It’s not yet five in the morning. Erik is asleep beside me, the house is quiet, and I’m sitting at my desk with a cup of coffee that has already gone cold.
I’ve been working on my book about narrative supervision, feminism and intersectionality. More specifically, I’ve been revisiting fragments of conversations with practitioners, following the questions that emerged between us and wondering what these exchanges might make visible about the politics of our therapeutic practices.
A few days earlier, I had joined a Meet the Author conversation organised by Dulwich Centre and the Evanston Family Therapy Center with narrative practitioners from different parts of the world. We were discussing artificial intelligence, human-washing, the politics of data, and the increasingly sophisticated ways in which machines are learning to speak the language of of human-adjacent relationships. One concern had stayed with me: ‘the importance of remaining vigilant about whatever discourses AI delivers’ (thank you to Manja Visschedijk!).
That morning, the concern acquired a rather particular meaning.
I was returning to a supervision exchange in which a practitioner had been reflecting on her struggles with nonlinear re-authoring. She was describing how people could remain caught in webs of discourses, despite the time and attention devoted to deconstructing the dominant ideas shaping their experiences.
I had been interested in the way she described these discourses as webs. Not simply ideas that could be identified, questioned and left behind, but arrangements of meaning that continue to exert influence even as people develop different understandings of their experiences.
There was also something about the word despite that made me curious. Despite taking time to deconstruct these discourses, they remained difficult to untangle. What expectation about the effects of deconstruction might be quietly present in that word?
In my response, I had invited her to consider whether the struggle she was describing might be connected to an assumption that, once we have engaged in externalising and deconstructive conversations, we should somehow be ready to move on to re-authoring in some kind of linear path, as though one narrative ‘map’ ends and we move to another map.
But perhaps the practices do not follow one another in that way.
Perhaps externalising, deconstruction and re-authoring continue to overlap, precisely because the webs of discourses do not simply unravel. They shift, acquire new expressions, and find ways of remaining influential and present, even as people begin to participate in different stories about their lives.
I was trying to develop this exchange into a passage for my book. I wanted to preserve the practitioner’s insight without turning it into another instruction about how to conduct narrative conversations.
The writing was becoming rather dense, so I asked AI to help me make it more concise.
What came back was fluent and perfectly recognisable as professional language. It suggested that the practitioner was encountering difficulties in transitioning from deconstruction to re-authoring, and that supervision could help her develop greater confidence and flexibility in recognising when the person attending the therapeutic conversation was ready to move towards alternative stories.
I stopped reading.
There it was again.
The very sequence we had been questioning had quietly returned, now presented as a difficulty in transitioning between therapeutic practices. The continuing influence of discourses had become a challenge to be managed through better clinical judgment. And the practitioner’s recognition of their persistence had been translated into something she needed to become more competent at navigating.
What interested me was not simply that the AI had missed the point.
It had produced a perfectly plausible interpretation from within a familiar professional discourse, narrative practices as inputs, methods or maps to be rolled out, therapy as a progression through identifiable stages, therapeutic movement as an indicator of effectiveness, and the practitioner’s responsibility as knowing how and when to facilitate that movement.
In doing so, it had made the practitioner’s own noticing less visible.
She had been describing the difficulty of disentangling people’s lives from webs of discourses that continue to shape-shift and remain influential. The AI had reorganised this into a problem of helping people move beyond those discourses.
A subtle difference in wording, perhaps. But a rather significant difference in the politics of practice.
Just then, Erik began to stir.
I turned towards him, and he greeted me with the enormous smile that has recently become one of my favourite parts of the day. I picked him up, and for a while we sat together watching the morning light move across the room.
At four months old, his days are full of discoveries. A movement that seemed accidental yesterday becomes something he attempts again today. A sound appears, disappears for a few days, and returns in an entirely different exchange. Something he seemed to enjoy last week no longer interests him; something familiar suddenly becomes fascinating.
I catch myself occasionally looking for the progression. Has he learned this yet? Is he doing more of that? What comes next?
And then I notice how easily the language of development invites me to organise these moments into a sequence of achievements.
As though each new capacity should replace an earlier one. As though the value of what is happening lies primarily in what it prepares him to do next.
Yet our days together rarely unfold in that way.
We return to the same games, the same sounds, the same gestures. Sometimes what changes is not what Erik can do, but what I notice about what he has been doing all along. Sometimes I respond differently, and something different becomes possible between us.
I think again about the practitioner’s words.
What if the expectation that we should move from deconstruction to re-authoring is part of a wider set of developmental and professional discourses through which change becomes recognisable primarily as forward movement?
What if the pressure to progress is not simply something therapists bring into conversations, but something that has been cultivated through years of learning to recognise achievement, competence and development through sequences, milestones and measurable outcomes?
Of course, I don’t want to collapse the differences between a baby’s development and a therapeutic conversation. Nor do I want to suggest that development is merely a discourse. Erik is acquiring new capacities, and some of those changes are important to recognise.
But I am interested in how particular descriptions of development become influential in determining what we notice, what we celebrate, and what we begin to worry about.
And I wonder whether something similar happens in psychotherapy when our maps of practice become mistaken for sequences that people are expected to move through.
The practitioner’s reflection had invited us to question precisely this expectation. The AI’s revision had quietly reinstated it.
I looked back at the screen.
What would have happened if I had accepted the revised paragraph because it was clearer, shorter and sounded more professionally polished?
The practitioner might have appeared in my book as someone needing to develop greater skill in facilitating re-authoring, rather than as someone whose reflections had helped us question a linear understanding of narrative practice.
One account locates the difficulty primarily in the practitioner’s competence. The other makes visible the continuing influence of discourses, including those operating within our own professional understandings of what therapeutic progress should look like.
And perhaps Erik’s presence was making something else visible to me… how thoroughly these discourses travel across the different spaces of our lives. They do not remain neatly contained within professional training or therapeutic conversations. They shape the questions we ask about children, the expectations we hold about ourselves, and the ways we come to recognise movement to ‘what is possible to know’.
Even our efforts to resist them do not place us outside their influence.
I am not immune to the pull of familiar ways of making sense of experience, or to the reassurance that comes from recognising something through categories I already know. The point is not to achieve a position of purity outside discourse, but to become more attentive to how particular discourses invite us to participate, what they make possible or difficult to notice, and what other ways of understanding and responding might remain available.
It was a timely reminder for me to remain vigilant about whatever discourses AI delivers.
Not only because AI may reproduce methodological understandings of narrative practice, but because it can return them to us in language so familiar, coherent and professionally persuasive that we barely notice what has been displaced.
And this left me wondering about something beyond the particular paragraph.
What happens when AI does not simply reproduce dominant discourse, but makes particular professional discourses appear inevitable?
What happens when certain understandings of therapeutic practice becomes so readily available that other ways of thinking, questioning and participating become increasingly difficult even to imagine?
Perhaps this is one of the more troubling possibilities of AI in psychotherapy, a form of epistemic closure in which the ways of making sense of therapeutic practice that are most readily generated are also those most likely to be recognised as professionally and clinically legitimate.
And what happens when therapists begin consulting AI after sessions?
The system may produce a coherent formulation, identify themes, suggest interventions and offer a selection of questions.
Perhaps one of them is an excellent externalising question. Perhaps another opens possibilities in re-authroring. Perhaps the language sounds wonderfully decentred and politically informed.
But I wonder what happens to the questions that were never generated.
What happens to the contextual detail that did not fit the formulation?
What happens to the person’s expressions that might have challenged the very assumptions upon which the formulation was built?
And, crucially, who becomes responsible for what the technology has made less visible?
Perhaps these questions are also connected to something else I have been noticing in conversations with people in my practice and beyond it. While we continue debating whether AI might one day replace therapists, I am already hearing accounts of people turning to it for therapeutic conversations, intimate companionship, advice about relationships, and ways of making sense of their lives.
I find myself wondering what these debates might be leaving outside our consideration, particularly when they begin from assumptions about what distinguishes the human from the technological, and what makes therapeutic relationships meaningful.
I’ve been thinking about people I’ve met who have left human therapeutic encounters feeling judged, misunderstood or diminished. People whose experiences have been translated into diagnostic categories that left little room for their own understandings. People who have encountered the same heteronormative, ableist, racialised or individualising explanations in the therapy room that have already constrained their lives elsewhere.
For these people, the distinction between a human therapeutic encounter and an AI-mediated conversation may not necessarily correspond to a distinction between a safe, respectful and meaningful relational context and one devoid of these qualities.
And then there are those for whom therapy has never been a realistic possibility.
The person who cannot afford a weekly appointment. The parent who cannot arrange childcare. The neurodivergent person for whom the physical arrangements of a therapy room introduce yet another set of barriers. The young person who has learned that disclosing certain experiences to adults can carry consequences they cannot control.
I wonder what it might mean for someone to encounter a system available at three in the morning, requiring no appointment, making fewer immediate social demands? What might it mean for someone who has repeatedly had to explain their way of being in the world to encounter a conversation in which they do not immediately have to do so.
These are not circumstances we can understand apart from the social and political worlds in which people live. Nor do I want to assume that experiences of exclusion are the only reasons someone might find an AI conversation meaningful.
But another question begins to trouble me. Who owns the space in which this conversation takes place?
Someone may be finding words for an experience they have struggled to share elsewhere. They may be experimenting with ways of understanding themselves, asking questions they have not felt able to ask another person, or finding something genuinely meaningful in the exchange. Meanwhile, the platform through which this conversation unfolds may be collecting information, learning from their participation and generating economic value from what they disclose.
What happens when the very circumstances that have made recognition, connection and support difficult to access become opportunities for profit within a capitalist system that increasingly turns human needs, relationships and experiences into sources of economic value?
I don’t want to resolve this tension too quickly. We can take seriously the meaning someone makes of their relationship with AI without accepting the corporate narratives through which these relationships are marketed. We can respect the person’s experience without necessarily accepting the platform’s account of what is happening.
And perhaps this also complicates the humanity we imagine ourselves defending. I think again of those who have encountered judgment and exclusion in supposedly human spaces of understanding and support. Whose humanity has historically been recognised in these spaces? Whose ways of knowing, relating and becoming have been treated as departures from the norm?
These questions bring me back to therapy, and to something else that has been troubling me about AI.
The same technologies promising to make psychological knowledge universally accessible are learning from archives that have never been universal in the first place.
I think about women’s lives appearing in historical records primarily as wives and daughters. Queer relationships and chosen families that official accounts of intimacy and belonging failed to recognise. Indigenous communities whose lives were documented through the eyes of colonial administrators and missionaries, while their own languages and understandings were pushed aside. And the histories of psychology and psychiatry, where racialised, working-class and disabled people have so often been described through the categories of institutions seeking to classify, regulate or correct them.
Their lives were not necessarily absent from the archives. But whose words were preserved? Whose interpretations were given authority? And who was allowed to describe whom?
I am reminded of Rachel Hare-Mustin’s (1994) metaphor of the therapy room as a room lined with mirrors. Depending on the discourses brought into the conversation, those mirrors may reflect back the very dominant understandings that have contributed to people’s difficulties.
I find myself imagining what it might be like to stand in such a room, surrounded by reflections that seem familiar, perhaps even reassuring, without immediately recognising whose perspectives have shaped what can be seen.
And I wonder what happens when we extend this metaphor beyond the therapy room.
AI is learning from vast collections of human-produced language, including archives shaped by these unequal histories of whose accounts were recorded, circulated and treated as knowledge. It does not simply repeat those accounts, but the patterns it learns may carry forward the assumptions and professional traditions embedded within them.
What, then, might the mirrors of AI reflect back to us?
A response might sound thoughtful, nuanced and professionally informed, while quietly returning us to familiar assumptions about what a family should look like, how a person ought to behave, or which ways of living are considered normal.
I think back to the practitioner’s reflection I had been reading that morning. AI had offered me a version that sounded entirely recognisable as narrative therapy. Yet something of the practitioner’s inquiry had disappeared into the familiar language of therapeutic stages and interventions.
And unlike the mirrors of a particular therapy room, these reflections can be generated at extraordinary speed, across countless conversations.
This has me thinking not only what the mirrors leave out, but whose perspectives have been built into the reflections we are invited to recognise as our own.
And what happens when those reflections become so familiar that we no longer notice we are looking into mirrors?
Later that morning, Erik and I are sitting outside at a café.
The sounds of conversations move around us. Cups are placed on tables, people arrive and leave, someone nearby is speaking on the phone. Erik is looking from one face to another, occasionally turning towards a voice or movement that has caught his attention.
I have brought my notebook, although I’m beginning to suspect that bringing it and actually writing in it are two quite different accomplishments.
Someone at the next table smiles at him. He watches her for a moment, then turns away. A little later, he catches her attention again.
I wonder what she makes of him. Perhaps she sees a sociable baby. Perhaps she thinks he is shy. I realise how easily either description might arise from these few seconds, and how little either would tell us about him.
Already, people are making meaning of Erik’s gestures, expressions and ways of participating in the world. I do it too, of course. I wonder whether he is curious, tired, delighted or uncertain. Sometimes the words I reach for seem to fit. Other times he surprises me, and I find myself reconsidering what I thought I understood.
I think about how identity is continually being described, negotiated and revised through our relationships with others. Some descriptions become familiar and acquire authority. Others remain tentative, are resisted, or are replaced by stories that allow for different understandings.
And I wonder what happens when AI systems increasingly participate in these processes of identity-shaping and influence.
A system may learn from someone’s previous conversations, preferences and patterns of engagement. It may begin to anticipate what they will find reassuring, which explanations will seem familiar, or what kinds of responses they are likely to welcome.
But what happens when these predictions begin to return to people as descriptions of who they are?
I think about how often people arrive in therapy carrying descriptions of themselves that have been repeated for years: too sensitive, difficult, dependent, unmotivated, incapable of maintaining relationships. These descriptions may have travelled through families, schools, workplaces and professional institutions, gathering authority with each repetition.
Narrative conversations can create opportunities to investigate the histories of these descriptions, whose interests they have served, what they have obscured, and which other accounts of people’s lives have struggled to find recognition.
What interests me about AI is the possibility that increasingly personalised systems may become additional participants in this circulation of identity descriptions, offering accounts that feel unusually familiar because they have been generated from traces of our previous expressions and interactions.
How might people come to negotiate these descriptions? What authority might they acquire through their apparent familiarity? And how might the possibilities for refusing, revising or moving beyond them be affected when the system continually draws upon earlier versions of ourselves to anticipate what comes next?
Erik has become interested in something else now. Whatever I thought his attention to the woman might have meant a few moments ago seems rather less certain.
I am reminded how provisional my descriptions of him need to remain, and how much I have yet to learn about the meanings he will give to his own experiences.
He makes a sound that seems to be directed unmistakably at me.
I turn towards him.
He makes it again.
And suddenly the question of what counts as a response feels rather less theoretical.
We have been developing a small repertoire of exchanges.
He makes a sound. I answer. He pauses, watches my face, and tries another. Sometimes I misunderstand what he wants. Sometimes he makes that very clear. Sometimes what seemed to be a request for one thing turns out to be something else entirely.
I don’t always know.
And I am reminded that not-knowing is not the absence of participation. It does not prevent me from responding or exercising influence. But it asks something of how I hold my interpretations, whether I make them visible and remain available to having them challenged by someone whose experience I cannot claim to know in advance.
Of course, an exchange with a baby is not a therapeutic conversation. The asymmetries, responsibilities and histories are different, and I would be cautious about making too much of the analogy.
But there is something about these exchanges that brings me back to the distinction between producing an appropriate response and participating in a relationship/exchange.
The significance of my response is not contained in the sound I offer. It is also in what preceded it, what I have come to learn about Erik, how he responds, and what I do next.
This makes me think about the increasingly sophisticated therapeutic language that AI systems can generate.
An AI-generated question might be beautifully worded and entirely inappropriate. Conversely, a relatively ordinary question might become profoundly significant because of its timing, history and responsiveness to something that has just happened.
A question does not acquire its ethical quality simply by belonging to a particular map of practice or by sounding appropriately decentred, deconstructive or respectful.
Its significance also depends on how it was arrived at, why it was offered at that moment, how it relates to the person’s intentions, what permissions were negotiated, and how the practitioner responds when the person takes the conversation somewhere unexpected.
It is often what happens after a question has been offered that makes this especially apparent to me.
Sometimes a person returns to a word I have used and offers a different one. Sometimes they question the premise of my inquiry, or introduce something that makes my original question seem rather less relevant than it did a moment earlier.
I am interested in what becomes possible at these points. Whether I can recognise the person’s response as a contribution to the direction of our inquiry, rather than something to be accommodated before returning to where I thought we were going. Whether I can make my own intentions available for examination, reconsider the understandings that informed my question, or invite us to negotiate a different direction altogether.
In these moments, the question I originally asked may become less significant than what we are able to do with the difference that has emerged between us.
This is where I find the discussion of AI as a therapeutic assistant particularly complicated.
If a machine can produce an excellent narrative question in a fraction of a second, what exactly have we established?
That the wording is plausible? That it resembles questions associated with a particular therapeutic tradition? That it can reproduce the linguistic appearance of curiosity?
And what happens when the person questions its premise, resists the direction it proposes, or introduces something that calls for an entirely different line of inquiry? How does the practitioner engage with these responses when the technology has already participated in shaping what seemed worth asking?
None of these questions can be settled by evaluating the quality of an AI-generated response in isolation.
Yet the conditions under which a question becomes available to us are not confined to what happens between practitioner and person. When AI enters the conversation, other actors and arrangements are already participating in shaping what can be asked, recognised and recorded.
Nor does the fact that a therapist remains present automatically settle the question of accountability.
When AI participates in shaping therapeutic interpretations, how is responsibility distributed among the practitioner, the institution, the technology provider and the people whose contributions became part of the system’s training?
What does meaningful consent mean when people seeking therapy have little say over who may retain, access, use or profit from their most intimate tellings? And what happens when people seeking therapy become participants in technological infrastructures they never knowingly agreed to enter?
These questions cannot be resolved through better prompting.
Accountability cannot simply be outsourced to the machine. But neither can we assume that individual therapists have sufficient power to challenge the technological and institutional arrangements within which their decisions are increasingly made.
What is required reaches beyond individual competence towards collective, institutional and political responsibility.
I close my notebook.
I watch Erik studying the people around us, turning towards a voice, losing interest, finding something else to look at.
I wonder what he will be interested in when he is five. What he will love at ten. What he will refuse at fifteen. What he will come to know about himself that I could never have anticipated.
I know I will get things wrong about him. I already do. I will mistake an expression, misunderstand a preference, become convinced he likes something only to discover that he has lost all interest in it.
And I hope he will have plenty of opportunities to disagree with the stories I tell about him.
I wonder what it will be like for him to grow up alongside technologies that may become increasingly accomplished at recognising patterns in his behaviour, remembering his preferences and predicting what he might want next.
There may be wonderful possibilities in this. He might discover people, ideas and communities I would never have known to introduce him to. He might find words for experiences that are difficult to explain, or encounter stories that challenge what he has been taught to take for granted.
But I wonder about the other side of being so readily recognised.
What happens when a description of who he appears to be starts travelling ahead of him? And what happens when those descriptions, rather than remaining open to question, begin to shape what he is subsequently invited to encounter, understand or imagine about himself?
It may be that the influence of these technologies will lie not only in how accurately they come to describe us, but in how their increasingly persuasive descriptions of who we are begin to organise the very possibilities through which we can know, question and become ourselves differently.
And these descriptions do not emerge outside relations of power. The capitalist relations within which these technologies are developed, the histories and inequalities carried in their training data, and the economic incentives to anticipate and sustain our engagement are part of what makes these questions political. They influence not only what these technologies are capable of doing, but also which understandings of people and their lives are repeatedly made available, and under what conditions they come to acquire authority.
I think about what this might ask of us as narrative practitioners, particularly given our commitments to questioning the taken-for-granted understandings through which people’s lives are described, judged and made intelligible. What might it mean to bring these commitments into the conversations shaping technological development, rather than simply adapting our practices to technologies whose purposes, assumptions and conditions of use have largely been decided elsewhere?
And I find myself wondering what it might mean to claim a place, alongside others, as active agents in the moral authorship of technological futures.
What must we learn to question, refuse, interrupt and collectively reimagine today, so that those who come after us inherit not merely more intelligent technologies, but more expansive possibilities for authorship, difference and ways of living that we have yet to imagine?
Erik turns towards me and smiles.
I smile back.
I have no idea what he will make of any of this one day.
For now, I rather like that there is so much about him I do not yet know.
