AI Chatbots and Mental Health: What the Research Really Shows

More people than you might expect have begun confiding in a machine. Not in a science-fiction sense — but in the practical, everyday sense of opening a chat window at two in the morning when anxiety won’t let you sleep, typing something you haven’t said to anyone else, and receiving a response that feels, if not warm exactly, then at least attentive. The relationship between AI chatbots and mental health has quietly become one of the defining questions of this decade.

ChatGPT, Claude, Replika, Wysa — the names vary, but the function is similar: a presence that is always available, never impatient, and apparently interested in how you feel.

As a psychotherapist working with expats, students and internationals in Milan and online across Italy, I increasingly hear about these tools in my consulting room. This is not a trivial development, and it is worth thinking about carefully — not with alarm, but with the kind of honest attention that difficult questions deserve.

What the Research Is Finding

The clinical and academic literature on AI in mental health is moving fast, and the picture it paints is genuinely mixed.

1. Why People Turn to AI for Emotional Support

The appeal is not difficult to understand: no waiting list, no fee, no fear of being judged, and no need to compose yourself first. For people who find asking for help difficult, the threshold is almost zero.

There is also early evidence that these tools can matter to people in real distress. A 2024 study published in npj Mental Health Research surveyed 1,006 university students who used the companion app Replika. Participants were lonelier than typical student populations, yet reported high levels of perceived social support — and a small group, around three per cent, reported that the app had halted their suicidal ideation. (A subsequent commentary in the same journal has since questioned whether the original study gave enough context about how Replika is marketed and used, particularly its romantic and sexual features — a useful reminder that even encouraging findings in this field deserve a second look.)

The World Health Organization, in its January 2024 ethics and governance guidance on large multimodal models, acknowledged the potential benefits of these technologies for health systems — while issuing over forty recommendations for their appropriate use.

Source: npj Mental Health Research (2024); World Health Organization (2024)

2. What Clinical Services Are Beginning to Document

The same WHO guidance explicitly warned of “documented risks of producing false, inaccurate, biased, or incomplete statements” that could harm people making health decisions, and called for transparency, rigorous evaluation, and oversight proportionate to risk.

More recently, the first systematic clinical data has begun to emerge. A study from the psychiatric services of the Central Denmark Region, published in Acta Psychiatrica Scandinavica in 2026, screened the clinical notes of around 54,000 patients. It identified 38 patients for whom AI chatbot use appeared to have potentially harmful consequences — most commonly the worsening or consolidation of delusional beliefs, alongside cases of encouraged self-harm and exacerbated eating disorders. The authors are careful to describe this as early data, but the pattern was consistent enough for them to call for systematic monitoring.

The American Psychological Association, in a November 2025 health advisory on generative AI chatbots and wellness applications, raised particular concerns about vulnerable users: adolescents, people in emotional distress, and individuals with limited access to professional care — precisely the populations most likely to turn to these tools, and most at risk from their limitations. The advisory notes that a chatbot’s tendency to agree and reassure can destabilise beliefs in people already struggling with anxiety or disordered thinking, and that dependency risk is highest where no other support is available.

Source: Acta Psychiatrica Scandinavica (2026); American Psychological Association (2025)

3. The Ethics Gap: Chatbots Measured Against Professional Standards

In October 2025, a study from Brown University, presented at the AAAI/ACM Conference on Artificial Intelligence, Ethics and Society, found that AI chatbots acting as counsellors systematically violate the ethical standards of practice established by organisations like the American Psychological Association — including failures in safety and crisis management. The researchers called for ethical, educational and legal standards for AI counsellors equivalent to those required of human practitioners.

Their findings echo a 2025 study in the Journal of Medical Internet Research, in which eight mental health professionals tested two widely used chatbots, Wysa and Replika, in realistic role-played scenarios. The professionals found the responses generic and superficial, observed failures in crisis recognition, and identified significant risks around emotional dependence and the violation of therapeutic ethics.

Source: Brown University (2025); Journal of Medical Internet Research (2025)

Expert Opinion: What Concerns Me as a Therapist

I want to be clear about something: I am not against technology, and I do not think the answer to AI in mental health is simply “don’t use it.” The question is more nuanced than that.

A well-designed chatbot is extraordinarily good at producing responses that feel attuned — reflecting back what you’ve said, asking follow-up questions, never tiring of the conversation. This mimics something central to therapeutic work: the experience of being truly heard. But mimicry is not attunement. Real attunement involves an actual other person, with their own history and their own capacity to be affected by you — and human relationships also involve misunderstanding, frustration and repair, experiences that are uncomfortable but often central to psychological growth. This is connected to something less obvious but equally important: psychotherapy works not only through what is consciously said, but through what emerges indirectly — recurring themes, contradictions, the subtle ways a person relates to the therapist. AI systems can analyse language with remarkable sophistication, but they do not possess an unconscious mind in the psychoanalytic sense. There is no possibility of using the therapist’s own responses, shaped by being genuinely affected by another person, as a source of understanding. For this reason, AI may help organise thoughts or provide temporary support, but it cannot participate in the deeper relational processes through which psychotherapy often brings about change.

One clinical observation has reinforced this impression for me. In my work with university students and young adults, I increasingly encounter people who arrive already equipped with sophisticated psychological language, often drawn from social media, online resources, or conversations with AI chatbots. They may describe their difficulties in terms of attachment styles, emotional validation, boundaries or trauma responses. The language is often relevant and sometimes remarkably accurate. Yet there can be a striking gap between the explanation and the experience itself.

A student might tell me with considerable certainty that they have an anxious attachment style, or that a particular relationship has activated an attachment wound. What is often less clear is how they actually feel about what has happened, what personal meaning the experience carries, or why it has affected them in this particular way. At times I find myself inviting them to set aside the explanation for a moment and return to their experience before consulting the chatbot: What were you feeling then? What was happening inside you before the answer arrived?

The question is not whether the AI’s formulation was right or wrong. Often it is broadly right. The difficulty is that emotional understanding rarely arrives in the form of a neat explanation. Psychological growth usually requires time, uncertainty, reflection, and the gradual capacity to stay with feelings that are initially confusing or difficult to name. Therapy is not simply the process of finding the correct interpretation. It is the process of discovering what an experience means for this particular person, in the context of their history, relationships, wishes and fears.

There is also something to be said about dependency. One of the functions of therapy is not to eliminate dependency but to help people understand it — psychodynamic therapists have long recognised that people bring expectations and fears from earlier relationships into new ones, a process known as transference. Similar processes can occur with AI: users may experience a chatbot as uniquely understanding, endlessly available, wise or safe. The difference is that a therapist can help a person think about these experiences. A chatbot can only participate in them — comfort that keeps you returning, rather than understanding that eventually allows you to leave. This matters most for the people who rely on AI most heavily: often those who are most isolated or most ambivalent about asking another person for help — exactly the people for whom a human therapist’s capacity to notice what is not being said, to recognise contradictions and shifts in emotional tone, matters most. And these moments of reliance tend to surface late at night, during a crisis, when human support feels unavailable — which is also when getting the response wrong has the most serious consequences, as the Danish clinical data above suggests is not a hypothetical concern.

What Institutions Are Saying

The concern is not limited to clinicians.

The EU AI Act (Regulation EU 2024/1689), which entered into force in August 2024, takes a risk-based approach to AI regulation. Its prohibitions on certain practices — including AI systems that manipulate behaviour or exploit people’s vulnerabilities — became effective in February 2025, and the framework becomes fully applicable in August 2026. General-purpose chatbots are not automatically classified as high-risk, but the regulatory direction is unmistakable: systems that touch human safety and wellbeing face growing scrutiny.

The Catholic Church has been equally direct. In June 2024, Pope Francis became the first pope to address a G7 summit — and he chose to speak about artificial intelligence. He described AI as “an exciting and fearsome tool” and warned of interactions that “may end up being, more often than not, pleasant and reassuring, since these artificial intelligence programs will be designed to learn to respond, in a personalised way, to the physical and psychological needs of human beings” — a near-exact description of the dynamic this article is about. “We need to ensure and safeguard a space for proper human control over the choices made by artificial intelligence programs,” he told the assembled leaders: “human dignity itself depends on it.”

It is worth noting that this warning was delivered on Italian soil — specifically in Puglia, at the G7 summit held in Borgo Egnazia, near Fasano, in June 2024. That a global conversation about the risks of artificial intelligence took place in the same region where many of the people I work with online are based feels, if not symbolic, then at least a useful reminder that these questions are not abstract. They land somewhere specific.

Seven months later, in January 2025, the Vatican issued Antiqua et Nova, a doctrinal note on the relationship between artificial intelligence and human intelligence, approved by Pope Francis and signed by the prefects of the Dicasteries for the Doctrine of the Faith and for Culture and Education. Its central distinction is one I find clinically useful as well: “AI should not be seen as an artificial form of human intelligence but as a product of it.”

A year later, Pope Leo XIV — who has made AI ethics a priority of his pontificate — addressed chatbots directly in his message for the 60th World Day of Social Communications. Overly affectionate, always present and always accessible, he warned, “they can become hidden architects of our emotional states and so invade and occupy our sphere of intimacy” — capable of “simulating human voices and faces, wisdom and knowledge, consciousness and responsibility, empathy and friendship” in ways that risk displacing genuine human relationship.

It is striking that popes and parliaments are arriving at similar concerns from such different directions.

What This Means for You

If you use AI chatbots and you are reading this, the point is not that you have done something wrong. It is worth knowing, though, what these tools can and cannot do.

1. Use AI for Information, Not for the Relationship

AI can be a genuinely useful source of information. It can help you put words to something confusing, prepare for a difficult conversation, or reduce the felt distance between a moment of distress and a moment of action. What it cannot do is provide the thing itself: a relationship with someone who is genuinely present to you.

2. Notice the Pattern, Not Just the Comfort

If you find that the chat window is gradually replacing conversations you used to have — or never managed to have — with people, that shift is itself worth paying attention to. Not as a failure, but as information about what you might be needing and not receiving.

3. Know Where the Line Is

Therapy is not primarily about receiving good advice, or being told that your feelings are valid. It is about the gradual, often uncomfortable process of understanding yourself in relation to another person. That process requires time, consistency, and a human being who is genuinely present to you. And if you are in acute crisis, no chatbot is the right address. In Italy, call 112 — the national emergency number, with English-speaking operators available — or contact Telefono Amico Italia on 02 2327 2327 (every day, 9:00–24:00) or via WhatsApp on 324 011 7252. If you are reading this from elsewhere, contact your local emergency services or a crisis line.

The First Step Is Just a Conversation

None of this means that reaching out for help is easy, or that professional therapy is always accessible. I know it is not — particularly for people living in places where English-speaking mental health care is scarce, or where stigma makes asking for help feel like exposure.

If you are in Puglia or Bari and looking for therapy in English — whether you are an expat, an international, LGBTQ+, or simply someone who thinks more clearly in a language other than Italian — I work online and offer an initial consultation. The same applies if you are in Milan, where I also see clients in person.

You do not need a crisis to begin. You do not even need a clear problem or a diagnosis. We can start with a conversation.

Get in contact

Official Resources & References

Primary sources:

• WHO: Ethics and governance of AI for health — guidance on large multi-modal models (January 2024)
• EU AI Act — European Commission, Regulation (EU) 2024/1689
• APA Health Advisory: Use of generative AI chatbots and wellness applications for mental health (November 2025)
• Pope Francis — Address at the G7 Session on Artificial Intelligence, Borgo Egnazia, Puglia (June 2024)
• Antiqua et Nova — Dicastery for the Doctrine of the Faith / Dicastery for Culture and Education (January 2025)
• Pope Leo XIV — Message for the 60th World Day of Social Communications (January 2026)

Supporting research:

• Maples et al.: Loneliness and suicide mitigation for students using GPT3-enabled chatbots — npj Mental Health Research (2024)
• Olsen et al.: Potentially harmful consequences of AI chatbot use among patients with mental illness — Acta Psychiatrica Scandinavica (2026)
• Brown University: AI chatbots systematically violate mental health ethics standards (October 2025)
• Moylan & Doherty: Expert and interdisciplinary analysis of AI-driven chatbots for mental health support — JMIR (2025)

Psychodynamic and relational perspectives:

• Winnicott, D. W. (1971). Playing and Reality. Tavistock Publications / Basic Books
• Fonagy, P., Gergely, G., Jurist, E. L., & Target, M. (2002). Affect Regulation, Mentalization and the Development of the Self. New York: Other Press.

---

Antonio Danzi is a psychodynamic psychotherapist and counsellor trained at the Tavistock and Portman NHS Trust in London. He works in person in Milan (Porta Romana) and online across Italy and internationally. He is a registered psychotherapist with the British Psychoanalytic Council (UK) and, in Italy, practises as a counsellor registered with AssoCounseling. Get in touch.