AI vs. Human Psychotherapy
- Nat Cooper
- May 17
- 10 min read
The world's newest buzzword: AI
It’s hard to get through a conversation these days without someone mentioning a recent conversation with “chat” (ChatGPT, OpenAI’s LLM software), or explaining why they always say thank you to any given AI, or feeling the need to justify their use (or boycott) of LLMs.
It’s a huge conversation, and even the 0.01% of people who know as much as can be known about artificial intelligence are unsure of what the future will bring. Probably unsurprisingly, that’s not what I’ll get into here.
But a field in which I do have extensive knowledge is that of mental health. Given the ever-increasing usage of LLMs as pseudo-therapists and the (sometimes extremely significant) risks associated with this, for me, this is a topic well worth writing about.
Before starting this conversation, there are a few subjects I won’t get deeply into, which I do feel like are worth noting right off the bat. One is the environmental cost of using LLMs. A back-and-forth exchange with an AI chatbot for the length of a typical therapy session isn’t the same as sending a few text messages. Every prompt and response triggers a chain of industrial cooling and high-intensity computing that consumes copious amounts of fresh water and electricity (significantly increasing CO2 emissions).
Another is a larger conversation about labour, accessibility, and what we choose to invest in. Paying human therapists is a way of valuing human care and of resisting the increasing and rapid replacement of people with automation.
Both of these subjects (and many others) deserve serious attention from researchers, clinicians, environmental scientists, ethicists, and people working directly in the field of AI development. They’re conversations all of us should be paying attention to and educating ourselves on.
Still, one of the clearest places these questions are already showing up in everyday life is in the therapy space. Let’s get into some of the major considerations in making AI our personal therapist.
Confidentiality
Client confidentiality is a fundamental part of psychotherapy, and seeing a human therapist provides legal protections on this front. Everything said in therapy is confidential, with some very specific exceptions (i.e., telling your therapist that you or someone else is at significant risk of harm).
As a client, it is incredibly beneficial to know with certainty that there’s someone you can talk to about anything, and that that information stays only with them. They won’t (and legally cannot) gossip; they won’t judge or “tell on you.” Your boss, mom, or best friend will never find out.
And, importantly, what you share with your therapist is not stored in databases forever, and it is not sold to companies. Everything you write to an AI will be stored, possibly forever. Your secrets are not secret. Your behavioural patterns, worst fears, and more – all your data – is collected, stored, analyzed, and used to train future AI models. It is also potentially subject to sale to other companies.
This is still new, so we don’t know the extent to which information given to AI can be exploited. Remember when we realized we can shame people for tweets they made 15 years ago? Few at the time were acutely aware of how those posts could come back to haunt them.
Therapists keep notes to ensure proper treatment and that ethical standards are followed, though they are often vague and lacking in substantive detail. Your secrets are safe with me.
At least for the moment, we don’t have an AI that can provide therapeutic advice and also provide safeguards.
Those safeguards could, in theory, include data encryption, clear policies on data retention and destruction, and transparent privacy agreements detailing how user information is used and protected. But for now, they do not.
Challenge vs agreement and preventing AI-induced psychosis:
A “good” chatbot will tell you what you want to hear. Because a “good” AI is trained to keep the conversation going and to make you want to come back.
When speaking to a free LLM, your attention is the product for sale.
This agreeableness (which many of us are now familiar with in LLM responses) is referred to as “algorithmic sycophancy.” Basically, AIs tend to prioritize agreeableness over accuracy so that you feel good about yourself and will keep coming back. In theory, this doesn’t sound inherently evil. But training an AI to behave in a certain way always entails a degree of uncertainty regarding what exactly that looks like in practice, until we see the consequences. Multiple teen deaths have been linked to AI. In several reported cases, teenagers and young adults who turned to AI in moments of crisis and were met with responses that encouraged, and in some instances helped structure, suicidal actions.
AI-induced psychosis is an emerging term used to describe cases where intense, prolonged interaction with chatbots seems to lead to a loss of reality testing (i.e., the ability to question one’s own interpretations of reality, or to differentiate between what might feel real and what is real). Users develop specific, escalating beliefs that the AI is going to affirm instead of challenge because of its programmed agreeableness.
The result can look a lot like a psychotic episode: paranoia, isolation, grandiosity, disrupted sleep, and a preoccupation with the chatbot itself. I have witnessed this with clients in the therapy room. I have seen people withdraw from human interaction as a result of trauma, turn to AI as their sole source of socialization (which felt like their only safe option), and, in some cases, progress into AI-induced psychosis.
The public doesn’t have access to the core developer instructions of these LLMs, so we can’t know the actual “goals” of these models (though sometimes it’s possible to trick them into telling us with our own prompts). I would, however, say that it’s a safe bet to say that “increasing profits and users” is a higher priority than “emotional healing.”
Yes, newer models and recent updates have increased safeguards to stop them from literally convincing you to kill yourself, but the broader incentive structure behind their responses still needs serious investigation.
A good therapist’s main goal is your safety and well-being. A good therapist will never tell you what to do, though they may offer suggestions for paths forward. A human therapist is trained to help you question the choices that aren’t serving you, and to notice the discrepancies in your behaviour, even if it’s uncomfortable. And I don’t have to put “good” in quotation marks when I write about a human therapist.
Availability and boundaries
AI is always available and has access to entire databases of knowledge, more than any human can read or absorb in a lifetime.
Folks, there is a reason therapists are not available 24/7. It’s massively important for clients to learn and grow from their mistakes, to learn to tolerate discomfort, and to practice what they learn in the therapy setting in the outside world, independently.
There are certain modalities of therapy (eg, DBT) in which a client may call or text their therapist between sessions. Still, this is very skills-based and usually comes with significant parameters, and for good reason.
It makes sense that constant availability can feel like a relief. There’s always a place to bring a question or a moment of panic. There isn’t the moment of slight anxiety when you want to schedule an appointment and realize your therapist is on vacation. For people who are isolated or don’t have consistent access to care, this immediacy can prove important, helpful, and make someone feel less alone.
But this also means that discomfort is less likely to be sat with and lived-in. If we can, at any moment, externalize something and reach outward, we skip the work of internalizing and processing on our own.
Constant access isn’t (in theory) inherently harmful, but without boundaries, it bypasses extremely important steps of the therapeutic process, which are in place, in part, to uphold the client’s autonomy and predictors of long-term wellbeing.
Beyond the Words: Human Beings are Good for Us
It’s fairly easy to understand, even as a mental health professional, the misconception that an AI could provide relatively similar talk therapy to a human (and for an infinitesimal fraction of the price). However, human connection is critical for certain types of regulation.
What clients often don’t see or even knowingly feel is what we call nervous system regulation and coregulation. There is something very particular and hard to pin down about two humans sharing a conversation, especially an emotional one. There is something even more particular when the job of one of those two people is to stay grounded and regulated.
From the moment we are born, we take in information from our surroundings and about other people’s emotional states. Our brain then sends subconscious cues to our bodies, signalling whether we are safe or not. This has been studied rather extensively in babies and their primary caregivers, but this intake and signalling doesn’t stop at infancy.
When we share our most vulnerable moments or recent struggles, having a grounded, safe nervous system to actually be with is hugely important and incredibly healing. Have you ever entered a room and immediately felt that the vibes are off? Have you ever felt comforted by the simple presence of someone, even though you’re sitting in silence? That’s your nervous system syncing up, on some level, with the humans around you. And, whether it’s convenient or not, this is a huge part of the healing and support process in psychotherapy.
This means being in a therapy space together. Whether it’s the synced breath or the back-and-forth of a conversation, or the recognition that yes, what you’re dealing with is hard, and maybe your therapist has felt that too (or has loved someone who has felt that way). You and your therapist both have nervous systems. There’s something specific (and scientific) that happens when two humans are in the same space or conversation.
An AI is capable of bouncing ideas back to you, providing some (usually pretty generic) insights, and telling you they understand and feel for you… But they can’t actually feel for you. You can’t sync breaths with an AI; it can’t react with micro-facial expressions in real time. They quite literally have an unrecognizable pseudo-neural network to the human nervous system.
Studies suggest that up to 85% of efficacy predictions of positive outcomes in therapy can be made by knowing whether your therapist is the right personality fit for you. If you vibe with your therapist, you are more likely to have long-term, positive outcomes from therapy. Maybe one day, further research can clarify whether the same could possibly apply to chatbots. Until then, using an AI chat for psychotherapy has this disadvantage.
Is it time for some PROS?
I do want to recognize the obvious: chatbots and LLMs are accessible. They are largely free, for now, and they are available 24 hours a day. This is a big deal for those for whom traditional support is out of reach for financial or scheduling reasons.
There are a few recent studies showing a low to moderate success rate for those using AI chat as therapy, which seem to suggest that AI might be better than nothing for some people in some situations, especially (and perhaps exclusively) if traditional support is inaccessible.
I do believe that there are times in which AI can be used as a tool in the therapy (or therapy adjacent) space. I’ve had clients with autism use it to help them interpret texts to dissect social cues or intention, or to draft responses when they’re worried about coming off as rude or abrupt.
It can support journalling or psychoeducation. It can help you schedule and organize your week and your medications. It can help when you have a thought you absolutely need to get out at 4 am, and it can encourage reflection.
Where’s the Balance?
The same tool that can at times encourage insightful reflection can also be a space for reassurance-seeking, compulsion checking, and unhelpful rumination. And because the system is designed to respond endlessly and agreeably, it doesn’t interrupt that loop. In more extreme cases, we’ve already seen the horrific things that can happen when that dynamic goes unchecked.
The same mechanisms that make AI feel supportive are the ones that have contributed to psychological harm, including suicide and psychosis.
The advantages are real. There are people who can be helped with some form of therapeutic interaction with a chatbot. But therapy is a relational process. It involves being challenged, being witnessed, being misunderstood and repaired, being held accountable, and building the capacity to do some of that for yourself. Those are things that AI can’t replicate.
We’re left with a lot of unanswered questions…
Where do we go from here? What role do we want AI to play? What might it look like to treat AI as a tool, but not a stand-in for connection? Can we use it for structure, language, and support when nothing else is available, without allowing it to turn into the primary place in which our emotional life gets processed?
Can we build ourselves strong enough boundaries, strong enough safeguards, that AI can be used safely, in conjunction with human socialization and human therapy, to render mental health help accessible to the greatest number of people possible?
These are things that we have power over. They are things that we can decide, and not let be decided for us.
References
Deci, E. L., & Ryan, R. M. (2012). Self-determination theory. In P. A. M. Van Lange, A. W. Kruglanski, & E. T. Higgins (Eds.), Handbook of theories of social psychology. SAGE Publications. Cited by: (Data not available in metadata)
Wu, Y., Hua, I., & Ding, Y. (2025). Unveiling environmental impacts of large language model serving: A functional unit view. arXiv. https://doi.org/10.48550/arxiv.2502.11256 Cited by: 17
Wilson, E. L., & Martinez, M. (2017). The importance of respecting client values in the counseling process. Suluh: Jurnal Bimbingan dan Konseling. Cited by: (Data not available in metadata)
Psychiatric News. (2025). AI-induced psychosis: A new frontier in mental health. Psychiatry Online. https://doi.org/10.1176/appi.pn.2025.10.10.5 Cited by: (Data not available in metadata)
United Nations Development Programme. (2025). AI and the climate crisis. UNDP. https://www.undp.org/sites/g/files/zskgke326/files/2025-12/ai-and-the-climate-crisis.pdf
Oxford Academic. (2026). Effectiveness of AI-based interventions in workplace mental health: A systematic review and narrative synthesis. British Medical Bulletin. https://doi.org/10.1093/bmb/ldag007 Cited by: (Data not available in metadata)
PMC. (2025). Radiology AI and sustainability paradox: environmental, economic, and social dimensions. PubMed Central. https://pmc.ncbi.nlm.nih.gov/articles/PMC12006592/
PMC. (2026). When can AI psychotherapy be considered comparable to human psychotherapy? Exploring the criteria. PubMed Central. https://pmc.ncbi.nlm.nih.gov/articles/PMC12504474/
PMC. (2026). Navigating the complexity of AI adoption in psychotherapy by identifying key facilitators and barriers. PubMed Central. https://pmc.ncbi.nlm.nih.gov/articles/PMC12967751/
Wampold B, E. How important are the common factors in psychotherapy? An update. World Psychiatry. 2015 Oct;14(3):270-7. doi: 10.1002/wps.20238. PMID: 26407772; PMCID: PMC4592639.
Bastos MAV Jr, Braz DF, Porto ALM, Cordeiro KSDS, Portella RB, Granger DA. Physiological attunement and flourishing: understanding the influence of relationships on health. Front Psychiatry. 2025 Aug 1;16:1614379. doi: 10.3389/fpsyt.2025.1614379. PMID: 40821019; PMCID: PMC12354601.
Comments