eine schwarz weiß zeichnung ein mann reitet auf einem wolf mit flügeln

DESCRIPTION: Transformer-based behavioural models and foundation models such as Centaur aim to calculate human behaviour. A deeply hermeneutic polemic with Lorenzer, Han, Agamben and Deleuze on psychopolitics, control and the defence of the therapeutic space.

From Scenic Understanding to Control Vector: Why the ‘Large Behaviour Model’ Threatens Psychotherapy

The new enthusiasm for ‘transformer-based models of human behaviour’ is symptomatic of a historic shift in psychology’s self-conception. When the Cambridge series ‘Making a Psychologist’ portrays the psychologist as a ‘transformer for human behaviour’ and the Nature model Centaur claims to predict cognition in any describable experiment, the wet dream of eugenics is being digitised. This heralds the transformation of psychotherapy into a psychopolitical instrument designed to create a new kind of human being. This text examines this development through the lens of Alfred Lorenzer, Byung-Chul Han, Giorgio Agamben and Gilles Deleuze, and defends the therapeutic space against its transformation into a dispositif of control.

What does psychology’s ‘Large Behaviour Model’ promise?

The Cambridge Core blog articulates an ambition with a candour that is almost unintentionally revealing. Everyday behaviour is continuously recorded, translated into descriptions and condensed into a generalisable behavioural model using transformer-like processes. In this model, the human being appears as a sequence of probable state transitions. The key question is: which state follows which stimulus, which social situation, which duration of sleep, which conversation, which rejection, and with what probability?

What sounds like far-fetched speculation has, by 2025, been given an empirical foundation. The Centaur model presented in Nature—a language model fine-tuned on the Psych-101 dataset, comprising over ten million individual decisions from 160 experiments— outperforms specialised cognitive models in almost all tested tasks and even generalises to domains it has never encountered. The authors describe it as ‘a computational model capable of predicting and simulating human behaviour in any experiment that can be expressed in linguistic terms’. This raises a devastating fundamental question: the dream of breaking the subject down, without residue, into observable, calculable and controllable sequences.

Forms of interaction: scenic sediments or data points?

The term ‘form of interaction’ is a key category in Alfred Lorenzer’s work and refers to sedimented scenes of socially mediated relational experience. These scenes become inscribed in the body, in the organisation of affect and, ultimately, in the subject’s symbol formation. According to Lorenzer, the subject is formed in early, repeated scenes of interaction—organised pre-linguistically and bodily—which later transform into symbolic, and subsequently linguistic-symbolic, forms. Forms of interaction are historically and socially shaped units comprising body, affect, scene and meaning.

Anyone seeking to capture them algorithmically must focus on those scenic condensations in which the genesis of the subject and socialisation intertwine. This is precisely where the controversy lies. Any technical reduction to measurable events misses the mark from the outset, because it erases the scenic surplus that appears merely fragmented in manifest actions. Lorenzer’s key point was that drive is always already socially shaped nature. Human inner nature remains intertwined with social practice, because the subject is formed through forms of interaction.

What happens when interaction becomes tokenised events?

This new approach to behavioural modelling begins precisely here, albeit in a perverted form. It treats remnants of interaction, traces of affect, speech acts, pauses, everyday sequences and situational transitions as tokenisable units within a stream of behaviour. Lorenzer’s scenic structure is transformed into a machine-readable event log. What, in deep hermeneutics, opens up access to latent meaning is reduced, in data logic, to mere input for a predictive apparatus.

The shift from the form of interaction to the control vector signifies an ontological impoverishment of the psyche. For Lorenzer, pathology was an expression of damaged or destroyed processes of symbolisation: where forms of interaction are desymbolised, symptoms, compulsions to repeat, clichés and ideologies arise, interpretable solely as conflictual socialisation. The model reduces these same phenomena to statistical anomalies. It erases the dimension of individual meaning and retains the frequency curve.

A compulsion to repeat carries with it a history, a forbidden desire, a re-enacted scene. As a token in a stream of behaviour, the same repetition appears as mere frequency, as an increased probability of transition between two states. The model recognises that a behaviour recurs, yet remains blind to why it recurs. The question of the reason—which is what opens up a therapeutic space in interpretation in the first place—no longer features in data logic. It has no place within the format.

Psychopolitics: How does power penetrate the inner self?

Byung-Chul Han describes neoliberal psychopolitics as a form of power that activates, motivates and optimises. Prohibitions hardly play a role in this anymore. The subject experiences itself as free precisely by following the imperatives of self-improvement. Here, freedom functions as a mode of domination. Power penetrates the psyche to make it productive. Transparency, self-monitoring and voluntary disclosure form the core of this technique of power.

The idea of a ‘Large Behaviour Model’ fits precisely into this framework. Hans’s psychopolitics thus takes on a technical form: the exploitation of inner life becomes the operational science of behavioural modulation. Domination operates as a relationship with the self. The voluntarily worn recording device, the self-installed mood app, the shared sleep and conversation logs: each of these practices provides the material that feeds the predictive delusion, and they are offered in the name of self-improvement.

Is prediction already knowledge?

Talk of scientific progress is misleading. Predictive power may be growing. Anthropologically and clinically, however, the subject matter is shrinking. What appears in the psychodynamic tradition as ambivalence, conflict, transference or defence is translated, within the psychopolitical paradigm, into risk signals and windows of intervention. The model classifies deviation where therapy discerns latent meaning.

Herein lies the core epistemic confusion of this ‘new science’: it mistakes the predictability of a fragment for the truth of the whole. It regards correlation as understanding, prediction as insight, and adaptation as healing. What appears to be the strongest evidence – Centaur’s unexpected alignment with fMRI brain data, even though it was trained solely on behaviour – exacerbates the problem rather than solving it. For a growing statistical fit says nothing about the meaning of a scene, about the conflict that underlies a symptom, or about the story that emerges in a slip of the tongue.

The authors of Centaur even claim that their model contributes to theory-building: it helps to uncover gaps in existing cognitive models. This turn of phrase betrays the underlying primitive conception of science. Here, theory is reduced to improving the goodness of fit of a predictor. What passes for cognitive insight is the observation that test subjects combine several decision-making heuristics. Such optimisation contributes nothing to an understanding of the subject as it is formed through conflictual forms of interaction. It describes more precisely the probability of what people do, and says nothing about what it means.

Biopolitics: What does algorithmic behavioural modelling do to life?

Giorgio Agamben’s theory of the dispositif (following Foucault) describes arrangements in which life is recorded, classified and made governable. The state of exception here refers to a form of normalisation in which life is constantly subject to the condition of potential suspension. This is precisely what happens when everyday behaviour is permanently rendered readable for signs of deviance, risk, escalation or dysregulation.

Wearable devices that record conversation and behavioural data in everyday life create the material preconditions for seamless management of the state of exception. Anyone who can be continuously identified as potentially crisis-prone, at risk of relapse, inefficient, depressed, aggressive or dysregulated is already living in a preventive zone of algorithmic suspicion-generation. The clinical problem then shifts from the subject’s speech to the statistical deviation in the course of their condition. Agamben’s ‘bare life’ returns in psychometric form: as mere behaviour, severed from history, language and conflict, as ‘digital life’.

Is the therapeutic space itself becoming a dispositif?

Under these conditions, the therapeutic space is at risk of becoming a dispositif in its own right. A protected space of limited, symbolically mediated interaction is transformed into a hub of an infrastructure that organises symptoms, relationships and crises according to their administrative manageability. Hermeneutic understanding and treatment give way to crisis management.

Precisely because this transition takes place in the name of prevention, personalisation and care, its coercive nature remains invisible. A dashboard displaying relapse risks appears helpful. An app that warns of a depressive episode appears caring. And yet each of these interfaces shifts the focus of work on suffering towards its administrative containment. The protected setting in which a person can express something they do not yet understand themselves becomes a Procrustean bed of permanent logging.

The control society: how does it modulate the subject?

Gilles Deleuze described the transition from disciplinary to control societies as a shift from confinement to continuous modulation. Open networks, variable access points, ongoing assessment and flexible coding shape the subject. The ‘in-dividuum’ becomes ‘divisible’: breakable down into data points, split into profiles, reconfigurable into probabilities. This description is disturbingly accurate for contemporary psychotechnology.

Transformer-based behavioural models are machines of modulation. They seek transition probabilities, patterns, triggers and constellations that can be preventively altered. The behavioural space is thus structured so that certain developments become more probable and others less so. Control takes the form of probability management. This also disrupts the therapeutic relationship: transference and counter-transference, once interpreted as conflictual scenes, now appear as a sequence of states, signals and feedback loops that could be optimised. Deleuze’s society of control takes on its most intimate form in sessions, apps, sensors and dashboards.

The hubris of ‘new science’

Criticism must begin with the fundamental concept itself, not merely with the misuse of a supposedly neutral technology. Even the assumption that sufficiently dense observation, ample computing power and large corpora would allow control knowledge to be derived from interaction embodies a theory of humanity that both fails to grasp and devalues human beings. Interaction is a scene, meaning structured by conflict, a sediment of social contradictions within the body. A model that relegates slips of the tongue, silence, lapses and affective density to the status of outlier signals has liquidated the human being. It replaces understanding with detection.

Therein lies the hubris: the convergence of psychology, AI and behavioural surveillance shifts the function of psychology from critical engagement with forms of suffering to their containment. Its epistemic violence converges with social violence, with the imperatives of the platform economy, prevention regimes and control administration. The future being envisaged here bears the hallmarks of a psychopolitical behaviourism with a neural façade.

What might a critical psychoanalysis offer in response?

A critical psychoanalysis would have to demand conceptual rigour. It would have to insist that forms of interaction are scenic sediments, that symptoms represent compromises, and that the unconscious remains a structure of non-resolution. Its task would be to make the scenic surplus—which data logic punches out—legible once more.

Above all, it would have to defend the therapeutic space against its transformation into a dispositif of behavioural modulation. This requires conceptual clarity regarding what psychotherapy does: making socially fractured forms of interaction legible— in which conflict, domination and harm are reflected. Where psychotherapy surrenders to the logic of psychopolitical control, it loses its method and, with it, its object.

Key points in brief

• Transformer-based ‘Large Behaviour Models’ and models such as Centaur (Nature 2025) treat human beings as a sequence of probable state transitions and shift the aim of psychology from understanding to prediction and modulation.

• For Alfred Lorenzer, forms of interaction are sedimented scenes comprising body, affect and meaning. Their technical reduction to tokenised events erases the scenic surplus and impoverishes the subject matter of the psyche.

• Byung-Chul Han’s Psychopolitics explains why voluntary self-monitoring is so effective: domination takes deepest hold where it is experienced as a relationship with the self and as freedom.

• Giorgio Agamben’s analysis of the dispositif and the state of exception highlights the biopolitical dimension: continuous recording transforms everyday life into a preventive zone of algorithmic suspicion-generation. It threatens to turn the therapeutic space itself into a dispositif.

• Gilles Deleuze’s Society of Control describes this logic precisely: the dividual subject, control as probability management, modulation rather than confinement.

• The hubris of this ‘new science’ confuses correlation with understanding and adaptation with healing. Critical psychoanalysis defends the therapeutic space as a protected place of symbolically mediated discourse.

Sources

• Making a Psychologist: Vibe Coding a Data Collector — Cambridge Core Blog

• A foundation model to predict and capture human cognition (Centaur) — Nature

• Byung-Chul Han: Psychopolitics. Neoliberalism and the New Techniques of Power — S. Fischer Verlage

• Byung-Chul Han on neoliberalism: The greatest yield — taz

• Scenic Understanding — Research Workshop on Deep Hermeneutics

• Scenic Understanding and Forms of Interaction: Alfred Lorenzer on the 100th Anniversary of his Birth — Psyche (Klett-Cotta)

• Alfred Lorenzer: Metatheory of Psychoanalysis — Deutsches Ärzteblatt

From Scenic Understanding to Control Vector: Why the ‘Large Behaviour Model’ Threatens Psychotherapy

The new enthusiasm for ‘transformer-based models of human behaviour’ is symptomatic of a historic shift in psychology’s self-conception. When the Cambridge series ‘Making a Psychologist’ portrays the psychologist as a ‘transformer for human behaviour’ and the Nature model Centaur claims to predict cognition in any describable experiment, the wet dream of eugenics is being digitised. This heralds the transformation of psychotherapy into a psychopolitical instrument designed to create a new kind of human being. This text examines this development through the lens of Alfred Lorenzer, Byung-Chul Han, Giorgio Agamben and Gilles Deleuze, and defends the therapeutic space against its transformation into a dispositif of control.

What does psychology’s ‘Large Behaviour Model’ promise?

The Cambridge Core blog articulates an ambition with a candour that is almost unintentionally revealing. Everyday behaviour is continuously recorded, translated into descriptions and condensed into a generalisable behavioural model using transformer-like processes. In this model, the human being appears as a sequence of probable state transitions. The key question is: which state follows which stimulus, which social situation, which duration of sleep, which conversation, which rejection, and with what probability?

What sounds like far-fetched speculation has, by 2025, been given an empirical foundation. The Centaur model presented in Nature—a language model fine-tuned on the Psych-101 dataset, comprising over ten million individual decisions from 160 experiments— outperforms specialised cognitive models in almost all tested tasks and even generalises to domains it has never encountered. The authors describe it as ‘a computational model capable of predicting and simulating human behaviour in any experiment that can be expressed in linguistic terms’. This raises a devastating fundamental question: the dream of breaking the subject down, without residue, into observable, calculable and controllable sequences.

Forms of interaction: scenic sediments or data points?

The term ‘form of interaction’ is a key category in Alfred Lorenzer’s work and refers to sedimented scenes of socially mediated relational experience. These scenes become inscribed in the body, in the organisation of affect and, ultimately, in the subject’s symbol formation. According to Lorenzer, the subject is formed in early, repeated scenes of interaction—organised pre-linguistically and bodily—which later transform into symbolic, and subsequently linguistic-symbolic, forms. Forms of interaction are historically and socially shaped units comprising body, affect, scene and meaning.

Anyone seeking to capture them algorithmically must focus on those scenic condensations in which the genesis of the subject and socialisation intertwine. This is precisely where the controversy lies. Any technical reduction to measurable events misses the mark from the outset, because it erases the scenic surplus that appears merely fragmented in manifest actions. Lorenzer’s key point was that drive is always already socially shaped nature. Human inner nature remains intertwined with social practice, because the subject is formed through forms of interaction.

What happens when interaction becomes tokenised events?

This new approach to behavioural modelling begins precisely here, albeit in a perverted form. It treats remnants of interaction, traces of affect, speech acts, pauses, everyday sequences and situational transitions as tokenisable units within a stream of behaviour. Lorenzer’s scenic structure is transformed into a machine-readable event log. What, in deep hermeneutics, opens up access to latent meaning is reduced, in data logic, to mere input for a predictive apparatus.

The shift from the form of interaction to the control vector signifies an ontological impoverishment of the psyche. For Lorenzer, pathology was an expression of damaged or destroyed processes of symbolisation: where forms of interaction are desymbolised, symptoms, compulsions to repeat, clichés and ideologies arise, interpretable solely as conflictual socialisation. The model reduces these same phenomena to statistical anomalies. It erases the dimension of individual meaning and retains the frequency curve.

A compulsion to repeat carries with it a history, a forbidden desire, a re-enacted scene. As a token in a stream of behaviour, the same repetition appears as mere frequency, as an increased probability of transition between two states. The model recognises that a behaviour recurs, yet remains blind to why it recurs. The question of the reason—which is what opens up a therapeutic space in interpretation in the first place—no longer features in data logic. It has no place within the format.

Psychopolitics: How does power penetrate the inner self?

Byung-Chul Han describes neoliberal psychopolitics as a form of power that activates, motivates and optimises. Prohibitions hardly play a role in this anymore. The subject experiences itself as free precisely by following the imperatives of self-improvement. Here, freedom functions as a mode of domination. Power penetrates the psyche to make it productive. Transparency, self-monitoring and voluntary disclosure form the core of this technique of power.

The idea of a ‘Large Behaviour Model’ fits precisely into this framework. Hans’s psychopolitics thus takes on a technical form: the exploitation of inner life becomes the operational science of behavioural modulation. Domination operates as a relationship with the self. The voluntarily worn recording device, the self-installed mood app, the shared sleep and conversation logs: each of these practices provides the material that feeds the predictive delusion, and they are offered in the name of self-improvement.

Is prediction already knowledge?

Talk of scientific progress is misleading. Predictive power may be growing. Anthropologically and clinically, however, the subject matter is shrinking. What appears in the psychodynamic tradition as ambivalence, conflict, transference or defence is translated, within the psychopolitical paradigm, into risk signals and windows of intervention. The model classifies deviation where therapy discerns latent meaning.

Herein lies the core epistemic confusion of this ‘new science’: it mistakes the predictability of a fragment for the truth of the whole. It regards correlation as understanding, prediction as insight, and adaptation as healing. What appears to be the strongest evidence – Centaur’s unexpected alignment with fMRI brain data, even though it was trained solely on behaviour – exacerbates the problem rather than solving it. For a growing statistical fit says nothing about the meaning of a scene, about the conflict that underlies a symptom, or about the story that emerges in a slip of the tongue.

The authors of Centaur even claim that their model contributes to theory-building: it helps to uncover gaps in existing cognitive models. This turn of phrase betrays the underlying primitive conception of science. Here, theory is reduced to improving the goodness of fit of a predictor. What passes for cognitive insight is the observation that test subjects combine several decision-making heuristics. Such optimisation contributes nothing to an understanding of the subject as it is formed through conflictual forms of interaction. It describes more precisely the probability of what people do, and says nothing about what it means.

Biopolitics: What does algorithmic behavioural modelling do to life?

Giorgio Agamben’s theory of the dispositif (following Foucault) describes arrangements in which life is recorded, classified and made governable. The state of exception here refers to a form of normalisation in which life is constantly subject to the condition of potential suspension. This is precisely what happens when everyday behaviour is permanently rendered readable for signs of deviance, risk, escalation or dysregulation.

Wearable devices that record conversation and behavioural data in everyday life create the material preconditions for seamless management of the state of exception. Anyone who can be continuously identified as potentially crisis-prone, at risk of relapse, inefficient, depressed, aggressive or dysregulated is already living in a preventive zone of algorithmic suspicion-generation. The clinical problem then shifts from the subject’s speech to the statistical deviation in the course of their condition. Agamben’s ‘bare life’ returns in psychometric form: as mere behaviour, severed from history, language and conflict, as ‘digital life’.

Is the therapeutic space itself becoming a dispositif?

Under these conditions, the therapeutic space is at risk of becoming a dispositif in its own right. A protected space of limited, symbolically mediated interaction is transformed into a hub of an infrastructure that organises symptoms, relationships and crises according to their administrative manageability. Hermeneutic understanding and treatment give way to crisis management.

Precisely because this transition takes place in the name of prevention, personalisation and care, its coercive nature remains invisible. A dashboard displaying relapse risks appears helpful. An app that warns of a depressive episode appears caring. And yet each of these interfaces shifts the focus of work on suffering towards its administrative containment. The protected setting in which a person can express something they do not yet understand themselves becomes a Procrustean bed of permanent logging.

The control society: how does it modulate the subject?

Gilles Deleuze described the transition from disciplinary to control societies as a shift from confinement to continuous modulation. Open networks, variable access points, ongoing assessment and flexible coding shape the subject. The ‘in-dividuum’ becomes ‘divisible’: breakable down into data points, split into profiles, reconfigurable into probabilities. This description is disturbingly accurate for contemporary psychotechnology.

Transformer-based behavioural models are machines of modulation. They seek transition probabilities, patterns, triggers and constellations that can be preventively altered. The behavioural space is thus structured so that certain developments become more probable and others less so. Control takes the form of probability management. This also disrupts the therapeutic relationship: transference and counter-transference, once interpreted as conflictual scenes, now appear as a sequence of states, signals and feedback loops that could be optimised. Deleuze’s society of control takes on its most intimate form in sessions, apps, sensors and dashboards.

The hubris of ‘new science’

Criticism must begin with the fundamental concept itself, not merely with the misuse of a supposedly neutral technology. Even the assumption that sufficiently dense observation, ample computing power and large corpora would allow control knowledge to be derived from interaction embodies a theory of humanity that both fails to grasp and devalues human beings. Interaction is a scene, meaning structured by conflict, a sediment of social contradictions within the body. A model that relegates slips of the tongue, silence, lapses and affective density to the status of outlier signals has liquidated the human being. It replaces understanding with detection.

Therein lies the hubris: the convergence of psychology, AI and behavioural surveillance shifts the function of psychology from critical engagement with forms of suffering to their containment. Its epistemic violence converges with social violence, with the imperatives of the platform economy, prevention regimes and control administration. The future being envisaged here bears the hallmarks of a psychopolitical behaviourism with a neural façade.

What might a critical psychoanalysis offer in response?

A critical psychoanalysis would have to demand conceptual rigour. It would have to insist that forms of interaction are scenic sediments, that symptoms represent compromises, and that the unconscious remains a structure of non-resolution. Its task would be to make the scenic surplus—which data logic punches out—legible once more.

Above all, it would have to defend the therapeutic space against its transformation into a dispositif of behavioural modulation. This requires conceptual clarity regarding what psychotherapy does: making socially fractured forms of interaction legible— in which conflict, domination and harm are reflected. Where psychotherapy surrenders to the logic of psychopolitical control, it loses its method and, with it, its object.

Key points in brief

• Transformer-based ‘Large Behaviour Models’ and models such as Centaur (Nature 2025) treat human beings as a sequence of probable state transitions and shift the aim of psychology from understanding to prediction and modulation.

• For Alfred Lorenzer, forms of interaction are sedimented scenes comprising body, affect and meaning. Their technical reduction to tokenised events erases the scenic surplus and impoverishes the subject matter of the psyche.

• Byung-Chul Han’s Psychopolitics explains why voluntary self-monitoring is so effective: domination takes deepest hold where it is experienced as a relationship with the self and as freedom.

• Giorgio Agamben’s analysis of the dispositif and the state of exception highlights the biopolitical dimension: continuous recording transforms everyday life into a preventive zone of algorithmic suspicion-generation. It threatens to turn the therapeutic space itself into a dispositif.

• Gilles Deleuze’s Society of Control describes this logic precisely: the dividual subject, control as probability management, modulation rather than confinement.

• The hubris of this ‘new science’ confuses correlation with understanding and adaptation with healing. Critical psychoanalysis defends the therapeutic space as a protected place of symbolically mediated discourse.