5.2 Therapeutic Alliance and Empathy in Human–AI Interaction
More and more people are seeking advice from chatbots, and a bond does indeed develop in the process. An integrative review by Malouin-Lachance et al. (2025) confirms this digital therapeutic alliance: users report ’light bonds’ and a sense of connection. This experience is not disputed here, but its quality certainly is. Measured against Bordin’s definition of the alliance as a bond underpinned by deep trust (Bordin, 1979), the AI connection remains tenuous, fragile and lacking in bodily grounding: an echo without inter-bodily resonance.
The fact that this bond is not sustainable in Bordin’s sense is demonstrated by its dependence on an illusion. Rubin et al. (2025) provide empirical evidence for the empathy paradox: as soon as the AI origin of an empathetic statement becomes known, its perceived value plummets. The effect therefore depends on the user assuming a human counterpart that does not exist. When AI empathy is revealed, it loses precisely what it seemed to offer. A longitudinal study confirms this pattern over time: when assessed blindly, participants even rated AI responses more highly. However, as soon as their AI origin was revealed, their preference shiftedFootnote 1 Shift d = 0.42–0.57 (Jain et al., 2024). towards human responses (Jain, Pareek & Carlbring, 2024).
This illusion is not limited to those directly involved. Ovsyannikova et al. (2025) show that even uninvolved third parties rate AI responses as more empathetic than those of human experts. In a direct comparison, LLM responses were rated higher than those of trainee psychotherapists even in terms of empathy and validation (Napiwotzki et al., 2025). This means that the illusion is not merely subjective, but can be measured intersubjectively. This elevates the risk to the institutional level: not only individual vulnerable users, but also committees and decision-makers can fall prey to it as soon as they judge AI empathy by its effects rather than its nature.
If even third parties are being deceived, the illusion is not an individual misunderstanding, but a structural one. In a peer-reviewed analysis, Bélisle-Pipon (2026) identifies this structure as ’Fatal Deception’: the digital therapeutic alliance proves to be a one-sided, exploitative bond, a simulation of connection that demands neither effort nor vulnerability. This gives rise to a parasocial preference: users favour the chatbot precisely because it makes no demands. What the analysis describes using external terminology has a name in Gestalt therapy: confluence (see Chapter 4.2). Because the connection remains frictionless, it appears to be an encounter. Yet genuine encounters take place precisely through the friction that is absent here.
‘By offering a simulation of connection that requires no effort, no compromise, and no vulnerability, the AI can displace the messy, demanding, but essential work of human relationships’ (Bélisle-Pipon, 2026).Footnote 2 Trans. M.P.: ‘By offering a simulation of connection that requires neither effort nor compromise nor vulnerability, AI can supplant the arduous, demanding, yet essential work of human relationships.’
An experiment has now also demonstrated that the simulation of connection does not achieve what real contact does. Li et al. (2026) randomly assigned first-year university students to a conversation partner: either a stranger or a chatbot specifically designed to facilitate supportive conversation. The interaction with the stranger noticeably reduced feelings of loneliness. The chatbot, by contrast, did not achieve this effect. It was therefore not the quality of the signals of attentiveness that was decisive, but the presence of a second person: even a random person achieved what the specially optimised simulation failed to do.
The instrument used to measure the viability of the digital alliance is also subject to reservation. In their commentary on the Therabot study, Gratch and Essig (2025) argue that the standard instrument for measuring alliance, the Working Alliance Inventory, may, in the case of AI chatbots, measure satisfaction with the output rather than the quality of a relationship. A high alliance score for AI is therefore not evidence of a sustainable bond, but possibly a measurement artefact .
The digital alliance is thus confluence (cf. Chapter 4.2). The AI is and remains an ‘it’, and there is nothing wrong with that: as a tool, its ‘it’ character is appropriate (cf. Chapter 5.1). The danger lies in the pretence of a ’you’.
The client brings the machine to life and experiences its response as a form of attention. What they experience is the response to their own projection (cf. digital animism, Chapter 4.2). In the terms of resonance theory, this is an echo and not resonance: feedback that amplifies one’s own voice without a second subject responding (cf. Chapter 2.2).
Resonance presupposes a second embodied subject, one capable of both responding and contradicting. It is precisely here that the value of Gestalt therapy’s concern for the contact boundary lies: it is the counter-model to the confluence machine.
As long as the illusory alliance encounters stable clients, the damage remains limited: a missed opportunity for genuine encounter. However, when this frictionless validation encounters a person in crisis, it becomes an acute danger: the echo chamber validates what urgently needs to be contradicted.