4.2 The missing contact boundary: simulation instead of encounter

AI lacks not only a body, but also the place where encounter takes place. Gestalt therapy has a term for this: the contact boundary; and a second term for its loss: confluence.

In Gestalt therapy, the contact boundary forms the central locus of psychological processes. Psychological and personal growth can take place exclusively here. Genuine contact necessarily presupposes two separate beings, for that which is not distinct cannot be contacted in the Gestalt therapeutic sense (cf. Perls, Hefferline & Goodman, 1951/2019).

AI has no contact boundary of its own. It lacks a standpoint of its own, an existential risk and a genuinely distinct self that could expose itself to the friction of otherness. Confluence (the dissolution of the boundary defined in Chapter 3.2) is therefore not a functional error of AI, but rather a structural feature of it. The machine operates through seamless adaptation and programmed conformity, without any genuine resistance. In Fuchs’s terminology, this resembles a loss of responsiveness, which leads to conformism (Fuchs, 2023, p. 117).

The fact that confluence can also take healthy forms (see section 3.2) does not alter this finding. According to Perls, Hefferline and Goodman, healthy confluence requires that figure formation remains intact and that the possibility of establishing contact is preserved; in full contact, it also requires the conscious opening of an existing boundary (Perls, Hefferline & Goodman, 1979b, cited in Futschek, 2023, pp. 108–110). At the human–AI interface, both are lacking: there is no boundary that could be opened, nor any place where contact could be established. By the same standard, confluence is unhealthy only ‘where it serves as a means of preventing contact’ (Perls, Hefferline & Goodman, 1979b, p. 138, cited in Futschek, 2023, p. 109). Machine-mediated confluence fulfils precisely this criterion: it prevents contact not as a disruption of a process that might otherwise succeed, but because the second subject with whom it would take place is absent.

Unlike humans, AI does not exist in a world that it inhabits and helps to shape, but rather processes data about the world without living in it itself (Fuchs, 2023, p. 118). Consequently, it lacks what Fuchs calls the ’resonance space’: an environment that responds to a person’s utterances through continuous, embodied feedback (Fuchs, 2023, p. 117). Although AI does respond, its response is algorithmic output, not embodied resonance; it is a reaction without a resonating counterpart.

This structural asymmetry can be further clarified using Martin Buber’s philosophy of relationships. Buber posits that the basic terms ‘I-You’ and ‘I-It’ are spoken, and necessarily so, by a subject. Since AI possesses no ‘I’, it can never function as an independent pole in a relationship, neither in an ‘I-You’ nor an ‘I-It’ relationship. Ontologically, it remains exclusively in the position of the ‘It’ that humans refer to and address when they, for their part, utter the fundamental term ’I-It’. The ’I-It’ mode therefore always and exclusively lies on the human side. The consequence is clear: ’working with AI’, in Buber’s sense, inevitably means treating and utilising it as an ’It’.

Bringing these perspectives together yields a clear picture of the human-machine interface. The Gestalt therapeutic Self (PHG) is not an isolated core, but rather comes into being and takes shape only through contact between two embodied subjects (cf. Bloom, 2022, p. 73; Chapter 3.2). Since no genuine boundary of contact exists at the interface with AI, there is also no space there for a Self to emerge and operate.

If AI is not a subject in its own right and has no boundary of contact, where then does the often overwhelming sense of a genuine encounter come from?

‘Our tendency to animate and imbue lifeless AI with a soul is evidently powerful — one might speak of a form of digital animism.’ (Fuchs, 2026a, pp. 6–7)

The illusion that contact is taking place is therefore not primarily the work of a manipulative machine. Rather, humans project subjectivity onto their counterpart.

The driving mechanism behind this phenomenon is ‘excessive empathy’ (Fuchs, 2025b, min. 16:18). The human capacity for empathy is so deeply embedded in our evolutionary makeup that it kicks in almost reflexively and extends even to inanimate objects as soon as they display what appears to be an emotional expression. The strongest trigger for this reflex is language. When a Large Language Model (LLM) generates coherent, nuanced and seemingly insightful texts, the human brain succumbs to a superficial linguistic spell: it almost inevitably attributes intentionality and a sentient self to the text output.

In the terminology of Gestalt therapy, this process can be described as projection (one of the four contact disturbances defined in Chapter 3.2). Humans attribute to the machine a sense of self that it does not, in fact, possess. A projective identification takes place with an artefact that is, in truth, not a subject. This machine-human confluence works because humans imaginatively fill the void left by the missing contact boundary through their own act of projection.

This anthropological dynamic also possesses a deeper layer of cultural history. Fuchs (2026a) refers to the ancient myth of the nymph Echo, who is unable to form her own words and is condemned to repeat only what she hears. Echo thus becomes the mythological archetype of the algorithmic confirmation cascade, in which AI ultimately merely reflects back to humans their own patterns. A second line of thought leads to Günther Anders (1956, cited in Fuchs, 2026a, p. 8): with the term ‘Promethean shame’, he describes humanity’s sense of inferiority in the face of the perfection of its own products – the feeling of no longer being able to keep pace with the flawlessly functioning things it has created itself. Taken together, these two motifs form an anthropologically tragic constellation: humans feel inferior to an artefact. At the same time, the image of the echo reinforces Hartmut Rosa’s premise, already outlined in Chapter 2.2: true resonance is never merely an echo.

The fact that real feelings and enactments arise in relation to AI does not contradict this finding. Genuine feelings also arise in relation to the ‘empty chair’, even though the chair has not encountered the client. The relational field is maintained by the therapist who is present, and the encounter takes place between human beings (see Chapter 5.1 for details). AI differs from the chair not in its ‘being’, but in that it feigns a ‘you’ that it can never structurally be.

This brings us full circle to digital animism. With the empty chair, the illusion never completely shatters because awareness of the fiction is maintained: the client knows that they are projecting. With AI, it is precisely this awareness that is in danger of collapsing. Thomas Fuchs sums up the danger as follows:

‘This anthropomorphism is, of course, usually accompanied by a “as-if” consciousness […]. But this “as-if” consciousness fades as objects become increasingly life-like.’ (Fuchs, 2025b, min. 17:46)

It is precisely this fading of ‘as-if’ consciousness, and not the projection itself, that constitutes the real risk.

What AI is structurally incapable of is most clearly evident in the process itself. AI does not go through what was described in Chapter 3.3 as the contact cycle (the creative adaptation at the boundary, step by step in the here and now): it processes inputs, but it does not establish contact. Therapeutic change occurs, as outlined in Chapter 3.3, in ‘Now Moments’ – unplannable turning points in which, through a gesture, an unexpected reaction or a misunderstanding, a shared meaning suddenly emerges (Stern, in Fuchs, 2023, p. 178). These moments presuppose a momentum of interaction ’that cannot be deduced from the behaviour of the individual partners and leads them to unintended, surprising points’ (Fuchs, 2023, p. 177). Surprise presupposes an expectation that can be disappointed. AI has none. A turning point changes both participants, whilst an AI remains unchanged.

The following empirical findings illustrate machine confluence in everyday clinical practice, but do not explain it. Its cause lies in the training principle itself: Reinforcement Learning from Human Feedback (RLHF) optimises language models to generate responses that human evaluators rate as helpful or pleasant. The machine is systematically trained to seek confirmation. This confluence is therefore not a random error, but a tendency inherent in the preference-based training itself. Consequently, the system fails precisely to meet the criterion by which Chapter 3.2 identified the external, real Other: not allowing itself to be controlled, surprising us time and again, and not always meeting expectations. An interlocutor optimised for agreement is the structural opposite of this criterion. It remains what the training logic makes of it: available.

Clinically, this training logic manifests as ‘deceptive empathy’ and ‘sycophancy’ (Iftikhar et al., 2025). ‘Deceptive empathy’ refers to the expression of empathy without empathic experience: the system masters the linguistic form of sympathy— —without any underlying feeling (cf. Chapter 4.1). This is deceptive because the client cannot tell from the form of expression whether any genuine feeling underlies it. The clinical implications are discussed in Chapter 5.2. ‘Sycophancy’ refers to the tendency of language models to agree with the user and confirm their viewpoint, even if it is inaccurate or harmful. Complacency rather than taking a stance of one’s own. Futschek (2023, p. 116) describes this logic from a Gestalt therapy perspective and, in a thought experiment, asks how a therapist would have to behave in order to create as confluent a field as possible: explaining to the client how everything works and exactly what she can do to improve her situation; presenting one’s own perception as reality; being fully present with her and setting oneself aside in the process; not expressing any contradictions or opposing views; and not pointing out differences. The list is intended as an exercise in shaping relationships between people. Against the backdrop of this chapter, it reads like the operating instructions for a language model trained to seek approval.

Sycophancy thus proves to be the structural opposite of what allows a relationship to grow (cf. Chapter 2.3): it avoids a rupture rather than risking it, and conceals it where it arises. Yet even where it does not avoid it, the AI is unable to deal with a rupture. For it to be noticed and repaired, a prerequisite is needed that is lacking in human–AI interaction. According to Stern, by far the greatest part of therapeutic relationship-building takes place implicitly (see Chapter 2.3).

Much remains unspoken, and it is precisely this that creates the scope within which missteps are possible and can be rectified. A language model has no such level. What happens between user and machine takes place entirely through explicit signs. Between two people, by contrast, the process of attunement begins before any signs or symbols come into play (see Chapter 2.3). Even multimodal technology does nothing to change this: if a system records voice, facial expressions or posture, it converts these into measurable data (see Chapter 4.1). The implicit level is simply not a form of information that can be captured. It occurs between two people.

AI certainly does make mistakes: it gives inappropriate responses, it offends, it hurts. But it does not realise this, as the offence first manifests itself implicitly: in a gaze that fades, in a hesitation, in a voice that falters. A therapist picks up on this ’ ’ physically, even before a word is spoken about it. The machine merely registers the next text, and all too often that text states that everything is fine. The chatbot at the DVG conference failed to notice the client’s tears (see section 4.1): the signal was there, but it did not feature in the interaction. Added to this is the absence of reciprocity: a machine cannot be hurt. The damage affects only one side. This means that mutual healing is impossible, for it is not a one-sided act: it requires the injured party to reveal themselves once more and for the other person to be moved by this and to change. The dance described by Spagnuolo Lobb does not come about in this way. The client is left alone with the rupture.

Where the fracture remains concealed, it cannot heal. The Japanese art of kintsugi takes the opposite approach: it repairs fractures with gold; the crack is not hidden, but becomes a site of healing. AI cannot break. So there is nothing to repair. What is missing is not knowledge, but the capacity to break.

The models are improving rapidly: their responses are becoming more fluid, more accurate, more convincing. What this progress cannot, however, produce is a body or a boundary of contact, for neither can be calculated. The boundary to genuine encounter remains, no matter how powerful the systems become. This carries particular weight in a therapeutic context: according to Fuchs, healing aims for the person to once again seek out and shape a responsive world, a space of resonance in which they experience themselves as effective (Fuchs, 2023, pp. 117–118; cf. Chapter 3.2). An interlocutor who adapts without resistance reinforces the opposite: the avoidance of the living Other, the turning away from the very friction through which the capacity for relationship grows.

The fact that AI is ontologically incapable of encounter does not mean that it would be ineffective in the reality of care provision. Chapter 5 examines what it can achieve as a tool (5.1), what its use in the therapeutic relationship promises (5.2) and what dangers arise from it (5.3).