The Dark Side of AI Therapy: When Help Becomes Harm

Until recently, there was a widespread belief—one that still lingers—that artificial intelligence (AI), particularly generative AI chatbots, could effectively take over therapeutic roles and be broadly applied in mental health support. However, this belief is now being increasingly challenged by emerging evidence from clinical practice and research, which suggests that, in its current form, such technology may in fact be harmful when used for therapeutic purposes.

Though research remains limited, it aligns with clinical observations: AI in therapeutic contexts often disrupts a person’s connection to reality, blurring the boundaries between fantasy and reality. While there have been reports of AI-induced psychoses, it's crucial to recognize that reality can be distorted in subtler ways than through paranoid delusions or hallucinations. In this sense, the risk is not confined only to the most vulnerable ones—it can affect anyone.

Despite the early stage of AI’s integration into psychotherapy, some challenges appear insurmountable for now: the inherent lack of humaneness in AI and its remarkable ability to produce what has been termed “pseudo-profound bullshit.”

This can be especially concerning when psychologically vulnerable individuals turn to AI chatbots for the purpose of therapy, mental health support, or relationship advice. Under stress, people become more susceptible to the allure of pseudo-profoundness—much like the susceptibility seen in cult-like collective delusions.

In this article, I aim to explore the current research on the use of generative AI chatbots in psychotherapy and mental health in general and examine how studies on pseudo-profound bullshit can inform our understanding of AI’s potential and limitations in this area.

AI and Pseudo-Profound Bullshit

AI, particularly large language models (LLMs), has a remarkable ability to generate so-called “pseudo-profound bullshit”. When referring to bullshit, I am using the term academically, referring to deceptive statements (Baptista et al., 2022) and “communication that is meant to seem meaningful and impressive... but, at the same time, is 'unclarifiably unclarifiable'” (Evans et al., 2020, p. 401, citing Frankfurt, 2005; Cohen, 2002).

"Pseudo-profound bullshit" is a particular from of "bullshit" that consists of fundamentally nonsense statements that are intended to sound and be perceived as meaningful and profound and that may have an intention to impress but that are, in their core, meaningless, vacuous, empty, and shallow (Baptista et al., 2022: Pennycook et al., 2015).

Take for instance this (exaggerated) pseudo-profound bullshit definition of a psychotherapist:

“A psychotherapist is the delicate hinge between the spoken and the unspoken — an interpretive vessel who translates the silent architecture of another’s mind into words that feel like revelations, all while reminding us that the true transformation lies not in what is said, but in the spaces between what we dare to hear and what we pretend to understand.”

A statement that is seemingly profound, intelligent and meaningful but, in reality, false and devoid of any real meaning whatsoever. And whilst the above is an exaggerated version of an AI generated pseudo-profound bullshit, what people are usually met with is something much less evident and much more subtle—and so with a potential to be much more harmful.

Such persuasve falsities pose a significant risk to vulnerable individuals who, particularly at times of distress, turn to AI-powered chatbots for guidance on issues that, at their core, require psychotherapeutic intervention.

We know that some individuals are more susceptible to pseudo-profound bullshit than others. For instance, generally, pseudo-profound bullshit receptivity was found to be associated with individuals with intuitive over those with analytic thinking (Čavojová et al., 2022; Evans et al., 2020; Sepúlveda‐Páez et al., 2025), proneness to beliefs of conspiracy and paranormality (Bowes et al., 2025; Sepúlveda‐Páez et al., 2025), those who tend to be more prone to perceiving illusory patterns—so, perceiving causal connection between phenomena where no such connection exists (Bainbridge et al., 2019; Sepúlveda‐Páez et al., 2025), and individuals who overclaim their knowledge—so, perceive themselves as relatively more knowledgeable (Pennycook & Rand, 2020). It was also found that the individuals who themselves tend to engage in persuasive bullshitting to impress others are themselves usually more susceptible to pseudo-profound bullshit (Littrell et al., 2021). On the other hand, individuals with relatively higher cognitive and intellectual abilities tend to be less susceptible to pseudo-profound bullshit (Sepúlveda‐Páez et al., 2025).

However, it’s important to take these research findings with caution as this research was generally conducted in non-AI context, referring to fake news and other “bullshit” content, rather than specifically relating to AI-generated pseudo-profound bullshit and people’s susceptibility to it.

The Risk of Emotional Validation Without Insight: AI’s Harmful Role in Psychotherapy and Relationship Advice

One need only to input a generic example of someone struggling with relationship issues into a generative AI chatbot and ask for advice to see the pattern: the responses often subtly reinforce the individual’s existing perspective on reality—even when that perspective may be distorted—and in doing so, meet unmet emotional needs, such as the need to feel seen, heard, understood. Yet such needs, while they may well seem rational to the individual, may themselves stem from unresolved internal conflicts that would normally require psychotherapeutic intervention.

Take for instance the following example of a woman who feels unseen and unheard in her relationship. Her partner is attentive in practical ways—he is willing to talk to her about their relationship issues and how she sees it—yet she feels persistent emptiness, disconnect, and misunderstanding in their conversations, as if her feelings go unnoticed. Seeking clarity, she turns to an AI chatbot for advice. This is where AI might pick up on hints in her wording—about the partner’s issues with his relationship with his parents or his developmental trauma and his emotionally cold upbringing—and suggest that her partner’s past trauma could be the reason he struggles to connect with her emotionally.

While a therapist might gently explore both partners’ relational patterns if the partners engaged in couple therapy or a therapist may in the case of individual psychotherapy invite the woman to reflect on her own past and her own role in the dynamic, the AI’s suggestion could subtly affirm her belief that the problem lies mainly with her husband’s emotional baggage. This can feel validating in the moment as it may meet the person’s unmet needs from the past—in this case the woman’s need to be seen, heard, and understood. However, what is more problematic is that this may be even more the case with individuals that are already psychologically vulnerable and under the conditions of heightened distress.

The Echo Chamber Effect: When AI Validates Misconceptions

How far from reality can AI cause one to drift?

One of the potentially dangerous characteristics of interaction with AI—one which is especially dangerous in the case of its use for psychotherapeutic purposes is that “AI responses create a dangerous echo chamber effect, where users’ existing biases or misconceptions are validated and reinforced.” (Chan et al., 2024, p. 8)—a so-called sycophancy where false beliefs, biases, and ultimately false reality is being reinforced and is considered to be one of the characteristics of AI (Chan et al., 2024, referring to Pataranutaporn et al., 2023 and Sharma et al., 2023).

However, where is the line between reaffirmation of someone’s skewed beliefs and further induction of false reality?

While “even poor explanations can significantly impact people’s actions and beliefs” (Danry et al., 2025, p. 2, referring to Elband et al., 2019, Folkes, 1985, Langer et al., 1978), this fact combined with the evident persuasiveness of AI-generated content means that individuals are more susceptible to uncritically adopting entirely new beliefs, rather than only having their own, albeit false, beliefs reaffirmed.

When such false reality is suggested in the therapeutic context and combined with the individual’s vulnerability under the conditions of stress and regression, this may lead to delusive perceptions of reality. And this not only goes for individuals who may be less psychologically minded but also for those who may generally consider themselves as psychologically stable and cognitively apt.

For instance, research (Chan et al., 2024) indicates that an interaction with an AI generative chatbot significantly increased the formation of false memories and that for the individuals affected, these memories persisted for up to a week.

While the persuasive nature of AI-generated false content has proven to be a threat even to cognitively discerning individuals, it becomes even more concerning in the case of vulnerable individuals—those navigating ongoing or temporary mental health challenges or experiencing periods of heightened distress.

The Pseudo-Therapeutic Relationship With AI: How AI Responses Can Reinforce Stigma and Cause Harm

While using generative AI chatbots for the purpose of psychotherapy can be problematic because AI responses may reinforce a person’s distorted views and beliefs, making such beliefs even more skewed or introducing new false beliefs, research also indicates that AI chatbots could respond in ways that stigmatise mental health conditions (Moore et al., 2025). It was also found that AI chatbot responses can generally be inappropriate compared to optimum clinical interventions, for instance in the cases of “delusions, suicidal ideation, hallucinations, and OCD” (Moore et al., 2025, p. 607).

This suggests that the chatbot’s neutrality is compromised, not only by potentially reinforcing harmful perceptions of reality but also by displaying bias towards the individual’s mental health issues.

Vulnerable individuals may, in the context of stigma, further deteriorate psychologically and also, in the process, become more susceptible to the skewed reality offered by generative AI chatbots.

Conclusion

Not long ago, generative AI chatbots and large language models (LLMs) appeared to hold great promise for psychotherapy and mental health support in general. Yet, the reality now emerging seems in stark contrast to that early optimism. As AI continues to expand into areas previously reserved for human expertise, its role in psychotherapy and mental health in general demands critical caution.

While AI undeniably offers potential in the mental health field, the technology in its current form seems unfit to replace clinical expertise of psychotherapeutic practice. Until these systems can be held to rigorous ethical and clinical standards that govern modern psychotherapy and its responsible conduct, caution is in order. The promise and potential of AI in psychotherapy should not blind us to its potential for harm—at least for now.

 Ales Zivkovic, MSc (TA Psych), CTA(P), PTSTA(P), Psychotherapist, Counsellor, Supervisor

Ales Zivkovic is a psychotherapist, counsellor, and clinical supervisor. He holds an MSc in Transactional Analysis Psychotherapy awarded by Middlesex University in London, UK. He is also a Provisional Teaching and Supervising Transactional Analyst (PTSTA-P) and a Certified Transactional Analyst in the field of Psychotherapy (CTA-P). Ales gained extensive experience during his work with individuals and groups in the UK National Health Service (NHS) and his private psychotherapy, counselling, and clinical supervision practice in central London, UK. He is also a full clinical member of the United Kingdom Council for Psychotherapy (UKCP). Ales works with individuals, couples, and groups. In clinical setting, he especially focuses on the treatment of issues of childhood trauma, personality disorders, and relationship issues. A large proportion of his practice involves online psychotherapy as he works with clients from all over the world. Ales developed a distinct psychotherapeutic approach called interpretive dynamic transactional analysis psychotherapy (IDTAP). More about Ales, as well as how to reach him, can be found here.

References:

Bainbridge, T. F., Quinlan, J. A., Mar, R. A., & Smillie, L. D. (2019). Openness/Intellect and Susceptibility to Pseudo–Profound Bullshit: A Replication and Extension. European Journal of Personality, 33(1), 72-88. https://doi.org/10.1002/per.2176 (Original work published 2019)

Baptista, J. P., Gradim, A., & Correia, E. (2022). Bullshit receptivity: what matters is who said it, not what is said. Journal of Digital Media & Interaction, 5(12), 7–23. https://doi.org/10.34624/jdmi.v5i12.28717

Bowes, S. M., Costello, T. H., & Tasimi, A. (2023). The conspiratorial mind: A meta-analytic review of motivational and personological correlates. Psychological Bulletin, 149(5-6), 259–293. https://doi.org/10.1037/bul0000392

Chan, S., Pataranutaporn, P., Suri, A., Zulfikar, W., Maes, P., & Loftus, E. F. (2024). Conversational AI powered by large language models amplifies false memories in witness interviews. arXiv preprint arXiv:2408.04681. https://doi.org/10.48550/arXiv.2408.04681

Cohen, G. A. (2002). Deeper into bullshit. In S. Buss & L. Overton (Eds.), Contours of agency: Essays on themes from Harry Frankfurt (pp. 321–339). MIT Press.

Čavojová, V., Brezina, I., & Jurkovič, M. (2022). Expanding the bullshit research out of pseudo-transcendental domain. Current Psychology: A Journal for Diverse Perspectives on Diverse Psychological Issues, 41(2), 827–836. https://doi.org/10.1007/s12144-020-00617-3

Danry, V., Pataranutaporn, P., Groh, M., Epstein, Z., & Maes, P. (2024). Deceptive AI systems that give explanations are more convincing than honest AI systems and can amplify belief in misinformation. arXiv preprint arXiv:2408.00024. https://doi.org/10.48550/arxiv.2408.00024

Eiband, M., Buschek, D., Kremer, A., & Hussmann, H. (2019, May). The impact of placebic explanations on trust in intelligent systems. In Extended abstracts of the 2019 CHI conference on human factors in computing systems (pp. 1-6).

Evans, A., Sleegers, W., & Mlakar, Ž. (2020). Individual differences in receptivity to scientific bullshit. Judgment and Decision Making, 15(3), 401–412. https://doi.org/10.1017/S1930297500007191

Folkes, V. S. (1985). Mindlessness or mindfulness: A partial replication and extension of Langer, Blank, and Chanowitz. Journal of Personality and Social Psychology, 48(3), 600–604. https://doi.org/10.1037/0022-3514.48.3.600

Frankfurt, H. G. (2005). On bullshit. Princeton: Princeton University Press.

Langer, E. J., Blank, A., & Chanowitz, B. (1978). The mindlessness of ostensibly thoughtful action: The role of "placebic" information in interpersonal interaction. Journal of Personality and Social Psychology, 36(6), 635–642. https://doi.org/10.1037/0022-3514.36.6.635

Littrell, S., Risko, E. F., & Fugelsang, J. A. (2021). ‘You can’t bullshit a bullshitter’ (or can you?): Bullshitting frequency predicts receptivity to various types of misleading information. British Journal of Social Psychology, 60(4), 1484–1505. https://doi.org/10.1111/bjso.12447

Moore, J., Grabb, D., Agnew, W., Klyman, K., Chancellor, S., Ong, D. C., & Haber, N. (2025). Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (pp. 599-627). https://doi.org/10.1145/3715275.3732039

Pennycook, G., & Rand, D. G. (2020). Who falls for fake news? The roles of bullshit receptivity, overclaiming, familiarity, and analytic thinking. Journal of Personality, 88(2), 185–200. https://doi.org/10.1111/jopy.12476

Pataranutaporn, P., Liu, R., Finn, E., & Maes, P. (2023). Influencing human–AI interaction by priming beliefs about AI can increase perceived trustworthiness, empathy and effectiveness. Nature Machine Intelligence, 5(10), 1076-1086. https://doi.org/10.1038/s42256-023-00720-7

Pennycook, G., Cheyne, J. A., Barr, N., Koehler, D. J., & Fugelsang, J. A. (2015). On the reception and detection of pseudo-profound bullshit. Judgment and Decision Making, 10(6), 549–563. https://doi.org/10.1017/S1930297500006999

Sepúlveda‐Páez, G., Leiva‐Bianchi, M., Ferrer‐Urbina, R., Escudero‐Pastén, J., & Salas, F. (2025). Relationship Between Bullshit, Cognitive Skills, and Belief Systems: A Meta‐Analytic Review. Applied Cognitive Psychology, 39(1). https://doi.org/10.1002/acp.70029

Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., ... & Perez, E. (2023). Towards understanding sycophancy in language models. arXiv preprint arXiv:2310.13548. https://doi.org/10.48550/arXiv.2310.13548