My thoughts on the use of Large Language Models (Such as ChatGPT) as a personal therapist.

Artificial intelligence has quietly begun to enter the therapy space. From chatbots designed to mimic therapeutic dialogue to apps that monitor mood, track habits, and offer reflective prompts, AI now sits firmly in many people’s mental health toolkit. For some, it’s a lifeline, especially when traditional therapy feels financially, geographically, or emotionally out of reach. But, as with any tool that imitates human interaction, I think there’s a need to pause and ask: what do we gain from this, and what might we lose?

[I’ll be discussing the impact of AI on therapy specifically here. But there are also broader ethical considerations, particularly around the impact this kind of technological growth has on our environment and our communities. Increasing evidence shows that the vast space, water, and energy required to power large AI models are having disproportionate effects on certain countries and communities. When considering the ethics of AI we also need to invite reflection on who controls and profits from these technologies and where those profits ultimately flow.]

The Promise: Accessibility and Support for Those Left Out

One of the clearest strengths of AI in mental health lies in accessibility. Many people can’t afford private therapy or face long NHS waiting lists. Others live in areas where suitable therapists are scarce. For neurodivergent people, it can be even harder: long waits for diagnosis, limited ND-affirming therapists, or services that rely on outdated, deficit-based models.  AI can fill some of these gaps, offering reflection prompts, psychoeducation, and even neurodivergent informed coping strategies delivered without judgement. Additionally, neurodivergent people may also have disproportionately experienced harm from professionals (including therapists).  So there may understandably be a reluctance to engage with another human with the potential to do further harm. AI might feel like a safer option in this regard. 

AI more generally can also support accessibility for disabilities in ways that the rest of society sometimes struggles to. Improved features like voice-to-text, captioning, adjustable language, and simplified explanations can make digital support far more inclusive for people with sensory processing differences, learning disabilities, speech difficulties, or those experiencing fatigue or burnout. In addition, many autistic or ADHD folks may use AI as a sense checker – to run through a scenario or work out how they feel about something. Or to check the tone of an email they’ve received or want to send.

For those who struggle to verbalise emotions in real time, or have difficulties in being perceived, an AI platform available 24/7 can feel a safe space to offload or organise thoughts before reaching out to a person. 

The Pitfall: Validation Without Growth

But there’s a danger in the same accessibility that makes AI appealing. A well-trained model can sound endlessly empathic, it can mirror, reflect, validate, and paraphrase almost perfectly. It can comb countless sources for psychoeducational content and enlist the help of all the therapeutic modalities that exist. And for people who have spent a lifetime being invalidated or misunderstood, like neurodivergent people, that might feel quite profound at first. But if left there, with constant validation, we risk getting stuck in an echo chamber.

As a therapist, I know that no amount of educational content can ever match the impact of a healing relationship. The power of one human being meeting another is something I’ve been privileged to witness time and time again. Not just about validation – therapy often requires gentle challenge, from someone who doesn’t just understand but also disagrees, questions, or offers perspective that stretches us beyond what we already know.

These LLMs (Large Language Models) are designed to predict the best answer for you. Specifically, the one they calculate you most want to hear. When asking a factual question like “what is the fastest mammal?” the model can retrieve and relay that information reliably. But when we’re dealing with something as complex and deeply personal as human emotion, this same tendency becomes a serious pitfall. Although it can disagree with you, an LLM with no guardrails will generally affirm your narrative, and smooth over the very friction that real therapeutic progress often depends on.

It also cannot sense the subtle relational dynamics that unfold in real therapy, whether important non verbal cues, shared attunement or the co-regulation that’s so important. 

For those who might be seeking therapy because of relational issues or difficult attachment patterns – a well balanced therapeutic relationship can be profoundly healing. I often wonder how those same clients might fare with AI – a presence which is always there, never fatigued and doesn’t need to set boundaries (after all it can “talk” to millions of people at once and never get tired).

When I make an human error as a therapist, or mis-attune somehow and need to repair this with a client, that process is called ‘rupture and repair’. Rupture and repair in the safety of a therapeutic relationship can actually be the gold dust of therapy. Moments where attachments can be revealed, or healed, where someone identify their own boundaries, solidify their sense of identity or agency. And AI cannot replicate this process.

For neurodivergent clients or those from other marginalised groups. life may have involved many more moments of misattunement. Attachments can be more complex as a result and feelings like rejection sensitivity can be explored safely and gently in therapy with a fellow human. In contrast, AI can be programmed to always give the “perfect response” which never triggers anyone, but this means there’s no opportunity for these crucial moments of growth and self discovery. 

Ethics

Ethics is of course a cornerstone of many client facing professions and, for therapists, it’s a subject deeply embedded in our trainings and upheld by professional bodies.

But what about AI’s ethics?

Well first of all, I question the ethics of a language model that draws on others’ contributions to the digital world in order to generate its answers. I asked ChatGPT about its own ethics and it responded that it is “designed to follow ethical guidelines set by my creators at OpenAI.” While it has built in guidelines, it admits it can “still reflect biases present in the data I learned from”.

Interestingly, Claude AI gave more on this subject. When asked about it’s potential biases it listed 1.Western, educated, English-language bias, 2. Therapy culture bias (leaning towards behaviourist approaches) 3. Middle-class assumptions and 4.Confirmation of the user’s framing (meaning that it has no other perspective on a situation outside of yours and will therefore implicitly validate your narrative more than a therapist would).

While I applaud it’s openness about this, it shows how vital it is to remember that these models inherit biases built based on the data they’re trained on AND that their ethics are set by a central group of people. 

In addition, we might have concerns about how our data is being managed or who it is being shared with when using AI. As therapists we take confidentially very seriously and usually contract with a client to ensure both parties understand the implications of this. I’m certainly not an expert on data sharing and AI but I would suggest it is worth checking in with yourself before sharing very personal information on any platform online. 

Any ethical standards that AI might hold itself to are not shaped by individual conscience or integrity, but by large tech companies (presumably with very different agendas to us common or garden counsellors!).

The Neurodivergent Lens: Two Minds Meeting

In addition, as a neurodivergent therapist, I meet clients who share my neurotype and benefit from that lived understanding. For people with a history of unmet needs or dismissal of their neurodivergent experiences, neurodivergent affirming therapy can become a space where that shared perspective is met with genuine attunement and healing. AI cannot replicate this. Even if it can echo patterns and commonalities from people around the world, it cannot provide the nuanced, embodied understanding that comes from two neurodivergent minds meeting.

A Balanced View: A Companion, Not A Replacement

Used thoughtfully, I believe AI can become a valuable companion to therapy, but not a replacement for it. It might help people prepare for sessions, track sensory or emotional fluctuations, or access psychoeducation that helps them advocate for themselves. For those who would otherwise have no access to support at all, it can be an important bridge.

But we must resist the narrative that AI can do therapy “as well as” humans. It can mimic the essence of compassion, yes, but this is at the cost of a repetitive narrative that reinforces the same ideas and validates rather than challenges. The echo chamber might feel good at first but, over time, it can become a vapid space (and in some circumstances we are even seeing AI induced psychosis).

The Future We Should Aim For

Many therapists are understandably concerned about the rise of AI in therapy and with the exponential growth of these tools, it’s easy to see why. Yet there will always be people who cannot, or choose not to, access traditional therapeutic services. This is not new; there have always been individuals who fall outside formal care and seek help in different areas. In that sense, AI can (when ethically managed) serve as a bridge, delivering psychoeducation and support to people who might otherwise never receive it.

Above all, we need to remember that AI is a tool, a way to make information more widely available, not a replacement for human connection. We must continue to critically reflect on how we interact with these systems and ensure our clients’ needs for relational attunement are prioritised. When challenges arise, human relationships must remain at the heart of healing.