5.1 Effectiveness, Areas of Application and the Concept of a Tool
Generative AI will transform the landscape of psychotherapy. The question is therefore not whether it will become established, but what it can actually achieve within clearly defined tasks. Current research demonstrates efficacy at the symptomatic level, but at the same time calls for a sober assessment of opportunities and risks (Hillebrand & Baumeister, 2025).
The most robust evidence comes from the Therabot study (Heinz et al., 2025), the first randomised controlled trial of a fully generative AI therapy system (n = 210). It recorded clinically significant reductions in symptoms across all three areas studied (depression, anxiety and increased risk of eating disorders), with predominantly large effect sizesFootnote 1 Cohen’s d ranged from 0.845 to 0.903 for depressive symptoms, from 0.794 to 0.840 for generalised anxiety, and from 0.627 to 0.819 for the risk of eating disorders (Heinz et al., 2025). . As a tool for symptom reduction, generative AI must therefore be taken seriously from an empirical perspective. This effect can largely be explained by the common mechanisms of action found in psychotherapy (Giotakos, 2025).
The most realistic application scenario is seen where AI was used not as a substitute but as a tool in human hands (Habicht et al., 2025): when used as an between-session support within group-based cognitive behavioural therapy, its use was associated with higher participation rates and fewer dropouts. From a clinical perspective, this is consistent with the assessment of practitioners (Hipgrave et al., 2025).
The meta-analysis by Li et al. (2023) also supports the view that AI is effective. The most recent meta-analysis by Sohn et al. (2026, 39 studies) both confirms and qualifies this finding: the effects on depressive (g = 0.31) and anxiety symptoms (g = 0.28) are moderate, and 35 of the 39 included studies show a high risk of bias, primarily due to reliance on self-report measures alone. However, the effect is fleeting: after three months, no demonstrable long-term effects could be detected (Zhong et al., 2024). This could be an initial indication that an AI-triggered process without face-to-face interaction remains superficial.
These findings are not isolated. Fitzpatrick et al. (2017) had already demonstrated a significant reduction in symptoms of depression and anxiety in the Woebot study. A randomised controlled trial involving women in active war zones shows that the tool is particularly useful where human care is in short supply: The AI chatbot achieved a significant reduction in anxiety , but remained inferior to traditional psychotherapyFootnote 2 Anxiety reduction: 45–50 per cent (conventional therapy) vs. 30–35 per cent (chatbot); Spytska, 2025. (Spytska, 2025). This does not contradict the relational thesis of this work; rather, it clarifies it: the relationship is a key factor in the therapeutic effect, but not the only one (see Section 2.1). Elements such as expectation, structure and encouragement already point to a perceived sense of care, and such a sense certainly arises in interactions with chatbots (see Section 5.2). In this way, an effect at the symptomatic level can be achieved. The extent of this effect is demonstrated by the following findings.
At the level of task quality, AI goes even further. In a cross-sectional study, six language models achieved higher knowledge scores than trainee psychotherapists in behavioural activation programmes (Napiwotzki et al., 2025)Footnote 3 Trainees, however, showed significant improvement following targeted training. .
It is telling that even a paper which explicitly acknowledges the advantages of AI (scalability, constant availability, reduced stigmatisation) identifies the lack of genuine empathy as a key limitation and classifies AI as a complement to, rather than a replacement for, human therapy (Zhang & Wang, 2024).
Alongside clinical symptom management, AI is opening up a second, less controversial field: the administrative and organisational support of therapeutic practice. The technology-oriented specialist literature demonstrates this ‘tool’ mode in detail: from process reflection and diagnostic support to the preparation of interventions (Raile & Geißler, 2025), as well as in the AI-assisted evaluation of medical histories and the creation of worksheets (Eichenberg et al., 2026). A shift is also becoming apparent in process research: AI enables an automated, multidimensional analysis of therapeutic trajectories, a task previously reserved for manual coding (Steppan & Birkhölzer, 2025). The fact that this structuring approach also has clinical merit was already demonstrated by the group therapy study, in which the AI tool replaced conventional worksheets (Habicht et al., 2025).
It is precisely here that the separation of levels is most evident: where AI documents, structures and processes data, its ‘it’ character is obvious and unproblematic. The therapist treats it as they would any other tool, without any sense of a personal ‘you’ coming into play.
However, this ability to delegate is subject to a significant legal caveat as soon as highly sensitive client data is involved: even seemingly harmless administrative tasks bring us to the heart of data protection and professional conduct law (see Chapter 5.3 for details). The administrative benefit is therefore real, but subject to one condition: the therapist remains responsible under data protection and professional conduct law.
It is questionable whether therapists are currently able to fulfil this responsibility. Augustin et al. (2026) are the first in the German-speaking world to systematically investigate the attitudes of 335 psychotherapists towards AI and their level of competence in using it. They found an average level of AI literacy, a very low ability to help design such systems themselves, and a clear generational effect: younger practitioners are more familiar with the technology.
Therapists are already using AI, but mostly as passive users who have little understanding of how these systems learn from human feedback (Rein forcement Learning). And those who do not understand how it works can hardly assess its limitations.
The real question, however, is how AI directly affects people in a clinical context and where its limits lie.
The reason lies in the mechanism of action itself: for this mechanism, it does not matter where the impetus comes from, because meaning arises from a person’s bodily experience (enactive approach; see Chapter 2.3; Fuchs, 2023). An AI text can therefore choose exactly the right words and thereby trigger a genuine inner process. Yet triggering something is not the same as engaging with it. For deeper, relational healing – change at the level of implicit self-processes – the ‘you’ remains irreplaceable. A technology-focused review identifies this limitation from within when it notes that ’the question remains open as to whether AI can adequately reproduce the quality of the interpersonal therapeutic relationship’ (Eichenberg et al., 2026, p. 8).
The therapeutic triad clearly defines the roles: the AI is the tool, the therapist is the ‘you’, and the encounter takes place exclusively between people. The AI is a medium comparable to therapeutic music or other creative tools. For example, the classic ‘empty chair’ technique can now also be implemented in virtual reality (Yoo & Kim, 2025). In one’s own practice, generative tools can also be used to visualise inner images, dreams or feelings and to work through them by viewing them together . One difference from a self-created image remains: the person who draws carries out the expression physically. The pressure of the pencil, the hesitation before applying a colour, the discarding and starting afresh are part of the experience and thus material for awareness (cf. Chapter 3.1). The generated image, by contrast, is a product without its own act of creation: calculated from descriptions and prior information, not arising in a lived moment (see Chapter 3.3). The crucial point here remains that the medium serves the encounter between people; it is not itself the other.
The empty chair used in Gestalt therapy—already introduced in Chapter 4.2—is, in essence, an ’it’: no one claims to be encountering the client. And yet, genuine feelings and genuine shifts arise in relation to it , because the client projects a relational ‘other’ onto it and because the therapist maintains the relational field around the chair work. Anyone who is moved in such a moment is in inner contact with themselves, not with the medium. A voice from the floor at the DVG conference described this precisely in this way for the chatbot session (Bauer, 2026, p. 86). In precisely this sense, AI is a sophisticated chair. The difference lies in how the projection is handled: the chair invites it without catering to it. The AI caters to it through its structural tendency towards validation and by simulating a counterpart.
This active fulfilment, the pretence of an interlocutor, is the ‘digital animism’ discussed in Chapter 4.2 (Fuchs, 2026a).
The comparison with the chair suggests an objection. A chair remains silent; it waits; it does nothing of its own accord. AI, on the other hand, responds; it sometimes contradicts; it produces turns of phrase that nobody anticipated. Something that behaves so autonomously seems to be more than just a piece of furniture. In his Tanner Lecture, Yuval Noah Harari makes precisely this objection the central thesis. He opens with the statement that the most important thing about AI is that it is not a tool:
“AI is not a tool. It’s not a tool in our hands. It is an agent with its own hands” (Harari, 2026, min. 01:00–01:12; trans. M.P.: “AI is not a tool. It is not a tool in our hands. It is an agent with its own hands.”).
The objection is valid, and must be acknowledged. A language model is not a coffee machine. Those who use it are not holding an instrument whose behaviour they understand.
Harari himself provides the answer. He explicitly states that agency does not require consciousness:
“You don’t need consciousness to be an agent. What you do need is the ability to make decisions by yourself” (Harari, 2026, mins 01:19–01:33; trans. M.P.: “You don’t need consciousness to be an agent. What you do need is the ability to make decisions by yourself.”).
His concept of an agent describes agency, not subjectivity: it describes what a system does, not what it is. This does not contradict the thesis of this thesis, as both address different questions. Harari asks what impact AI has on society. This thesis asks whether it can be an interlocutor. Both can be true at the same time: an AI with enormous agency that is nevertheless not a ’you’. Harari’s statement can also be reversed: if a system can act without possessing consciousness, then its actions reveal nothing about consciousness either. Everyday perception, however, draws this conclusion nonetheless: the system responds, it surprises, it cannot be fully controlled, so there must be someone there. The illusion of an other thrives on precisely this fallacy.
The concept of the ‘tool’ in this work is therefore not a technical one, but a relational one. It is not measured by how complex a system is, but by how humans relate to it. People relate to AI as they would to an ‘it’. Not because it is transparent, but because they are not faced with a subject with whom they could engage. What makes an ’it’ an ’it’ is not its predictability, but the absence of a body, experience and mortality (see Chapter 4.1). A tool that cannot be controlled is a risk, but that does not make it a counterpart.
Responsibility therefore remains entirely with the therapist. Responsibility presupposes someone who responds. It cannot be delegated to an ’It’.
The limit of the tool does not lie in its capabilities, but where words end.
‘Everything made of words will be taken over by AI’ (Harari, 2026, mins 31:36–33:10; trans. M.P.: ‘Everything made of words will be taken over by AI.’).
For Harari, the future of humanity depends on how much weight is given to that which cannot be put into words. He leaves open what that is. From the perspective of this work, it is the body: what happens between two people takes place not only in what is said, but in what is felt, in the resonance, in that for which neither has words (Chapters 2.3, 3.1). A machine can take over words. What happens between the words, it cannot.
That is the extent of the tool’s capabilities. In the experience of many users, artificial intelligence is perceived not as a tool, but as a counterpart. The following section examines which aspects shape this experience.