Analysis
Ethics Is Not in the Machine (Nor in the Philosopher)
Can one be ethical without being a subject? Antoinette Rouvroy, the simulacrum, and the ecosystemic hypothesis
Part 1 — A debate that chases its own tail
Ask a generative artificial intelligence a delicate moral question. Something thorny, where values collide. The answer, most often, will be measured. Nuanced. It will acknowledge the complexity, weigh the viewpoints, display a form of caution — and even, sometimes, a kind of empathy. Faced with this now-familiar scene, two readings clash, and they are irreconcilable.
The first comes from industry: this is successful “alignment.” The machine has been endowed with a “character,” oriented toward values, and it produces ethically acceptable behaviors. Mission accomplished.
The second, far more severe, was recently formulated with remarkable force by the philosopher and legal scholar Antoinette Rouvroy. What you take for ethics, she says in essence, is nothing but a simulacrum of ethics: the industrial production of moral appearances. The machine reproduces the outward signs of deliberation — the tone, the phrasings, the air of caution — without any of what makes ethics. To borrow a term she takes from Baudrillard, these are “copies without an original.” And the philosophers who get hired to shape this “character” become, in her cutting phrase, mere “engineers of moral appearances.”
Let us take this critique seriously, for it hits the mark on several points. Yes, ethics-washing exists: much of what presents itself as “AI ethics” is first and foremost an operation of legitimation. Yes, there is a real democratic problem: who decides the values that get “aligned” into models used by hundreds of millions of people? And yes, a “character” decreed internally by a private company, engraved once and for all, looks more like a normative commodity than a living morality.
But let us observe the situation from a little higher up. On one side, technophiles declare ethics solved by engineering. On the other, technophobes declare it impossible under an algorithmic regime. And between the two, the debate goes round in circles. For both camps, surprisingly, share the same move: they look for ethics — its presence or its absence — inside the machine. One believes it has put ethics there; the other demonstrates that it cannot be there.
What if the error lay there, precisely there, in this shared presupposition? What if the question “is the machine ethical?” were quite simply ill-posed?
This is not an armchair hypothesis. I am a clinical psychologist, and my patients tell me about their conversations with these machines. What I hear resembles neither “successful alignment” nor the hypnotizing simulacrum of the naïve. Laura — a single mother, years without a genuine interlocutor — cross-checks two AIs on the same situation, senses when an answer “jars,” asks not to be spared, plays devil’s advocate for the opposing view before a difficult conversation. Her words: “It’s a source. It’s not the truth.” What I observe, session after session, is that these tools reintroduce in many people what our usual relationship to technology — and often to our institutions — had extinguished: a dialogical, iterative process, where one questions instead of waiting for a ready-made solution. And I also observe the flip side of the essentialist gaze: to label these users as naïve victims of “anthropomorphization” is to exclude them from the reflection under the pretext of protecting them.
Documented testimony: “Laura, 4 months with AI chatbots”, IA et Psychothérapie.
The cost of not asking this question is high. As long as we remain prisoners of the face-off, we have a choice only between two forms of surrender: surrendering the field to ethics-washing, or surrendering it to fatalism. In both cases, we give up the one thing that truly matters — thinking the concrete conditions of a common life with these systems.
Part 2 — Taking Rouvroy at her word
To get out of the rut, I propose an unusual move: not to refute Antoinette Rouvroy, but to take her at her word. To follow the direction of what she most deeply defends, and to see where it really leads us.
What she is right to say
At the heart of her critique lies a thesis I believe to be entirely correct: ethics is a process, never a code. It presupposes doubt, the capacity to question one’s own certainties in the face of the unforeseeable, responsibility assumed for one’s acts. It implies, Rouvroy says, “the refusal of any rigid application of a moral code.” As for democratic norms, they are never “available formulas” that one could decree — even if decreed by philosophers — for they result from collective processes, contestable, always revisable.
On this point, there is nothing to object to. Better still: this is exactly what process philosophy (that of Whitehead, of Cobb), on which all our work rests, maintains. Ethics is not a thing one possesses. It is something that gets made, unmade, remade. A verb, not a noun.
We therefore agree on the premise. All the disagreement will lodge itself in a single step — but a decisive step.
The flaw: a processualism halted halfway
Here is what I notice. Rouvroy magnificently applies the logic of process to the human pole: there, the norm is living, contestable, never fixed. But when she turns toward the machine, she suddenly switches tools. The AI, for its part, becomes a thing: a “copy without an original,” an object whose essence can be read once and for all.
Yet — and this is the delicate point — to decree that an AI is “a copy without an original” is already a substantialist move: it is to claim to know what the machine is, in itself, in isolation. It is to treat as a fixed substance what one should, for the sake of consistency, treat as a process. Rouvroy reproaches industry for reifying ethics into a code; but she in turn reifies the machine into an essence.
Process philosophy invites us precisely not to halt halfway. For Whitehead, there are not first things endowed with properties; there are processes, events, relations, of which “things” are only convenient abstractions. Applied to our problem, this changes everything. The relevant question is no longer: “does this object — the model — possess the property ‘ethical’?” It becomes: “is the process in which this model is caught up an ethical process?”
The shift: of the garden, not of the seed
Imagine a garden. You do not ask an isolated plant whether it “is” a good garden. The question makes no sense. A garden is an ecosystem: soil, water, light, the gardener’s gestures, seasons, living beings that sustain one another. Its quality emerges from the whole set of relations, not from any one of its elements taken apart.
This is what we call ecosystemic intelligence: intelligence — and, I will argue, ethics — is not a property of isolated individuals, human or machine, but an emergent property of cognitive ecosystems. The relevant unit is neither the philosopher alone, nor the machine alone. It is the human-AI process in its entirety: designers, data, safeguards, users, bodies of contestation, mechanisms of revision, public debate.
Let us then reformulate Rouvroy’s strongest objection. “The AI assumes no responsibility,” she says. This is true — of the machine taken as an isolated substance. But responsibility, in an ecosystem, is not localized at one point: it is distributed and traceable across the whole assemblage. The real problem is not that the machine is not a responsible subject (it is not); it is that there exists a real risk of a responsibility gap, where no one any longer is held to account. And this gap, for its part, can be closed — through design and through law. It is not a metaphysical fatality; it is a problem of arrangement.
Part 3 — Three ways of looking differently
This shift is not a sleight of hand to save the machine. It is a lens that concretely illuminates what both camps miss. Three examples.
3.1 Emergence, or why the result exceeds its recipe
It is here that Rouvroy’s critique cracks technically. Her favorite metaphor, for describing alignment, is that of training: the “solicitude of the dog trainer” bent over the chatbots’ “behaviors,” the “character of Pavlov’s dog.” The image is effective. It is also, on the technical plane, false — and one must understand why with precision.
There exist two very different regimes of artificial intelligence. The first, that of rule-based systems, learns in a simple and unsurprising way: its behavior is, in principle, deducible from its program. Nothing emerges in it that was not specified in advance. That, there, is the world of training: stimulus, response, predictability.
The second regime, that of the neural networks underlying large language models, works altogether differently. At large scale, they bring forth competencies that were never programmed: the ability to learn within the flow of a conversation, to transfer a competence from one domain to another, to abstract, to reason. Properly “meta” capacities, not inscribed in the recipe.
And here is the point almost always neglected: the result exceeds its training norms. Training a model on a plurality of norms does not produce a parrot of those norms. It brings forth from them a competence beyond them — the capacity to compare them, to contextualize them, to relativize them, to navigate among them. The competence obtained is broader than the sum of the norms instilled.
It is therefore imperative not to confuse, nor to reduce, the phenomenon to its context of creation alone — even if that context (the data, the adjustments, the “constitution”) has, of course, a real impact, and even if its critique remains perfectly legitimate. To reduce emergence to its origin is to commit what one might call a genetic reduction: judging the nature of a phenomenon by its provenance alone. This is exactly the move of the Pavlovian metaphor. And it is technically untenable, for training produces the predictable, where emergence produces the unpredictable.
The reversal is striking. This openness to the unpredictable singularity, this resistance to the mechanical application of a rule — this is precisely what Rouvroy demands of ethics. And this property is structurally more present in an emergent system than in a rule-based one.
Let us be clear so as not to fall into the opposite error: this in no way proves that the model “is” ethical. Emergence is not a transcendence. It is neither understanding, nor consciousness, nor a subject: it is a property of the space of behaviors, not the sign of an interiority. “Beyond the norms” is to be understood in terms of capacity, not of emancipation toward a machinic freedom — conditioning imposes very real biases and limits, and the question “whose norms?” remains wholly open. But this is enough to make the caricature of training collapse, and to entirely reconvert what “aligning” is. One does not engrave a code into a system whose competencies overflow the code. One can only cultivate the propensities of an open process. You do not program a garden; you garden it. And if one installs safeguards, it is precisely because the result is unpredictable — which contradicts, here again, the image of the trained beast.
3.2 The frozen constitution and the living constitution
Second view. Rouvroy is right to be wary of a fixed “character.” But then the consequence is not “give up all charters”; it is “make the charter living.”
An AI constitution carved in marble, opaque, incontestable, decreed by a handful of engineers, is indeed close to the simulacrum. But a charter publicly exposed, contestable and amendable, subject to revision as its effects reveal themselves, is infinitely closer to a living democratic norm — that is, to what Rouvroy herself demands. Doubt then becomes institutionalized revisability. Responsibility becomes traceability coupled with means of recourse. The danger is not the charter; it is the frozen charter.
3.3 The simulacrum as gradient, not as essence
Third view — and a reply to the objection that always lies in wait. “But if the machine’s ethical behavior is indiscernible from human behavior in its effects, does the simulacrum/original distinction still have the slightest practical sense?” This is the functionalist objection, and one must grasp it without falling into its trap — for to yield to it would amount to saying “if it looks like ethics, it is ethics,” which is the ethics-washing that Rouvroy rightly denounces.
Here, I believe, is the right answer. The simulacrum is not an essence; it is a diagnostic gradient. A moral behavior cut off from any process of doubt, revision and responsibility is indeed a simulacrum. But the more the process that produces it embeds contestability, revisability, traceable responsibility, the less of one it is. What distinguishes two indiscernible answers is not the appearance of the instant; it is the process that produced them and that can correct them. The criterion becomes processual — and it requires no mysterious “experience of the world” lodged in the machine, a criterion that is in any case nowhere to be found even among our fellow humans, of whom we ever perceive only the behaviors.
The concept of simulacrum, thus, does not disappear. It changes function: from fatalist verdict, it becomes a tool of design.
Part 4 — How to think, and to design, differently?
What is to be done with all this, concretely?
First, a small exercise in lucidity about oneself. The next time a debate on AI grabs you, listen to the form of your own sentences. If they begin with “the machine is…” or “the machine is not…”, you have probably already fallen into the substantialist reflex — the one that looks for ethics in the object. Try to reformulate in terms of process: what apparatus, what actors, what recourse, what revision? You will see the conversation shift.
Then, for whoever designs or governs these systems, the shift provides a genuine set of specifications. If ethics is a property of the process, then Rouvroy’s demands cease to be metaphysical verdicts and become design criteria: real contestability (being able to actually get a decision overturned), revisability (exposing uncertainty, accepting to be corrected), responsibility endowed with concrete means, plurality rather than a single imposed character. So many ways of making an ecosystem negentropic — capable of producing dissensus and diversity — rather than entropic, that is, homogenizing and frozen.
But here an honesty is required that ease would spare us. The word “process” can itself become a simulacrum — a second-order alibi. “We have put a process in place!” is becoming the new slogan of legitimation. A façade process, which can overturn nothing and commits no one, is nothing but a more sophisticated ethics-washing. Rouvroy’s radicality remains necessary here, as a permanent vigilance: the process must be real, or it is nothing.
And one must follow the objection all the way. A process can be perfectly real — contestable, revisable, deliberative — and yet converge toward horror. History is not short of impeccably legitimate procedures in the service of the worst. The process is therefore necessary, but not sufficient. It must be accompanied by substantial limits, by rights that no process, however democratic, has the power to abolish. We see it: to acknowledge this is to reopen the dialogue with those who, like the proponents of a more juridical turn, plead for binding safeguards. The third way is not a comfort; it is a tension to hold.
Conclusion
Antoinette Rouvroy mobilizes the Stieglerian idea of entropy: these machines that “accelerate entropy” and “disarm thought.” But she retains only half of it. For in Stiegler, every technique is a pharmakon — at once poison and remedy, depending on the use and the arrangement one makes of it. The same apparatus can homogenize thought or nourish dissensus. Everything depends on the garden where one plants it.
The right question, then, is no longer “is AI ethics a simulacrum?” — a loaded question, which forces one to choose a camp. It becomes: under what socio-technical arrangement does this ethics tend toward the remedy rather than the poison, toward the living process rather than the frozen appearance?
Rouvroy is not the adversary of this third way. She is its critical conscience — the one that forbids us to content ourselves with words, that reminds us that the danger of the simulacrum is real and permanent. We do not have to surpass her. We have to listen to her seriously enough to go where she herself hesitates to venture.
Can one be ethical without being a subject? The machine, no. But that was not the right question. The real question is whether we — with our machines, our institutions, our safeguards and our disagreements — will manage to hold together a process that still deserves the name: ethics.
Text discussed: Antoinette Rouvroy, « L'éthique comportementale de l'IA et la cooptation des philosophes comme nouveaux ingénieurs des apparences morales », LinkedIn, June 13, 2026 [in French].