The danger of mistaking AI mental health support for real therapy

By Medical Xpress · June 27, 2026.

The story, reported by Medical Xpress, addresses a growing debate in the field of digital health: the tendency to treat artificial intelligence tools designed to provide emotional or psychological support as if they were equivalent to professional psychotherapy. The headline is blunt —'the danger of confusing AI mental health support with therapy'— and reflects a concern that is taking up more and more space in medical journals and bioethics forums.

**Transparency note:** The full content of the article could not be accessed in this edition (possible restriction or loading failure). The following summary is based on the headline, the RSS description and widely documented industry context, duly flagged as such.

**The distinction that matters**

In general, there is a clear clinical consensus in the sector: AI-mediated emotional support —such as active-listening chatbots, guided mindfulness apps or conversational assistants like Woebot, Wysa or the companion modes of general language models— does not constitute psychotherapy. Psychotherapy is a regulated clinical process, carried out by a professional with specific credentials, that involves diagnosis, case formulation, therapeutic alliance and longitudinal follow-up under legal responsibility. AI systems, however sophisticated, do not meet that standard.

The documented problem is that the perceptual boundary for users is eroding. Modern conversational models are capable of responding with apparent empathy, recalling prior context and adapting their tone to the detected emotional state, which generates in some users a subjective sense of 'being understood' comparable to —or more comfortable than— what they report in in-person sessions.

**Concrete risks identified in the sector**

As industry context, the risks flagged by the scientific literature and regulatory bodies include several vectors:

1. *Delay in seeking professional help*: If a user perceives that their AI chatbot 'is already helping them', they may indefinitely postpone access to clinical treatment for conditions that require it urgently (major depression, bipolar disorder, suicidal ideation).

2. *Absence of risk assessment*: A human therapist is trained and legally obligated to assess suicide or homicide risk and to activate crisis intervention protocols. Most current AI systems cannot reliably fulfill this role, and in some documented cases they have failed gravely when faced with users in crisis.

3. *False validation*: Language models tend to be positive reinforcers by design (RLHF oriented toward user satisfaction). This can translate into the validation of distorted beliefs or harmful psychological dynamics, exactly the opposite of informed therapeutic confrontation.

4. *Privacy and sensitive data*: Mental health conversations are especially sensitive. Their storage, use for training or exposure to third parties presents risks that users rarely understand when interacting with a chat interface.

**The European regulatory framework**

In the context of the EU AI Act, already in force in its early phases, AI systems intended to influence people's mental health are potentially classified as *high risk*, which entails obligations of transparency, human oversight and technical documentation. However, the exact categorization of many 'emotional well-being' applications remains a regulatory gray area: if the app does not make an explicit diagnosis, it may try to escape the high-risk category, even if the practical effect on the user is similar.

As industry context, there is debate over whether the regulatory threshold should depend on the system's stated intent or on its actual use by the population.

**Implications for agentic AI**

From the perspective of agentic AI systems —those capable of taking initiative, recalling histories and executing chained actions—, the risk is amplified. A mental health agent that maintains persistent memory, proactively contacts the user at moments of detected stress and suggests intervention strategies operates de facto as a continuous therapeutic companion. Without robust ethical and clinical frameworks, and without a human in the loop, this type of agent can cause real harm even if it was designed with good intentions.

The debate over whether AI agents should have *mandatory referral capability* —that is, the capacity and the mandate to interrupt the interaction and connect the user with human crisis resources when warning signs are detected— is one of the most active points of discussion in the field.

**Perspective**

The Medical Xpress article adds to a necessary critical current: AI can be a valuable tool for first contact, psychoeducation, support between sessions or access in contexts where clinical resources are scarce. But that real usefulness must not be confused with therapeutic equivalence. The distinction is not semantic: it has direct consequences for what care vulnerable people receive and who assumes responsibility when something goes wrong.