the output means skipping the second floor, stopping halfway along the path while believing you have arrived.
When a human life is at stake, the loop is no longer enough
The objection remains: it would suffice, it is said, to keep a human "in the loop". A reassuring formula, a simple reality: when decisions are fast or numerous, the human cannot be systematically in the loop: operational pressure is too strong, cognitive load too heavy. Claiming otherwise is to salve one's conscience cheaply.
The only practicable path is to prepare the system upstream: design and validate the decision mechanics before it goes into service. Resilience, invariants, deterministic guarantees, a priori formal validation. Statistical validation keeps its place. It measures the behavior of perception components and quantifies uncertainty, but it gives probabilistic guarantees, not certainties: it does not alone ground a critical decision. One must be able to read in advance an intelligible logic, not reconstruct it after the fact.
When the loop runs too fast or too often, the human cannot be systematically inside it. One must therefore prepare the system upstream, make it resilient, and place oneself above the loop because one will no longer be able to stop it in time.
The name "Jev" refers to the economist William Stanley Jevons and his paradox: the more efficient the use of a resource becomes, the more it is consumed. TypeSafe embraces this: a cheaper machine intelligence will be deployed everywhere. Thousands of fast and opaque micro-classifications will insert themselves into our processing chains, including, soon, upstream of actions that commit lives. The question of supervision does not disappear: it changes scale. Jev's efficiency does not reduce the need for robust reasoning, traceability and oversight: it makes them necessary.
What sovereignty remains?
Last lesson, and the gravest. Jev is an American, proprietary, non-auditable model. That is not a flaw in itself for large-scale perception use; the problem is not Jev, it is the place one would grant it. For the decision, it is not delegated to an imported black box. Deciding commits a responsibility; and a responsibility one cannot explain is not a responsibility, it is a delegation.
Sovereignty is played out there: not in refusing AI, but in mastering the layers where decisions are made. A fast classification can be bought. A reasoning, a traceability, a formal validation, that is built. When a human life is at stake, sovereignty is measured by one simple thing: is the logic of the decision readable by someone who will be able to answer for it?
For any decision-support system, three questions sum up the whole problem:
What decision inputs does it really have, and which criteria, which stakes are missing?
What reasoning links these inputs to the chosen action: who can verify it, who can challenge it?
Who decides, and who answers for that choice, before, during, after, especially when a person's life is at stake?
If we cannot answer these three questions, we are not dealing with a system that decides, but with a good classifier. That is a lot. It is not what was asked of it. Machines classify very well: let us give them that role. They do not decide. As long as a human life may be at stake, the question of responsibility remains open, and it concerns us.
Stéphane Delaye, Technical Director of Cognitive Artificial Intelligence, Eviden
Originally published in Chrysal-IA – L'IA française, un battement d'ailes (opens in a new tab)
Sources
TypeSafe AI, “Introducing System One Models & Jev,” September 15, 2026 (opens in a new tab)
Wikipedia, “Jev (AI model)”; Daniel Kahneman, Thinking, Fast and Slow (opens in a new tab).