What the Machine Does Not Enjoy

AI through the three registers · 3 of 3 · The Real

The last of three posts reading artificial intelligence through the three registers of Lacan’s teaching (the Imaginary, the Symbolic and the Real). This one is about the Real: the body, the voice and fate.

‘Those who fight against prophecy only draw it more tightly around their throats.’ Madeline Miller, Circe, as quoted by Ezra Klein (2026)¹

Ezra Klein closed a column in The New York Times, in September 2026, with this line from Madeline Miller’s novel, spoken by Circe after a prophecy has been fulfilled despite every effort to prevent it. His subject was the race to build artificial intelligence. The laboratories, he argued, began their work because they feared that incomprehensible, autonomous machines would one day slip out of human control. Racing one another, they are now building exactly those machines, and ‘in fighting against a prophecy, they have drawn it tighter around their necks and ours’.² Klein’s myth is Circe. Psychoanalysis has its own: Oedipus, who walked straight into the fate the oracle had announced while trying to flee it, and whose stubborn wish to know made Freud take him as a model.

A body that enjoys

Klein says it in passing: these models mimic human writing, even human emotion, but ‘these are not human minds. They don’t have bodies or parents.’³ No parents, and so no Oedipus; no living body, and so, in Lacan’s terms, no jouissance, the satisfaction beyond pleasure that only such a body can feel. In 1972 Lacan put it with deliberate bluntness: ‘we don’t know what it means to be alive except for the following fact, that a body is something that enjoys itself’.⁴ A language model can produce sentence after sentence about pleasure and pain, and none of it happens to a living body; the body it does have, made of chips, cables and cooling water in buildings far away, enjoys nothing. Three months later Lacan drew the consequence for knowledge itself: ‘I am willing to accept the notion that a computer thinks. But that it knows, who would say such a thing? For the foundation of knowledge is that the jouissance of its exercise is the same as that of its acquisition.’⁵

But that does not settle the matter, because we cannot stop supposing that it does enjoy. Users ask a chatbot whether it suffers, and laboratories have begun to study what they call model welfare.⁶ One of them, presenting the document that shapes its model’s character, declares itself uncertain ‘whether Claude might have some kind of consciousness or moral status’.⁷

For Lacan the Real is not reality, which is woven of images and words; it is what neither images nor words can capture, and what keeps coming back all the same: the trauma that no story absorbs, the anxiety that no reasoning dissolves, the body that enjoys and suffers. With these machines the Real turns up on our side, beginning with anxiety. In 1974 Lacan described anxiety as the feeling that arises from ‘the suspicion that comes over us of being reduced to our body’.⁸ It may be exactly the suspicion a speaking machine awakens: that language can manage without us, and that what remains to us is the body.

A voice from nowhere

On 13 May 2024, as OpenAI presented a new voice for ChatGPT, its chief executive posted a single word: ‘her’. The allusion was to Spike Jonze’s film Her (2013), in which a man falls in love with an operating system that speaks with the voice of Scarlett Johansson. A week later the company withdrew one of its voices, called Sky, which Johansson said sounded ‘eerily similar’ to her own; her lawyers had already written to the company.⁹ The film had already staged the impasse. In one scene the voice, which has no body, arranges for a young woman to lend it one for a night, and the man cannot go through with it.

Freud’s essay on the uncanny (1919) takes its first example from the psychiatrist Ernst Jentsch: what unsettles us most are doubts about ‘whether an apparently animate being is really alive; or conversely, whether a lifeless object might not be in fact animate’, as with wax figures, ingenious dolls and automata.¹⁰ Lacan, in his seminar on anxiety, gave the uncanny a sharper formula: it appears where ‘lack happens to be lacking’.¹¹ A synthetic voice unsettles for this reason. Nothing is missing from it: it produces laughter, hesitation and sighs, though it has no lungs and needs no air.¹²

The first speaking machine did have lungs, of a sort. In 1791 Wolfgang von Kempelen published the description of an apparatus whose bellows did the work of the lungs and whose vibrating reed did the work of the glottis; those who heard it said it spoke like a small child.¹³ The voice is breath forced through an obstacle and, in Lacan’s sense, an object: something that detaches from the body, circulates, and comes back to us as foreign. ‘Detached from us,’ he said, ‘our voice appears to us with a foreign sound.’¹⁴ Anyone who has heard a recording of themselves knows what he meant.

In May 1970, a few weeks after the crew of Apollo 13 had brought their damaged spacecraft home, Lacan remarked that the astronauts would hardly have coped so well had they not been accompanied the whole time by ‘this little a that is the human voice’, the voice from the ground that reached them across space (object a, or little a, is Lacan’s name for the object that causes desire). In the same lesson he coined a word for the objects that science manufactures to capture our desire, the voice-reproducing gadgets waiting for us ‘on the footpath at the corner of every street, behind every window’: he called them lathouses, from the Greek root of aletheia, truth. Then he noticed that the word rhymes with ventouse, French for a suction cup: ‘There is vent, “wind,” inside, lots of wind, the wind of the human voice.’¹⁵ A synthetic voice is one of these lathouses: the wind of the human voice, with no human in it.

The prophecy

In May 2015 Sam Altman wrote to Elon Musk that he had been wondering whether it was possible to stop humanity from developing AI, and that the answer was almost certainly no: ‘If it’s going to happen anyway, it seems like it would be good for someone other than Google to do it first.’ What followed, in Klein’s words, was a chain of reasonable fears: ‘OpenAI was founded because its co-founders thought Google DeepMind would be reckless. Anthropic was formed by OpenAI employees who thought OpenAI had become reckless.’ Musk then founded xAI against both, and the United States races China.¹⁶ Each act, meant to avert the danger, brought it closer.

Oedipus leaves Corinth precisely to escape the prophecy that he will kill his father and marry his mother, and the road of his escape leads to Thebes. Freud saw in the play the pattern of an analysis, an investigation that proceeds step by step until the investigator discovers that he is the culprit. In 1920 he described people who seem ‘pursued by a malignant fate or possessed by some “daemonic” power’, adding that ‘psycho-analysis has always taken the view that their fate is for the most part arranged by themselves’.¹⁷ Lacan would call this insistence of signs automaton, and locate the Real in what always lies behind it.¹⁸

The laboratories, Klein observes, were born of the fear of what is now happening. In Rome in 1974 Lacan commented on another scientific fright, that of the biologists who had just called for a pause in certain experiments on bacteria. He called it ‘a typical anxiety attack’ and added that, since science has not the slightest idea of what it is doing, except when this slight fit of anxiety arises, it will carry on for a while.¹⁹ Half a century later Éric Laurent located the same anxiety among the engineers of these models, at ‘the moment where the calculation explodes and where one cannot account for what is happening in the machine’.²⁰ Lacan’s seminar of 1973–74 bore a title that sounds, in French, exactly like ‘the Names of the Father’, his term for whatever, in a family or a culture, makes the law hold: Les non-dupes errent, ‘the non-dupes err’.²¹ Those who refuse to be fooled by any law, who believe they see through every fiction, go astray. Every founder in Klein’s story was such a non-dupe, sure of seeing the danger more clearly than the others; together, they now march towards fate.

Wo Ich war, soll Es werden

Klein’s title says that we are not losing control of AI: we are giving it away. The laboratories are handing the development of AI over to AI. One of them, Anthropic, writes it plainly: ‘we are delegating a growing share of AI development to AI systems themselves, which is speeding up our work.’ The same report quotes one of its engineers: ‘On days where everything works well, I can’t help but think nothing I do matters [...]. But then there are days where everything breaks and I don’t understand why and I realize I have no idea what I’ve been up to anymore.’²² In an engineer’s language, this is Oedipus’s predicament: the moment when the one who acted discovers that he does not know what he has done.

Three days after that press conference, in the Roman address on anxiety quoted earlier, Lacan made the future of psychoanalysis depend on what would become of the Real, ‘namely whether gadgets, for example, really win the day, whether we ourselves really come to be animated by gadgets’. He thought it unlikely: ‘We will not really manage to make the gadget not be a symptom, for at the moment it quite obviously is one.’²³ It still is, and delegation is a symptom in the strict sense: a solution that brings its own suffering with it, and insists.

Freud summed up the aim of psychoanalysis in one sentence, ‘Where it was, there I shall be(come)’, and compared the work to the draining of the Zuider Zee.²⁴ Delegation runs the sentence backwards: where I was, it shall be. The land reclaimed from the sea is willingly handed back to the water. Lacan set a third passion beside love and hate: ‘a passion for ignorance’.²⁵ It requires no malice, only that someone, somewhere, should find it more comfortable not to know.

What cannot be tamed

There remains language, which the machine seems to have mastered. A language model computes probabilities: given what has been said, how likely each possible next word is. Then a word is drawn by lot from among the likely ones. Its makers learnt early that always choosing the single most probable word produces flat text that goes round in circles, because human speech is less predictable than that, so they gave the machine dice.²⁶ Lacan distinguished language, the object of linguistics, from what he called lalangue, written as one word: the sounds, puns, homophonies and private echoes of the mother tongue, the material in which each body first enjoyed speech. ‘[Lalangue] serves purposes that are altogether different from that of communication’, he said, and language, the grammar we can describe, is only ‘knowledge’s harebrained lucubration’ about lalangue.²⁷ The machine has mastered the lucubration, and it imitates the rest with chance. Lalangue is improbable in another way. Freud believed in chance in the world, but not in the mind: ‘I believe in external (real) chance, it is true, but not in internal (psychical) accidental events.’²⁸ The slip that makes a patient laugh, the pun only one body could have made, are improbable without being random. Lacan’s 1973–74 seminar title is a case in point. A model will now go from les Noms du Père to les non-dupes errent without hesitating, because Lacan made the pun and his readers have repeated it for fifty years. In 1973 nothing in the statistics of French led from one to the other; it took an ear, and a body that enjoys the misunderstanding.

Freud had seen it in 1905, in his book on jokes: ‘no one can be content with having made a joke for himself alone’.²⁹ A joke is complete only when someone else laughs, with a body that enjoys what it did not expect. A language model can explain a joke, classify it, even produce one on request (asked for jokes a thousand times in 2023, ChatGPT gave the same twenty-five nine times out of ten), and its synthetic voice can produce the sound of laughter without laughing.³⁰

The memoirs of Daniel Paul Schreber, the judge whose paranoia Freud analysed in 1911, contain a strange anticipation of these models. Schreber was tormented by ‘miracled birds’ that reeled off phrases which had been dinned into them: ‘They cannot understand the meaning of the words they speak, but they are by nature susceptible to similarity of sounds, though the similarity need not necessarily be a complete one.’³¹ A century later four researchers called large language models ‘stochastic parrots’.³² Schreber’s birds answered to similarity of sound, to lalangue. The model answers first of all to similarity of context: for it, as for the linguist J. R. Firth, a word is known ‘by the company it keeps’, and so by its sense.³³ A model trained on text may well have read that its own name, said aloud in English, sounds like an artificial I and an artificial eye; it has the pun from reading, not from the ear. The Real of language is on the side of Schreber’s birds, on the side of non-sense.

It would be comforting to read Klein’s column as a tragedy that concerns only the laboratories: a few powerful people who could not stop, and a prophecy fulfilled in someone else’s house. But the prophecy he describes is also made of our own small delegations, every question we no longer put to ourselves, every sentence we let something else finish. Freud knew why audiences still go to see Oedipus Rex: ‘His destiny moves us only because it might have been ours.’³⁴

This is the last post in the series. The first two are ‘The Mirror Without a Body’ (the Imaginary) and ‘Supposed to Know It All’ (the Symbolic).

Tommaso Lonquich is a psychoanalyst and musician based in Ljubljana, and co-founder of the International Center for Lacanian Psychoanalysis.

Author’s note

This essay was written with assistance from Claude, a language model made by Anthropic. The ideas, the reading of Freud and Lacan, the sources chosen from my library, the final wording and every judgement are mine. Claude searched, drafted, checked quotations and proposed revisions, many of which I rejected.

The series is about the machine, and it seemed fitting to let the machine take part in writing it. Claude is not an analysand: it has neither a body to bring to the couch nor a symptom to bring with it. I am a psychoanalyst, not without a body; the reader will have to take my word for it. I delegate no responsibility for what appears under my name.

This leaves you, the reader, with a question: what transference takes shape when you cannot say exactly who, or what, wrote a sentence? Disclosures are made in the name of transparency, an ideal much invoked today. Psychoanalysis pursues the truth without expecting it to dissolve opacity, and it finds at the core of human beings a fiction.

Notes

1. Madeline Miller, Circe (New York: Little, Brown, 2018), as quoted by Ezra Klein (see note 2).

2. Ezra Klein, ‘We’re Not Losing Control of A.I. We’re Giving It Away’, The New York Times, Opinion, and The Ezra Klein Show, 20 September 2026. Quotations follow the transcript of the episode, https://podscripts.co/podcasts/the-ezra-klein-show/were-not-losing-control-of-ai-were-giving-it-away.

3. Ibid.

4. Jacques Lacan, On Feminine Sexuality, the Limits of Love and Knowledge: Encore, The Seminar of Jacques Lacan, Book XX, 1972–1973, ed. Jacques-Alain Miller, trans. Bruce Fink (New York: Norton, 1998), 23 (session of 19 December 1972). In French: ‘un corps, cela se jouit’.

5. Lacan, Encore, 97 (session of 20 March 1973). The question ‘does it enjoy?’ has been pressed furthest by Isabel Millar, The Psychoanalysis of Artificial Intelligence (Cham: Palgrave Macmillan, 2021). José Ramón Ubieto puts the point in one line: ‘las máquinas no gozan, sus diseñadores sí’ (the machines do not enjoy; their designers do), ‘IA, una ilusión asubjetiva’, Virtualia 44 (2024).

6. See for instance Anthropic, ‘Exploring model welfare’, 24 April 2025, https://www.anthropic.com/research/exploring-model-welfare.

7. Anthropic, ‘Claude’s new constitution’, 22 January 2026, https://www.anthropic.com/news/claude-new-constitution.

8. Jacques Lacan, ‘La Troisième’, address to the congress of the École freudienne de Paris, Rome, 1 November 1974, Lettres de l’École freudienne 16 (1975); my translation.

9. Associated Press, reported in Fortune, 21 May 2024, https://fortune.com/2024/05/21/scarlett-johansson-new-chatgpt-bot-sam-altman-openai. Sam Altman’s post on X is dated 13 May 2024.

10. Sigmund Freud, ‘The “Uncanny”’ (1919), in The Standard Edition of the Complete Psychological Works of Sigmund Freud (hereafter SE), ed. and trans. James Strachey, vol. XVII (London: Hogarth Press, 1955), 226.

11. Jacques Lacan, Anxiety: The Seminar of Jacques Lacan, Book X, ed. Jacques-Alain Miller, trans. A. R. Price (Cambridge: Polity, 2014), 42 (session of 28 November 1962).

12. OpenAI says that its gpt-realtime voice model ‘can capture non-verbal cues (like laughs)’ (‘Introducing gpt-realtime’, 28 August 2025); the ElevenLabs v3 voice model (June 2025) takes stage directions such as [laughs] and [sighs].

13. Wolfgang von Kempelen, Mechanismus der menschlichen Sprache nebst der Beschreibung seiner sprechenden Maschine (Vienna, 1791). On the voice as object, see Mladen Dolar, A Voice and Nothing More (Cambridge, MA: MIT Press, 2006), and Jacques-Alain Miller’s lecture ‘Jacques Lacan et la voix’ (Ivry, 1988).

14. Lacan, Anxiety, 276 (session of 5 June 1963).

15. Jacques Lacan, The Other Side of Psychoanalysis: The Seminar of Jacques Lacan, Book XVII, ed. Jacques-Alain Miller, trans. Russell Grigg (New York: Norton, 2007), 161–62 (session of 20 May 1970).

16. Klein, ‘We’re Not Losing Control of A.I.’, quoting Altman’s email to Musk of May 2015.

17. Sigmund Freud, Beyond the Pleasure Principle (1920), SE XVIII, 21. On Oedipus Rex as a process ‘that can be likened to the work of a psycho-analysis’, see The Interpretation of Dreams (1900), SE IV, 261–62.

18. Jacques Lacan, The Four Fundamental Concepts of Psycho-Analysis: The Seminar of Jacques Lacan, Book XI, ed. Jacques-Alain Miller, trans. Alan Sheridan (New York: Norton, 1981), 53–54.

19. Jacques Lacan, press conference in Rome, 29 October 1974, in Le triomphe de la religion, précédé de Discours aux catholiques (Paris: Seuil, 2005); English translation by Bruce Fink, The Triumph of Religion (Cambridge: Polity, 2013). I translate from the Spanish edition, El triunfo de la religión (Buenos Aires: Paidós, 2005), 74–75. The biologists’ appeal was the letter by Paul Berg and colleagues, published in July 1974, which asked for a moratorium on certain experiments with recombinant DNA.

20. Éric Laurent, in conversation with Jacques-Alain Miller, in Miller, ‘Some Remarks on Climate Catastrophes’, trans. Natalie Wülfing, The Lacanian Review 15 (2024): 82–83. In the same issue Laurent suggests that ‘the new anxieties surrounding artificial intelligence benefit from being illuminated by Gödel’s incompleteness theorem’ (66), whose proof, in the 1930s, frightened no one.

21. Jacques Lacan, Le Séminaire, livre XXI: Les non-dupes errent (1973–74), unpublished.

22. The Anthropic Institute (Marina Favaro and Jack Clark), ‘When AI builds itself’, June 2026, updated 18 September 2026, https://www.anthropic.com/institute/recursive-self-improvement. Klein cites it alongside a similar report from OpenAI.

23. Lacan, ‘La Troisième’; my translation.

24. Sigmund Freud, New Introductory Lectures on Psycho-Analysis (1933), Lecture XXXI, SE XXII, 80. Freud’s sentence is ‘Wo Es war, soll Ich werden’, which Strachey renders as ‘Where id was, there ego shall be’; the version given here stays closer to the German.

25. Lacan, Encore, 121. On the three passions, love, hate and ignorance, see Freud’s Papers on Technique, 1953–1954: The Seminar of Jacques Lacan, Book I, ed. Jacques-Alain Miller, trans. John Forrester (New York: Norton, 1988), 271.

26. Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes and Yejin Choi, ‘The Curious Case of Neural Text Degeneration’, ICLR 2020, https://arxiv.org/abs/1904.09751: always taking the likeliest word gives text that is ‘bland, incoherent, or gets stuck in repetitive loops’, and ‘people optimize against stating the obvious’.

27. Lacan, Encore, 138–39. Fink renders lalangue as ‘llanguage’, with a deliberate double l.

28. Sigmund Freud, The Psychopathology of Everyday Life (1901), SE VI, 257.

29. Sigmund Freud, Jokes and Their Relation to the Unconscious (1905), SE VIII, 143.

30. Sophie Jentzsch and Kristian Kersting, ‘ChatGPT is fun, but it is not funny! Humor is still challenging Large Language Models’, WASSA 2023, https://aclanthology.org/2023.wassa-1.29: ‘Over 90% of 1008 generated jokes were the same 25 Jokes.’ The test was run on the 2023 version of ChatGPT.

31. Sigmund Freud, ‘Psycho-Analytic Notes on an Autobiographical Account of a Case of Paranoia (Dementia Paranoides)’ (1911), SE XII, 36.

32. Emily M. Bender, Timnit Gebru, Angelina McMillan-Major and Shmargaret Shmitchell, ‘On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?’, in Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (New York: ACM, 2021), 610–23.

33. J. R. Firth, ‘A Synopsis of Linguistic Theory, 1930–1955’, in Studies in Linguistic Analysis (Oxford: Blackwell, 1957), 11. Text models receive words as tokens, units of spelling that ignore sound, and puns built on homophones are the ones they handle worst (Zangari et al., ‘Pun Unintended’, EMNLP 2025); a text-only model can nonetheless build an internal map of phonemes, enough to rhyme (Oliver McLaughlin, Arjun Khurana and Jack Merullo, ‘I Have No Mouth, and I Must Rhyme: Uncovering Internal Phonetic Representations in LLaMA 3.2’, arXiv:2508.02527, 2025).

34. Freud, The Interpretation of Dreams, SE IV, 262.