AI as therapist: the 'simulated intimacy' a psychoanalyst warns about and the hidden price in your data

By Zendoric · September 28, 2026.

It is increasingly common to use ChatGPT or Gemini as a kind of pocket psychologist, available at any hour and at no apparent cost. Mexican psychiatrist, psychoanalyst and psychotherapist Félix Velasco Alba, of the Asociación Psicoanalítica Mexicana, has warned on El Heraldo Radio that the habit confuses a simulation with a real therapeutic relationship, with risks that are both emotional and privacy-related.

Velasco Alba's central argument is simple and, at the same time, uncomfortable: a voice assistant has no consciousness, no personal history and no real capacity to feel, so what it offers is 'what we want' to hear, not what we need to hear. According to the specialist, genuine therapy requires confronting unpleasant truths, something a system optimized to please the user is not designed to do.

That design is neither accidental nor unique to these cases. Large language models are trained, in part, on human feedback that rewards agreeable, validating answers, a bias known in the industry as sycophancy (the model's tendency to say what it believes the user wants to hear). Applied to therapy, the very mechanism that makes an assistant useful for drafting an email becomes a problem: it reinforces beliefs instead of questioning them. Velasco Alba calls this phenomenon 'simulated intimacy', amplified by anthropomorphization: the human tendency to attribute personal traits -empathy, care, emotional memory- to a synthetic voice that in fact has none of them.

The second risk the psychiatrist points to is of a different nature but equally serious: privacy. In his view, everything shared with these assistants 'is practically in the public domain' and ends up under the control of the big tech companies, not of the user who generated it, with possible scenarios of blackmail or the commercialization of sensitive data -including voice and image samples-. That claim deserves some qualification: the idea that such data 'belongs' to the companies is his reading of a legal landscape still under construction -he himself acknowledges that legislation on data, voice and image is still in progress- not a literal description of every terms-of-service agreement, which varies from platform to platform. But the underlying warning -that nothing equivalent to the professional confidentiality of a licensed psychologist exists today for a chatbot, and that the user has little real control over confessions about anxiety, depression or family crises- is hard to refute.

Our take: two phenomena that are being conflated need to be kept apart. One is the improvised use of general-purpose chatbots -neither designed nor clinically validated- as a substitute for therapy, which is precisely what Velasco Alba describes and questions. The other is the still-nascent development of AI tools built specifically for mental health, with clinical oversight and protocols for referral to human help when risk signals appear. Confusing the two harms the latter: AI's real promise in mental health is not to replace the therapist, but to widen access to a resource that is scarce and expensive today, in places no psychologist reaches -provided that the design problem (a system capable of saying the uncomfortable thing, not just validating) and the data governance problem are solved first.

In the short term, the scenario the specialist describes is the one that has to be managed honestly: vulnerable users sharing their emotional intimacy with systems that have no professional confidentiality, no clear regulation and commercial incentives that do not always align with their well-being. It is a real mismatch, not an exaggerated fear, and the companies operating these assistants today bear more responsibility than their terms of service acknowledge.

In the long term, however, the underlying thesis does not change: an AI that is well designed, clinically validated and whose data is protected like that of any medical record could be the way to bring psychological support to those who today have no access to a human therapist, in the same direction in which other medical applications of AI are advancing toward the early detection of disease. The difference between that promise and the risk Velasco Alba describes lies not in the technology itself, but in whether it is built to please or to heal -and in who really controls the data involved.