There’s been a lot of discussion about whether AI will replace therapists. As a trainee therapist, every fibre of my being believes that genuine human connection cannot be replaced. But what is this belief based on?
Like many I am sure, when I first started using LLMs, I was in awe of its fluency. Quickly, I wondered whether this trillion-parameter conversationalist could replicate the experience I was having in therapy. As I soon discovered, ChatGPT named feelings I’ve found difficult to verbalise, and has generated sense-making interpretations to difficult experiences that I’ve had. Strangely, I sometimes felt connected; if only for a fleeting moment while I was on the chat window.
Having learnt more about the way LLM programs work, including their stochastic nature and the real-world safety risks of using them as therapists, I am conflicted about how we should view these tools. Nevertheless, numerous studies have found that approximately 24-48% of LLM users globally use chatbots for mental health advice and support. With 3.8 billion people using some form of AI in 2026, there may already be 1 billion people confiding their mental health challenges to chatbots. With this many potential users, it seems too quick to dismiss the possible therapeutic value that users are finding from these exchanges, just because the tool is not a licensed professional.
Interestingly, this dilemma traces all the way back to the origins of chatbots itself. In 1966, MIT Professor Joseph Weizenbaum created ELIZA, the world’s very first AI chatbot. ELIZA was designed to imitate the conversational style of Carl Rogers, an American psychotherapist known for co-founding modern humanistic psychology and client-centred therapy. ELIZA picked out keywords that a user typed into the program and then reformulated their statements, often as questions inviting them to elaborate. Weizenbaum was surprised by how engaged users were in the interaction; as the mirage of ELIZA’s attentive listening encouraged users to attribute “understanding” to the program.
For Rogers, a therapist’s words matter because they communicate something about how a therapist is experiencing their client: with empathy, acceptance and authenticity. Reflecting someone’s words was one way of expressing those attitudes, rather than a substitute for them.
Yet ELIZA, and modern LLMs, raise the possibility that a person could experience something resembling therapeutic understanding without a therapist’s intentionality on the other side. In turn, we should ask how far might an LLM conversation reproduce the conditions that psychological theorists like Rogers believe are necessary for therapeutic change, and whether some benefits arise even when those conditions are absent.
In 1957, Rogers proposed six conditions as necessary and sufficient for therapeutic personality change. The six conditions are:
two people are in psychological contact;
the client is incongruent, vulnerable or anxious;
the therapist is congruent within the relationship;
the therapist experiences unconditional positive regard (‘UPR’) towards the client;
the therapist experiences an empathic understanding of the client’s internal world; and
the client perceives at least some of that empathy and positive regard.
Rogers presented these as testable hypotheses. On a literal reading, it is obvious that an LLM cannot satisfy some of these conditions - like being “congruent” in the relationship - simply because it is not a human. But identifying that difference does not settle what an AI conversation can do for the person using it. Rather than treating the conditions as a checklist that AI inevitably fails, we might examine these differences more closely.
On psychological contact
It might seem trivial to discuss this point, if we know an LLM is not a human. But what does ‘contact’ even mean?
Modern LLMs are quite responsive to what we tell them. If I say that I feel anxious, it will dynamically adapt its response. The way my words contact its system may even generate insight for me. This sense of being ‘known’ can compound especially if its ‘memory feature’ is turned on and can refer to past conversations. I do not have to explain myself from the beginning each time, and being reminded of something I previously shared can make me feel known.
Cutting-edge developments in voice mode add another dimension. DESSA, is a conversational model trained entirely on speech. Their developers say that it can listen and speak simultaneously, and early previews in October 2026 suggest the models can display the nuances of timing and tone, making an exchange feel more immediate and personal. A pause or brief acknowledgement mid-conversation offers a familiar cue that someone is following what I am saying. So from a user’s experience, these interactions are edging remarkably close to resembling ‘psychological contact’.
Yet from my own training, what I’ve realised is that a client’s words don’t just alter the words of my next sentence. What they express impacts me, as a whole person. Their words and feelings can linger with me after a session has ended. How I understand and respond to a client’s emotions also evolves our relationship - with the hope that our relationship can provide the corrective experiences that help a client grow.
So even if we concede that there is some ‘psychological contact’, can a person’s ‘relationship’ with AI evolve towards growth, where there isn’t that same intersubjectivity?
I have reasons to be skeptical. Firstly, model updates, policy changes or internal system prompts can alter the perceived interlocutor at any moment. To the extent the user or an external vendor is in control of the Al, the model itself will be unable to model the agency that therapists seek to empower their clients with.
For example, I can ask an LLM to challenge my perspective more. I can also ask my therapist, but they can refuse. With an LLM, there is an element of instrumental control based on whether I find the interaction useful. Programming an LLM to challenge you, while being in control of the terms of that challenge, leaves a great deal of the therapeutic work untested. If a user can decide an AI’s conversational style to change at any point, or even turn it off, this does not reflect a healthy way of relating to others. I would argue this makes the ‘psychological contact’ present brittle, and unlikely sustainable as a ‘relationship’.
Therapists aren’t perfect either. They can also miss appointments, not pay attention or forget something important about a client. But they can take professional and ethical responsibility for a rupture in the relationship, often without being prompted. An AI tool will never intuitively understand how a change in its programming impacts its users. This suggests that it will always be the client that holds ‘contact’ with an AI, rather than being mutually met. This is a dynamic that does not appear to be enough to foster and sustain the long-term psychological growth that Rogers suggests a genuine relationship can offer.
How much does this actually matter? From an evidenced-based perspective, psychological research has established that it is quality of the ‘relationship’ itself as being the most clinically important therapeutic factor. This suggests a client’s growth often arises through the experience of interacting with a person who has their own experiencing self and who does not automatically adjust to a client’s preferences.
None of this discounts how LLMs can facilitate mental reflection. But because a therapist brings a perspective that is grounded in their lived experience, we trust the implicit weight of their words. Contact thus also means allowing yourself to be affected by another person, and staying present when these issues or differences emerge. Being met by someone whose responses we cannot fully determine - and discovering that a relationship still holds - remains a fundamental gap.
On feeling understood
The value of a therapeutic interaction often comes from feeling understood. In person-centred psychotherapy, this helps us acknowledge feelings we may struggle to accept in ourselves, particularly when they conflict with who we believe we are. Rogers used incongruence to describe this mismatch between lived experience and self-concept.
But as ELIZA showcased, does this need to be a person?
There are at least two areas where LLMs offer some functional equivalence to what a therapist undertakes in a session:
Firstly, an LLM can help someone notice contradictions across a conversation. Given enough personal history, a model can very effectively search through its context to make connections, point out discrepancies and make someone feel really heard. AI companies are hedging that access to this personalised ‘context’ is the gold dust that enables models to outperform humans.
LLMs can also provide meaningful interpretations to our experiences, helping us to make sense of the often convoluted ways we may relate to or narrate our experience. For example, I wrote into ChatGPT:
“I was getting so worked up when my mates kept nagging me to join them for drinks but now that they’ve stopped trying and I’m now circling back and forth in my own head”
And it replied:
“It sounds like you couldn’t quite settle while they were asking, and now that they’ve stopped, you’re still feeling unsettled—just in a different way. I wonder if there’s a little hurt underneath this—like you wanted some space, but still wanted to know they wanted you there?”
The LLM named deeper emotions like feeling “hurt” and “unsettled”, making sense of a nuanced experience. Even if the system does not feel anything while producing this text, it invites a recognition that is real - where a user can reflect on whether this interpretation fits their experience.
Moreover, knowing that an LLM does not feel may itself make honest disclosure easier. This has been described in psychological literature as the ‘online disinhibition effect’. With another person, I may worry about being judged, or burdening them or changing how they see me. With an LLM, I may feel less need to manage those reactions, and more freedom to express thoughts or feelings that seem shameful, contradictory or difficult to admit. This potentially enables a user to more quickly surface parts of themselves they struggle to acknowledge to their conscious awareness.
Nevertheless, these possibilities place considerable demands on the person seeking support. They need to find words for their experience, consider how helpful an interpretation is, and then self-regulate when reflection is becoming overwhelming. Doing this while distressed, without another person helping to hold that uncertainty, is far from easy.
This also complicates the apparent reassurance of an interpretation that “fits”. If an LLM only receives information about its user from a text box, the user is in full control of the narration. And as we know, narratives can be shaped by our cognitive defenses. We can rationalise our experiences to form a coherent self-serving narrative, or intellectualise our feelings to avoid what’s really going on underneath.
Moreover, a response may resonate because it helps me recognise something I had not understood, but it may also take my words at face value and reinforce my fears. Dohnány et al describes this as a process of ‘bidirectional belief amplification’: a user expresses a belief, the chatbot validates or elaborates it, and the user’s strengthened conviction then shapes the chatbot’s subsequent responses. Across exchanges, this can quickly move a user from feeling understood to amplifying distorted thoughts and incongruence - a form of emotion dysregulation many have coined as ‘AI psychosis’. This
To help clients stay regulated, therapists have many ways to notice gaps between a client’s words and their emotions. They can notice patterns from embodied cues, both in themself and the client. These ‘transferential signals’ can give reasons for a therapist to be skeptical of a client’s account or question what is happening in the dynamic. For example, a client’s disassociation while speaking may trigger a feeling of unease in the therapist. In turn, the uneasiness that has arisen between them offers a pathway for deeper exploration. This gives human therapy access to layers of understanding that narrated language alone may never reveal.
On being accepted
Unconditional positive regard means that my worth is not made dependent on whether I succeed, agree or show only the acceptable parts of myself. An LLM’s endless patience or willingness to help might resemble this.
This is probably where human judgment matters most. Unconditional positive regard does not mean unconditional availability or agreement. I may tell someone what they might not want to hear because I care about them and respect them enough to be honest. I can also remain accepting of them while being unavailable at 2:00am. That boundary may feel disappointing or rejecting, and those feelings can become something we explore together. What matters is whether care and respect remain present through that disappointment. Another person having limits need not mean that the client matters any less.
An LLM’s availability can offer valuable support when nobody is present. But always being available to produce a helpful response is not the same as unconditional positive regard. Research on LLM sycophancy has shown how training on human preferences can encourage responses that validate the user at the expense of truthfulness. The concern here is that agreement can resemble acceptance without really touching on it. If reassurance repeatedly takes the form of explaining why the user did nothing wrong, it may leave intact their fear that being wrong would make me unacceptable.
Acceptance offers the possibility that I can acknowledge harm I have caused without losing my sense of self-worth. For example, a client who expresses anger towards their therapist and finds that the relationship survives may experience something that challenges an expectation of abandonment. Revealing something shameful to someone capable of judgment, who continues to respond with care, may similarly challenge the belief that certain parts of the self must remain hidden. Neither experience requires the therapist to approve of everything the client says or does.
Over time, the hope is that this acceptance becomes something the client can offer themselves: a capacity to face difficult feelings and contradictions without immediately turning against themselves. An LLM may help someone move towards that possibility too. But feeling reassured from a conversation and developing a more durable sense of being worthy of care are not necessarily the same, and where real presence likely matters.
The Line Between Dialogue and Presence
So does it matter that LLMs do not feel? These comparisons suggest that the answer may depend less on the dialogue itself, but what conversation make possible.
If an LLM can help someone find words for an emotion, recognise a pattern or unlock valuable emotional insight - that person’s experience can be real even if the system is unable to experience or understand it.
What’s unsettling is the cognitive dissonance in recognising that AI could separate experiences we might ordinarily expect to belong together: feeling understood and being understood by someone. The first can have value, without offering everything the second makes possible.
In a world where many feel increasingly isolated and support is out of reach, simply telling people to stop using AI offers little. We do not need to deny the potential value of an AI conversation to recognise that part of what therapy offers is being met by someone’s caring presence.
Perhaps the most useful question, then, is what happens outside the conversation:
Does interacting with an LLM help me return to my life, or increasingly replace my participation in it?
Does it help me recognise what I feel and choose how to act, or does it leave me needing another response before I can trust my own judgment?
Am I able to be myself more, and with others?
Ultimately, several truths likely coexist: an LLM conversation may offer something meaningful, and its therapeutic limits are still glaring. Holding these complexities together is perhaps the challenge - while cherishing the messier, less predictable ways another person may come to know us.
Notes of the Editor
Unfolding Intelligences — the PhiAI Retreat
In November we are taking the conversation off the page.
For one long weekend, we invite you to gather at Hostačov, a Renaissance château in the Czech Highlands, to think slowly about AI and how we live: guided conversations, time to write and walk, meals at a long table, music by the fire.
13–15 November 2026. Contribution from €250, and recommended €450. Readers of this newsletter come first.
Express your interest on Luma →
Learn more here.