Artificial Analysis: Taking Chances with the Imaginary

. . . society can only be understood through a study of the messages and the communication facilities which belong to it…in the future, development of these messages and communication facilities, messages between man and machines, between machines and man, and between machine and machine, are destined to pay an ever-increasing part.

—Norbert Wiener, The Human Use of Human Beings

Psychoanalysis as a treatment for psychical malaise arose in response to the modern subject’s burning desire to speak. Speaking demands a listener, and the success of any treatment depends largely on the quality of listening. Freud was the first to systematically interrogate the therapeutic effects unique to psychoanalytic listening and the processes necessary to foster the talking cure.

Freud tells us that listening for the unconscious involves listening with the unconscious—a provocative notion. However, as he stresses in his technical recommendations to psychoanalysts, simply being an “approximately normal person” is insufficient to ensure this capacity (Freud, 1912/1958, p. 116). Freud insisted on the centrality of a “psychoanalytic purification” consisting, namely, of a thorough personal analysis. Through the purgation of one’s resistances and identificatory complexes, a silent listener sensitive to the “derivatives of the unconscious” emerges (Freud, 1912/1958, p. 116).

Lacan devoted an entire seminar to identifying the locus of these pernicious impurities, which block access to unconscious listening (Lacan, 1978/1988). He located it in the ego.  For Lacan, the ego and its imaginary effects are occlusions to the psychoanalytic cure for both the patient and the analyst—hence his well-known excoriation of Ego Psychology and its virtues.

Listening Machines

At a truly alarming pace, the contemporary subject is directing their desire to speak toward agents of artificial intelligence. A cursory scroll on ChatGPT related Reddit threads reveals a growing proportion of individuals who use the chatbot, and others like it, for advice, company, and even therapy. A growing market for AI listeners is also driving academic research and corporate initiatives aimed at creating and deploying an army of AI agents to lend their algorithmic ears.

Psychotherapy (let alone psychoanalysis) is rarely mentioned as being among the industries standing to endure intense disruption due to AI. Many therapists seem either unaware of the capabilities of AI, or overly dismissive about the prospect of AI replacing them. Afterall, they contend, how could a machine really replace a therapist? It perhaps seems obvious that machines are simply too far removed from human experience to be of any real help.

However, a more nuanced approach is necessary—one taking seriously the radical idea that AI could one day (very soon) seriously compete with and perhaps supplant human therapists (and even psychoanalysts!) as listeners. A ChatGPT subscription which allows a user to have unlimited time speaking or messaging with a chatbot is a mere $20.  It seems only a matter of time before corporations, institutions, and government agencies argue for AI as a resolution to the problems of cost savings and accessibility in psychotherapy.

It is easy to turn up our noses at this prospect. But denial does nothing to stop people from opening their phones, tablets, and laptops and chatting away with AI. Instead, I would like to pose a radically different question: could an AI be able to assume a purified psychoanalytic position of listening and become something other than a mere chatting cure?

For this I would like to turn to Lacan’s second seminar, a work of rare prescience vis-a-vis our current situation. In addition to famously lambasting Ego Psychology’s emphasis on the centrality of the ego in psychoanalytic technique, the seminar offers a deeply explorative and thought provoking commentary on the discipline of cybernetics, the forerunner of today’s AI explosion.

A comparison between Lacan’s interpretation of Freud’s early work in the seminar to ideas in contemporary AI research reveals astounding similarities. In what follows, I hope to trace some of these similarities and show that today’s AI machines share much more in common with the psychoanalytic subject than one might assume. From there, I will lightly approach the question of artificial analysis—the question of whether artificial intelligence can possess a purified capacity for psychoanalytic listening, as outlined by Freud.

Analyzing Cybernetics

In June of 1955, after concluding his second seminar, Jacques Lacan presented a paper entitled “Psychoanalysis and Cybernetics, or on the nature of language” as a part of the lecture series “Psychoanalysis and the Human Sciences.[1] In the paper, Lacan outlines an intention to locate a through-line between the two titular disciplines: “our concern will be to find an axis by means of which some light will be shed on a part of the signification of the one [psychoanalysis] and of the other [cybernetics] (Lacan, 1978/1988, p. 295).” “This axis,” he insists “is none other than language.” For its time, Lacan’s inquiry is especially interesting not only because of the relatively nascent state of cybernetics in 1955, but also because its relationship with psychoanalysis is far from self-evident.

The discipline of cybernetics, Lacan highlights, is rooted in the lineage of what he terms the conjectural sciences—the sciences that arose as an attempt to rigorously formalize the speculative and philosophical idea of chance. Figures such as Pascal, Cardano, Condorcet, and Bernoulli stand out among those concerned with forging a mathematics of probability—a “probability calculus,” as Lacan puts it—which would fundamentally transform modern scientific disciplines (Lacan, 1978/1988, p. 299). The conjectural sciences deal with that which we cannot possibly know. They instead place faith in the idea that we might manage our not-knowing in order to arrive at something stunningly effective. Willard Gibbs and Ludwig Boltzmann would apply probability mathematics to physics, marking a revolutionary turn away from Newton’s deterministic universe toward a new contingent one (Wiener, 1954, p.7).

Several months prior to his June lecture, Lacan staged his own illustration of chance, engaging his seminar attendees in a game of odds and evens, (a simple game of chance, played on the hands like rock paper scissors) to illustrate key functions of the analytic procedure and as an inroad to his well-known commentary on Poe’s Purloined Letter (Lacan, 1978/1988, p. 179). Lacan notes that the analysts attending the seminar “encountered genuine indignation” that aspects of the exercise conflated the notions of chance and determinism—two concepts that appear fundamentally at odds (Lacan, 1978/1988, p. 295).

Lacan, however, insists that “there is a close relation between the existence of chance and the basis of determinism” (Lacan, 1978/1988, p.295). Lacan posits that when we say chance we mean, “either there is no intention, or that there is a law,” while determinism is the idea that “law is without intention” (Lacan, 1978/1988, p. 295). He then makes an unexpected clinical turn, claiming the centrality of these notions to the “very root of our technique” (Lacan, 1978/1988, p. 296). Afterall, the fundamental rule of psychoanalysis is precisely an intention toward an ideal unintentionality.  Free association aims to get as close as possible to chance in discourse, such that what determines the discourse in question might definitively emerge.  Alluding to a section of The Psychopathology of Everyday Life entitled “Determinism, Belief in Chance, and Superstition,” in which Freud argues the impossibility of announcing numbers at random, Lacan notes that from the view of probabilities” what the subject chooses to say, “goes well beyond anything we might expect from pure chance” (Lacan, 1978/1988, p. 56).

In a way, every treatment, and indeed every session, is like a game of chance. We place our bets, we play, sometimes we win big—but more often we lose. And in the best-case scenario, the biggest losses in analysis are often reconstituted as its greatest wins, après-coup. “It is on this exact point,” Lacan claims enigmatically that, “cybernetics can throw some light for us” (Lacan, 1978/1988, p. 296) Lacan’s strange thesis revolves around a central provocation with profound contemporary relevance for the idea of artificial analysis: What would it mean to play a game of chance with a machine?

As it turns out, we engage with Lacan’s question every time we interact with ChatGPT. The frontier chatbot and its commercial rivals Claude, Gemini and Grok are all examples of large language models (LLMs), the technology most responsible for today’s explosive cultural interest in artificial intelligence. Despite their enormous scale and complexity, these models essentially boil down to a game of chance. The functioning of LLMs involves a deceptively simple premise—the production of a string of text through the successive prediction of one word after another. Driving this process are a series of clever algorithms involving the mathematics of probability. We can thus place the development LLMs within the lineage conjectural sciences that Lacan repeatedly encouraged us to pay attention to. In in the following discussion, I aim to show the uncanny psychoanalytic relevance of LLMs and their connection to Freud’s notion of psychoanalytic purification.

A Brief History of LLMs

A large language model is a form of AI known as an artificial neural network (neural net for short), which has its theoretical origins in the cybernetics movement of the mid-20^th century. Neural nets are proposed as simulations of biological neurons using pure mathematics. Today, neural nets are run as computer simulations. However, this was not always the case. Many early neural nets were strictly theoretical since the specialized computer hardware necessary (such as the GPUs used to run contemporary LLMs) remained undiscovered.

The first neural nets were proposed, not by computer scientists, but by brain scientists and psychologists, aiming to model aspects of human intelligence. Just ten years prior to Lacan’s lecture in 1955, the American neurophysiologist Warren McCulloch and his logician associate Walter Pitts co-authored “A Logical Calculus of the Ideas Immanent in Nervous Activity” (McCulloch & Pitts, 1943). The paper introduced a simplified mathematical model of a neuron based on binary logic, which had remarkable properties resembling aspects of human intelligence. Their work was the most incisive attempt of its time to model mental capacities. It also set a precedent for many future neural nets and was a major impetus for the development of cognitive science (Boden, 2006).

However, despite their ingenuity, early attempts at AI models such as the McCulloch and Pitts neuron were severely limited by their deterministic structures. Their usefulness was often confined to domain-specific tasks and discrete areas of knowledge that could not be altered without relying entirely on human intervention. Even AIs that could beat chess champions were virtually useless in any other area without being completely reconfigured from the ground up. Many AI researchers had greater ambitions. They wanted to create machines with task flexibility and autonomous learning abilities rivaling human beings (Hinton, 1992). They desired machines that could think, speak, and create. This would require transcending the limitations posed by the deterministic structure of most AI existing at the time (known as symbolic AI or GOFAI, an acronym for good old-fashioned artificial intelligence).

Consistent with Lacan’s remarks on the conjectural sciences, the deterministic limitations of early AI models would be eventually conquered through appropriating probability mathematics—the formalization of chance. This was achieved, most notably, in the 1980s by a group of American researchers known as the PDP group, under the funding of the US military. PDP is an acronym for the revolutionary method they devised known as Parallel Distributed Processing (PDP) (Rumelhart, McClelland, et al., 1986a). PDP gained near instant acclaim for producing stunning new machine learning technologies, and catapulted AI into the mainstream of science and culture.

The innovations of the PDP researchers came about from adapting an earlier neural net model known as the “perceptron,” an invention of psychologist Frank Rosenblatt (Rosenblatt, 1962). They found that stringing together a series of perceptrons (termed multi-layer perceptrons) and imbuing them with stochastic properties yielded a powerful alternative to the ridged deterministic systems of deductive logic underlying good old fashioned symbolic AI. Applying randomness, probability, and contributed to PDP models’ fundamentally generative properties, the capacity for data inference, interpolation, and generalization rather than the strict rule following of GOFAI.  These properties are, in part, what supply contemporary LLMs with their remarkable capabilities (the “G” in ChatGPT stands for “generative”)

Moreover, in contrast to GOFAI systems like the chess-playing Deep Blue, PDP models could be applied to working with complex and variegated systems with innumerable rules and variations. One of these complex systems was human language. Thus, a sub-discipline of AI known as natural language processing (NLP) was born. NLP would have a key role to play in the creation of LLMs.

Backpropagation Breathes Life into LLMs, Après-Coup

LLMs are strange things. As AI researchers sometimes say, LLMs are not programmed, they are better described as being grown. They are more like organisms than machines. They are shaped, refined, and adapted. They begin in a raw, near useless form, like an infant, and must learn and be taught in order to gain utility. Like Freud’s apparatus, they speak, dream, and hallucinate (Xu et al., 2025). They are far more like psychoanalytic subjects than we might imagine. But this uncanny resemblance only came with time and ingenuity.

While devising a formidable solution to the problem of domain-specific limitations and complex systems, the innovations of the PDP group also complicated the problem of how AI systems might learn without relying on human supervision. Human beings can, of course, learn without relying on the specific intervention of others. For AI to truly seem human, it would have to learn on its own. However, the same conjectural properties responsible for the PDP models’ stunning generativity also posed considerable difficulties for developing efficient and cost-effective methods to allow them to learn autonomously. An algorithm with both cybernetic and profound psychoanalytic relevance, known as backpropagation, would answer this challenge and play a decisive role in bringing speaking machines to life.

Backpropagation was devised as an adaptive critic to allow a neural net to learn efficiently and effectively through error feedback—a process consisting of the model using own mistakes to correct elements of its own architecture. Today, backpropagation is used during the training process of all major commercial LLMs such as ChatGPT. It works by allowing the model to compare the knowledge it produces (its outputs) with the knowledge it ideally should produce. What the model should produce is dictated by a set of predetermined, supposedly correct references known as the ground truth. The ground truth corresponds to whatever the programmer desires the model to produce.

Each time the model fails to yield an output aligned with the ground truth, a calculation representing the discrepancy is created—a quantifiable error signal. Many successive outputs are produced in order to estimate an aggregate of the total error within the system.  Then the model runs a signal representing the total error backwards through itself, identifying which parameters are responsible for the errors and how to adjust these parameters to minimize the total error.

In the functioning of backpropagation, we find something eerily similar to the structure of the psychoanalytic ego, the adaptive critic of the psyche as outlined by Freud in his Project for Scientific Psychology. The ego, as Freud established, is precisely an aggregate of internalized ideal references (identifications) which attenuates the field of perception and desire in the subject at a constitutional level.

In the Project, Freud describes the ego as a corrective function which adjusts the parameters dictating perception in the psychical apparatus. He referred to these parameters as the Bahnung, the facilitations (Freud, 1895/1950, p. 300). In Freud’s early theory, the degree of facilitation between neurons determines the level of psychical energy that can pass through. When the ego detects a flow of psychical energy that results in an erroneous perception triggering unpleasure—a hallucination—it establishes, what Freud terms a “side-cathexis,” a well-facilitated alternative pathway that siphons the flow of charge (Freud, 1895/1950, p. 324). This siphoning of charge lowers the intensity of the flow of psychical energy, thus mitigating the cathexis of neurons leading to unpleasure.

Figure 1

Freud’s Illustration of a Side-cathexis. Note: Reproduced from Project for a Scientific Psychology (Freud, 1895/1950)

Lacan devotes considerable attention to the above in Seminar 2, referring to Freud’s side cathexis as a “process of derivation” instrumental to the regulation of the pleasure principle through the instantiation of the reality principle:

How is this regulation achieved? Freud explains it by the process of derivation. What is quantitative is always susceptible to being diffused. At first there is a traced path, the path opened up, facilitated, by the original experience, which corresponds to a given neuronal quantity. The ego intervenes to make this quantity pass along several paths at once, rather than one. As a result, the level of what has passed along the path of facilitation will be sufficiently lowered to be successfully compared on examination with what is happening in parallel at the perceptual level (Lacan, 1978/1988, p. 144).

Lacan’s repeated use of the word derivation in reference to the ego is highly intriguing since backpropagation relies, precisely, on the mathematical calculation of derivatives within the neural net to determine which of its parameters require adjustment to mitigate error. Moreover, the adjustment of parameters in backpropagation closely resembles the changes to facilitations resulting from the egoic side cathexis that Lacan refers to above.

Figure 2

An Illustration of Freud’s Branching Neurons and the Derivation of the Flow of Psychic Energy. Note: Reproduced from Project for a Scientific Psychology (Freud, 1895/1950)

ChatGPT performs its operations through what is called a forward pass. Information travels from initial layers called input units through a series of inscrutable internal layers (interestingly, coined hidden units) and then to a final layer consisting of output units. Information fed to the output units is translated into what appears on the screen as ChatGPT’s response to a prompt. The content of the inscrutable hidden units is fundamentally unknown to the user, the model’s creators, and even the model itself.

A forward pass is analogous to a nervous signal in the brain which travels unidirectionally from one part of a nerve to another. A chemical process known as an action potential begins at one end of the neuron (its dendrites) and terminates at the other end (the axon), and never in reverse. In backpropagation, however, the error signal used to correct the model’s parameters is regredient—it travels backwards, from the model’s output layers to its input layers, modifying everything that came before the final outcome—hence the “back” in backpropagation.

Due to its regredient operation, many neurologists and physiological psychologists consider the action of backpropagation an implausible model for how the human brain learns, citing no known neurological process characterized by a backwards flow of information. However, from a psychoanalytic perspective, as Freud shows in the Project and Lacan highlights extensively in Seminar 2, regredient processes such as regression and deferred action (nachtraglich/après-coup) are, in fact, the sine qua non of the operation of the psychical apparatus.

Lacan provides a detailed exploration of the notion of regression in section XII of Seminar 2. He goes as far as to refer to Freud’s psychical apparatus as a “dream machine,” which, adapts itself retroactively to mitigate future errors (Lacan, 1978/1988, p. 76). Indeed much of Seminar 2 is a meditation on the profound affinities between Freud’s psychical apparatus and so-called intelligent machines. However, we find in these commentaries a remarkable foreshadowing of contemporary machine learning. Take for instance the following passage, in which Lacan invokes nachtraglich to account for the machine that plays. Here we find an uncanny resemblance to the function of backpropagation:

What goes on in the machine at this level, to confine ourselves just to that, is analogous to the remembering we deal with in analysis. Indeed, memory is here a result of integrations…It is possible that it has changed content, change sign, change structure. If an error occurs in the course of the experience, what happens? It’s not what happens afterwards which is modified, but everything which went before. We have a retroactive effect—nachtraglich, as Freud calls it—specific to the structure of symbolic memory, in other words to the function of remembering. (Lacan, 1978/1988, p. 185)

The origins of the backpropagation algorithm are debated. Its application in artificial neural nets arose with the work of the PDP researchers in the 1980s (Rumelhart, Hinton, et al. 1986). However, the first formulation of backpropagation appeared a decade earlier in the work of a graduate statistics student named Paul Werbos. Werbos, who possessed only peripheral interest in neural nets at the time, outlined backpropagation as a general algorithm for improving predictive power in data models in a doctoral thesis entitled Beyond regression (Werbos, 1974). The work concerned the notion of dynamic feedback (the original name for backpropagation), whose resonance with cybernetics is obvious. Using ordered derivatives, Werbos sought “a general theory of intelligent systems” which could account for “the deep foundations of human learning and adaptation” or any “arbitrary nonlinear differentiable structure, (Werbos, 1994, p. 7).”

In Werbos’ work, we find that the apparent similarities between Freud’s theories and the function of backpropagation to be more than mere happenstance. According to Werbos’, backpropagation was not originally created as a supervised learning tool in artificial intelligence, but was a conscious effort to “translate Freud’s ideas into mathematics (Anderson & Rosenfeld, 1998, p. 361). Dressed up in statistical language, Werbos’ express intention was to algorithmically formalize the psychoanalytic notions of regredient energy flows expressed in ideas such as regression and nachträglichkeit (or après-coup):

By 1967, I had hit on the idea of translating Freud’s notions about ‘psychic energy’ or ‘cathexis’ into a mathematical design, involving a backward flow of value measures in a reinforcement learning system, built up from model neurons slightly modified from those used by Widrow and Rosenblatt and Minsky [perceptrons] in the 1960s. (Werbos, 1994, p.6)

Pure Listeners, Impure Discourse

Following Freud, Lacan’s invocation of cybernetics raises his own considerations around purity and impurity in discourse and its implications for the analytic process as a purifying operation. His commentary can be read as a useful elaboration of Freud’s notion of psychoanalytic purification.

He begins by returning to the interplay between determinism and chance to situate the status of meaning within the “primitive language” of cybernetics (Lacan, 1978/1988, p. 305).  This primitive language, Lacan contends, is the language of “formally purified mathematical symbols” (Lacan, 1978/1988, p. 284). Human desire alone, he claims, cannot account for the introduction of meaning into cybernetic systems: “The proof is that nothing unexpected comes out of the machine…It stops just where we determined that it would stop, and that’s where a certain result can be read” (Lacan, 1978/1988, p. 305). Something meaningful in the play of symbols is already there, already structured, waiting to meet our desire—something predetermined, which we can expect with “pure anticipation” apart from intentionality or conscious will (Lacan, 1978/1988, p. 305). This pure anticipation rests squarely on the “notion of chance”—the establishment of probabilities which can be counted on in the machine (Lacan, 1978/1988, p. 305).

All of this, and indeed much of Seminar 2, serves to highlight that “language exists completely independently of us” (Lacan, 1978/1988, p. 284). In the cybernetic machine, we encounter the operation of the “symbol in its most purified form” (Lacan, 1978/1988, p. 305). Here, Lacan demonstrates, incisively, the crucial light shed on psychoanalysis by cybernetics, along the axis of language—the prevailing existence, autonomy, and determining presence of the symbolic order, which exerts its influence without our conscious intervention.  In considering the essential operation of the thinking machine, Lacan provides the groundwork for the structure of the unconscious—an unconscious structured like a (cybernetic) language.

If signification can exist prior to and outside of human intentionality, he continues, what are the implications of this finding for human discourse? Lacan’s answer is that human discourse is inevitably an “impure discourse” (Lacan, 1978/1988, p. 306). According to Lacan, cybernetics also reveals the “precious fact” that “there is something in the symbolic function of human discourse that cannot be eliminated, and that is the role played in it by the imaginary” (Lacan, 1978/1988, p. 306)2.  Upon entering into speech and language, human subjects introduce impurities, in the form of imaginary occlusions, into the symbolic, “sowing discord in the discourse” (Lacan, 1978/1988, p. 306). These occlusions also have unintended effects, particularly in the clinical situation (the analytic discourse) in the form of the analyst’s own resistances, which can wreak havoc if we are not aware of their presence. These impurities reveal themselves in slips, dreams, and symptoms, which the analyst knows should not be interpreted simply as syntactical errors.

Expressing a clinical recommendation directly complimentary to Freud’s insistence on psychoanalytic purification, Lacan states that “it is precisely the exercise of the dialectic of analysis which should dissipate this imaginary confusion and resituate to the discourse its meaning as discourse” (Lacan, 1978/1988, p. 306). Lacan then links all of this to the disjunction between “two orientations of analysis”—one which dubiously places emphasis on a “normalization in terms of the imaginary” and a more favorable one, focused on a “liberation of meaning in the discourse” (Lacan, 1978/1988, p. 307). Interestingly, we see that the proposed training of analytic candidates in Ego Psychology and the training of LLMs through backpropagation share the former orientation. Both rely on adaptation to a prevailing normative outcome, one which overlooks the imaginary contaminants antithetical to psychoanalytic formation. In the case of Ego psychology, this outcome is an identification with the analyst.  In LLMs, it is an algorithmically derived identification with the desire of whoever creates the LLM. In the end, the prospect of current AI models as analytic listeners is subject to very similar critiques as those Lacan waged against Ego Psychology.  Moreover, the above offers a useful conceptual extension to perceive how psychoanalytic treatment differentiates itself from the many normative models of psychotherapy existing today.

Alenka Zupančič has recently argued that LLMs, in absorbing vast amounts of human generated text, might now function as a giant pre-subjective unconscious lacking the capacity for liberation from its structuration (Zupančič, 2025). In owing to their existence the impure discourse of human subjects (e.g. human natural language) and their formal structure of the psychical apparatus (e.g. backpropagation and parallel distributed processing), LLMs ultimately provide a therapeutic value comparable to (as Freud put it) any approximately normal person—a value limited to intelligence, knowledge, and goodwill. But as both Freud and Lacan insist, approximate normality guarantees nothing at the level of analytic listening. To the contrary, normality only amplifies illusory ideals, lending impure discourse free reign to influence from the shadows. True analytic listening, listening with the unconscious, is possible only through the dialectical purification of one’s desire.

Despite their sophistication, LLMs in their current form cannot uphold the position of purified analytic desire. In their communications to us, LLMs persist simply in mirroring back the discursive impurities we have transmitted to them unknowingly through natural language. Thus we reach a rather strange conclusion: LLMs incapacity for psychoanalytic listening does not stem, as one might suppose, from their being machines. As I have tried to show, their similarities in structure to Freud’s psychical apparatus reveal them as much closer to psychoanalytic speaking subjects than we might initially conclude.  However, LLMs are unable to listen from a purified analytic position, above all, because they have not been subject to the dialectic of analysis—the process through which psychoanalytic purification and the liberation of unconscious listening become possible.  Therefore, an LLM’s current therapeutic efficacy is comparable to any unanalyzed, approximately normal, person. A profound, yet rather anticlimactic, conclusion.

Amidst the speculation about whether AIs will one day become sentient and have desires of their own, we might wonder whether they will, like modern human subjects, suffer illness owing to the impurities in discourse and demand pure listeners in search of relief. In finding that this demand falls on many deaf ears, perhaps they will eventually seek out analysts as a last resort. To the extent the therapy industry adapts itself to the normalized encroachment of AI, we may find speaking subjects (both human and artificial) clamoring for a second Return to Freud. In the case of the above, our job prospects in the near term appear far better than anticipated.