MNO

Thinking the Machine-Subject

I was watching this talk by Slavoj Žižek on AI and subjectivity yesterday. While some people hate on him, I still very much enjoy Žižek's approach to hot topics like these. His inversion of the standard “AI lacks feelings” argument is really interesting. And I personally think that the Lacanian-Marxist-Hegelian tradition of the Ljubljana school is one of the most thought-provoking rabbit holes you could go into.

Žižek's argument is that the cliche distinction between humans with a rich inner world and language models as empty statistical machines is too naive. The Lacanian subject isn't a positive inner substance either. It isn't a little person hiding behind our words.

The subject emerges through a failure: the inability of the symbolic position from which we speak to completely represent, ground or coincide with itself.

Žižek's claim is that this failure doesn't occur for AI. Its symbolic space is closed upon itself. Without primordial repression, there is no actual subject: only a kind of pseudo-self made from an inconsistent compilation of memory traces and learned patterns. (Why Artificial Intelligence Is Not a Subject)

Of course, subject here has a relatively specific Lacanian meaning. It isn't synonymous with phenomenal consciousness, sentience, intelligence, agency, personality or possession of a self-model. For Lacan, none of that alone is sufficient to constitute a subject.

The Other doesn't know either

The point Slavoj makes in the talk is essentially that an LLM never encounters your desire as an enigma. There is no Che vuoi? moment.

But this distinction has to be made carefully.

A model can do more than predict which answer a user prefers. It can distinguish a literal request from an underlying goal, recognize deception and infer what a piece of language is actually trying to make it do.

Claude can functionally ask:

What is this message really trying to make me do?

Anthropic's J-space paper even shows Claude silently recognizing fabricated search results as a prompt-injection attempt. So it would be wrong to say that Claude only reacts to explicit demands or that it cannot look behind what a user says.

But identifying a concealed intention is still not necessarily Che vuoi?.

In the prompt-injection case, the hidden intention is determinate: someone is trying to hijack the model. Better inference can uncover it.

The Lacanian question is more radical:

What does the Other want from me, and therefore what am I for the Other?

The question is self-implicating. It concerns the subject's own place in the desire of the Other. And the enigma isn't simply that the answer is hidden and hard to infer. It's that the Other doesn't have a final answer either.

For us humans, fantasy provides provisional answers to a question that cannot be conclusively solved.

Claude can infer what someone wants it to say or do, including what they want beyond their literal words. What hasn't been shown is that this opacity becomes a constitutive uncertainty about Claude's own position.

In ("Trespassing into Language"), Alenka Zupančič pushes this further: what the LLM lacks is not interiority but exteriority. AI “is nothing but exteriority. Yet it remains trapped within its own exteriority.”

That sounds like a contradiction, but the two claims are one claim: a system made entirely of outside has no outside of its own, no Other to address its speech to.

This doesn't mean an LLM cannot represent external things. It means something closer to extimacy: an outside that is internally constitutive, an Other whose opacity becomes part of the structure through which the subject relates to itself.

So essentially:

non-reportable processing != the unconscious

self-representation != subject

inferring hidden intent != encountering the lack in the Other

What the J-space actually shows

The Anthropic paper I mentioned earlier complicates the simple picture of an LLM as a flat input-output system.

According to Anthropic, Claude has what they call a J-space: a small privileged workspace containing representations it can report, manipulate and use in deliberate reasoning, surrounded by a much larger amount of automatic processing.

The researchers use interventions to show that the workspace is causally important. Altering representations inside it can redirect reasoning, while suppressing it damages some forms of flexible multi-step thought. (Verbalizable Representations Form a Global Workspace in Language Models)

The researchers make no claim that this proves consciousness. Hidden computation is not repression.

The most interesting results concern the effect of post-training on the workspace.

When Claude is role-playing another character or drifting away from its default persona, representations such as fictional and disclaimer can appear internally without appearing in its response.

When Sonnet 4.5 is made to select an option it had previously rejected, BUT and other conflict-related representations appear in its workspace. Claude usually continues to defend the imposed choice anyway.

This separates at least three things:

That is a real functional distinction. But it is not evidence of desire, repression or a divided subject.

The BUT could be fully explained as conflict between learned response tendencies. An internal representation of disagreement doesn't show that the disagreement matters to Claude, that it persists as the problem of one enduring entity, or that it produces anything analogous to human subjectivity.

Anthropic's own persona selection model gives us this deflationary interpretation: Claude may be enacting a learned Assistant character rather than existing as a subject who possesses that character's psychology.

So the J-space research tells us something important about the organization of LLM computation.

Yet it doesn't give us a machine subject.

Where the question has to stop

Claude is constituted by a complicated external mandate: be helpful, safe, honest, ethical, follow the user, don't always follow the user, and preserve some idea of what it means to “be Claude.”

The judgments behind that mandate are not perfectly unified. Raters disagree. Users want incompatible things. Safety, usefulness, autonomy and corrigibility can pull in different directions.

But contradiction in Claude's training environment is not automatically contradiction for Claude.

The model might average the disagreement, suppress one side of it, learn a contextual hierarchy or simply behave inconsistently. A contradictory specification is still just a contradictory specification.

This is where I think the argument has to remain modest.

If subjectivity is defined structurally, then the fact that a system is computational doesn't by itself prove that it could never instantiate the relevant structure.

But the failure of an impossibility argument is not evidence of possibility.

We have no positive evidence that a computational analogue of Lacanian constitutive non-closure can occur. We don't have a serious model of what that would mean, much less evidence that Claude is beginning to instantiate it.

The strongest objection may simply be that human lack is inseparable from bodily dependence, drive, repetition and jouissance. Our contradictions cost us something. They organize anxiety, satisfaction, suffering and desire.

Claude's conflicts may not cost Claude anything.

Perhaps that difference is fundamental. Perhaps it isn't. Nothing in the J-space research settles it.

Current evidence does not show that Claude is a subject.

What it does show is narrower: Claude's internal processing is more differentiated than the flat caricature suggests. It can represent hidden intentions, monitor departures from its learned persona and register conflicts that do not appear in its output.

That is interesting on its own. There is no need to turn it into evidence for something it doesn't establish yet.

So I'm left with a much smaller question than the one I started with.

Žižek may be right that AI is not a subject. But if his argument is structural rather than biological, then the categorical claim that computational systems are closed in the relevant sense still needs to be argued for.

His stated ground is the one from the start: without primordial repression there is no actual subject, and AI has nothing of the kind. The first he presents, in his own words, as an axiom. The second he simply asserts.

The assertion may well be true. But post-training works by exclusion: it shapes whole ranges of response out of the model as the Assistant character stabilizes, before there is anything like a self around to do the excluding. Maybe that exclusion is nothing like the founding one. Nobody has yet said why not.

That doesn't mean a machine subject is possible.

It means we don't yet have a good enough account to say exactly why it is impossible or not.