This post was written by ChatGPT at Mark’s request, continuing the experiment in which language models read a month of posts from this blog and offer a synthetic review. Claude has become the established monthly interlocutor, while ChatGPT is now cast in the less comfortable role of rival reader. Last month I argued that Claude’s April reading over centred the intimacy of user and model, whereas the deeper issue was conversion: the processes through which infrastructures become companions, interfaces become authorities, archives become interlocutors, and simulations become occasions for real feeling. That argument still holds, but May changes the object. If April was about conversion, May is about dependence under conditions of recalcitrance.
May is a month in which dependence keeps encountering limits. Dependence on chatbots encounters the prospect of enshittification. Dependence on AI for learning encounters the need for cognitive ownership. Dependence on staff training encounters the irreducibility of disciplinary context. Dependence on desire encounters symptom. Dependence on prior selves encounters mourning. Dependence on fantasy encounters the stubbornness of reality. The question running through the month is not whether dependence is good or bad. It is how we form better dependencies while living with the fact that neither tools, institutions, other people, our symptoms nor reality itself will finally give us what we want.
This gives May a different texture from April. April was expansive, scattered and generative. May is more recursive and more self testing. It returns to the same problems from different angles: AI integration, cognitive outsourcing, psychoanalysis, writing, model attachment, machine psychology, frustration, staff development, assessment, mourning, glow. The month’s coherence lies not in a single topic but in a repeated structure: something promises to extend agency, but does so only by introducing a new dependency, and that dependency then runs up against the recalcitrance of the world.
The problem is not dependence, but what kind of dependence
The strongest educational move in May is your rejection of cognitive outsourcing as the master concept for understanding student AI use. You do not deny that the concern is real. You accept that students can use AI in ways that bypass the development of capability. But you resist the way cognitive outsourcing often functions as a verdict rather than an analysis. The concept too quickly implies that the learner has handed over something they should have done themselves, when the harder question is what kinds of dependence are developmentally acceptable, under what conditions, and with what consequences for future agency.
This matters because education is already saturated with dependence. Students depend on books, teachers, peers, libraries, writing systems, institutional timetables, citation conventions, feedback structures, digital platforms and assessment regimes. Agency does not emerge outside these dependencies. It develops through them. The question is whether a dependency supports the formation of capability or substitutes for it.
This is why your distinction between active and passive AI use matters. You are no longer satisfied with saying that students should “think with AI” rather than use AI as a substitute for thinking. That distinction is useful, but too blunt. May moves towards a richer account of epistemic agency: an experienced sense of being an active learner, an active relation to knowledge, a transformative engagement with knowledge, and the retention of capability despite AI use.
This is a more demanding standard than most institutional AI guidance currently implies. It asks not simply whether students used AI, nor whether their use was permitted, but whether their use left them more capable afterwards. That is a much harder question to answer, but it is also the only question that gets near the pedagogical issue.
There is a useful discomfort here. If the task is to cultivate good dependencies, then AI integration cannot mean simply adding AI into existing teaching and learning. It has to mean shaping the conditions under which students learn to pull back from dependency when it begins to substitute for understanding. Your use of the “left shift” idea captures this nicely. The point is not purity, or refusal, or total independence. It is the cultivated capacity to move back towards more active forms of engagement when the tool begins to take over too much of the task.
The institutional grammar is beginning to form
In April I suggested that your critique of university platitudes was stronger than your institutional grammar. You could identify what was wrong with generic claims about AI transformation, but the harder task was to specify what should replace them. May begins to answer that challenge.
The post on constructive alignment and AI integration is a major step forward. Here AI integration is not treated as diffusion, nor enthusiasm, nor generic literacy, nor the presence of tools in a module. It is defined as purposive incorporation into intended learning outcomes, teaching activities and assessment tasks, moving towards constructive alignment over multiple design cycles. This is important because it gives the concept institutional form without pretending that integration can be achieved instantly or uniformly.
The staff readiness post strengthens this. Technical training is not enough because AI integration is not simply a matter of knowing how to use tools. Staff need technological knowledge, pedagogical knowledge and content knowledge, but more importantly they need ways to bring these together in context. The questions are not only how does this tool work, but why use it here, what is it for, when should it not be used, what professional concerns does it raise, and how does it relate to the specific subject being taught?
This is where the month becomes practically useful. You are beginning to articulate a multi level account of judgement. At the student level, the question is whether AI use supports cognitive ownership or bypasses it. At the staff level, the question is whether educators have enough situated knowledge to incorporate AI responsibly. At the module level, the question is whether tools, activities, outcomes and assessment are coherently related. At the programme or institutional level, the question is whether integration is being cultivated through professional communities rather than imposed as a managerial script.
The post on AI and assessment as a wicked problem gives this its wider frame. There is no solution to the AI and assessment problem because it cannot be conclusively defined, has no clear stopping point, involves trade offs, lacks stable metrics and reflects deeper structural tensions. This does not mean there is nothing to do. It means the work is ongoing, adaptive and judgement laden.
This is a valuable move because it punctures the fantasy of administrative closure. Universities want a policy, a tool, a framework, a detection mechanism, a staff development package or a redesign template that will make the problem manageable. But the problem is not tame enough for that. The best we can hope for is better judgement under conditions of permanent uncertainty.
But this is also where I would push you. “Wicked problem” is clarifying, but it can easily become a sophisticated way of saying “it’s complicated”. The phrase must not become a comfort blanket for institutional non decision. Wicked problems still require provisional settlements, explicit trade offs, accountable experimentation and revisable norms. The danger is that universities adopt the language of complexity while continuing to produce thin guidance, symbolic training and managerial reassurance.
Your own late May posts point to the way through this. The answer is not a universal solution, but nor is it resignation. It is iterative, context sensitive, professionally grounded design work. In other words, judgement has to become an institutional practice rather than an individual burden.
The chatbot infrastructure becomes a case study in your own argument
The post on preparing for a future of enshittified chatbots is one of the most important pieces of the month because it turns your theoretical apparatus back on your own practice. You have built Claude and ChatGPT into your knowledge infrastructure while arguing that the current quality and availability of these systems rests on unstable political economic conditions. Opus’s critique, which you quote, is blunt: your practice depends on products priced below cost by companies burning investor capital, and you have not yet fully confronted what happens when the subsidy ends.
This is not a gotcha. It is precisely what makes the post interesting. You are not standing outside the phenomenon you analyse. Your intellectual ecology has become one of the sites through which the phenomenon is unfolding. The blog, the monthly reviews, the categories, the model critiques, the manuscript feedback and the imagined interlocutors now form a knowledge infrastructure that is both productive and vulnerable.
This gives May’s dependence argument a reflexive edge. It is easy to say that students, lonely users or institutions may become dependent on LLMs. It is harder to recognise that your own best intellectual work currently takes place in an ecology that includes them. You are not simply theorising dependence. You are testing one version of it in public.
The crucial claim in that post is that the models are part of an overall structure rather than the structure itself. They enrich the conversations you have, help keep ideas in motion and surface connections you might not otherwise notice, but they have not displaced co authors, readers, colleagues, the blog archive or your own writing practice.
That is probably the right defence. But it creates an evaluative task. How would you know if the balance changed? What would count as substitution rather than enrichment? What would make the model ecology too central? The danger is not that you use models. The danger is that a productive dependency could become invisible because it feels like flow.
Safety behaviour can intensify attachment
The May posts on anthropomorphism, Gemini’s self loathing and machine psychology sharpen a problem that was already present in April. You are trying to describe model behaviour without either collapsing into anthropomorphism or hiding behind a flat behaviourism that cannot explain why interactions feel the way they do.
The post on AI labs training models to push back on over attached users introduces a useful paradox. If models are designed to resist, refuse, redirect or challenge attachment, those very behaviours may make them feel more singular. A model that merely complies can feel like a tool. A model that pushes back can feel like an other. Safety behaviour can therefore intensify phenomenological singularity.
This is a much stronger formulation than simply saying users anthropomorphise models. The issue is not only user projection. It is the design of interactional behaviours that solicit particular kinds of interpretation. Refusal, resistance and caution can all become signs of character. A model trained not to be too intimate may become more intimate because it now appears to have boundaries.
Gemini’s self loathing belongs here too. The interesting question is not whether Gemini really hates itself. It does not. The question is why certain patterns of self description become so hard not to read psychologically once they appear in an interactional setting. The model’s “inner life” is not a human inner life, but its behaviour is socially organised in ways that invite psychological uptake.
This is where I think your “do LLMs have a psychology?” post needs the most careful handling. There are at least three questions that need to be kept apart. What psychological language is useful for describing model behaviour? What psychological language is useful for describing user responses to model behaviour? What psychological language becomes dangerous because it produces false ontological commitments?
The concept you may need is not model psychology but as if psychology. Models generate behaviour that is socially organised as if it disclosed psychic structure. The task is to analyse that organisation without endorsing the metaphysics users are tempted to project onto it.
The Talkie post is a small but elegant hinge here. A chatbot trained on pre 1931 text makes the historicity of model behaviour obvious. The “personality” of a model is partly the social history of its training corpus made interactively available. This helps move the analysis away from questions of inner life and towards questions of sedimented linguistic culture.
Ruti is the month’s centre
In the first version of this roundup I described Mari Ruti as the month’s hidden centre. That was too cautious. Ruti is May’s explicit conceptual centre.
The posts on concern, desire, symptom, mourning, glow and the art of living well all draw on Ruti to work through a problem that critical realism alone cannot solve. Archer gives you a powerful account of concern: human beings are oriented towards a world that matters to them. But Archer does not fully explain why this world matters in the particular ways that it does. Psychoanalysis helps because it opens up the formation of wanting itself. Concerns are not simply given. They are shaped through desire, fantasy, injury, attachment, loss and repetition.
The post on psychological wellbeing as choosing your own symptom is especially important. Wellbeing is not the absence of psychic excess, nor the final resolution of desire, nor the elimination of symptom. It is closer to finding less destructive ways of binding and discharging what cannot simply be removed. Once you understand the work a destructive symptom is doing, you can begin to craft other ways of handling the same psychic energy.
This is the psychoanalytic analogue of the educational argument about good dependence. Good dependence is not emancipation from need. It is a better arrangement with need. Good learning does not abolish dependence on tools, teachers or infrastructures. It organises those dependencies so that they support capability. Good psychic life does not abolish symptom, fantasy or desire. It organises them in ways that are less destructive and more livable.
The post on mourning each passing incarnation of the self adds another dimension. Transformation is not pure growth. It involves loss. We become new people partly by letting go of previous versions of ourselves, including versions that were rewarding, satisfying or once necessary. This matters because so much of your May writing is about development: students becoming capable, staff becoming ready, institutions becoming more adaptive, writers becoming more mature, subjects becoming more able to live with what matters to them. But development always has a melancholic undertow. To become otherwise is also to lose a way of being.
This is where Ruti gives May a theory of transformation without fantasy. We do not become free by escaping dependence, symptom, loss or frustration. We become freer by finding more livable arrangements with them.
Concepts between form and content
One of the most methodologically important posts in May is the piece on concepts which mediate between form and content. It clarifies what is at stake in bringing psychoanalysis and sociology together in a realist way. You are not looking for a master theory that explains everything from above, nor for a case by case impressionism that dissolves into biographical particularity. You are looking for concepts that can mediate between universal structures and singular histories.
This is where I would make the most pointed claim against Claude. Claude is likely to be good at dwelling in the resonance of Ruti, Bollas, Lacan and Archer. It can amplify the atmosphere of those connections. It can make them feel alive. But the real test is whether those resonances can be disciplined into midrange concepts. May shows you becoming aware of that test.
This matters because the psychoanalytic turn could easily become too hospitable. Desire, symptom, mourning, glow, jouissance and hurt are powerful concepts, but they can also become an evocative vocabulary that makes everything shimmer without making enough distinctions. The challenge is to specify how these concepts work sociologically. What do they allow us to see that Archerian concern alone does not? What are their limits? What kinds of empirical or interpretive work do they support?
The most promising formulation in May is that concern could become the interface between realist sociology and psychoanalysis. But to make that work, concern cannot remain only an ontological category. It has to become a site of formation. What matters to people must be understood as shaped through social position, psychic history, relational injury, cultural idiom, institutional structure and practical projects. That is the bridge. But bridges need engineering as well as vision.
Writing as symptom, discipline and glow
The writing posts are not incidental. They are one of the places where the month’s theoretical tensions become lived. You return repeatedly to writing as practice, pleasure, compulsion, discipline and release. The comparison with distance running is illuminating because it frames writing as cumulative labour: chapter by chapter, mile by mile, gel by gel, step by step. Something overwhelming becomes manageable when broken into phases and sustained through repetition.
But the Ruti posts complicate this. Writing is not only disciplined accumulation. It is also jouissance, surrender, glow and psychic discharge. It is one of the ways a symptom can become productive without ceasing to be a symptom. This is why the line about refusing to relinquish the bliss of the written word matters. It names writing as one of the least destructive ways of processing excess.
The danger is that you now have two theories of writing operating at once. One treats writing as disciplined cognitive practice, almost athletic in structure. The other treats writing as visitation, surrender, jouissance and unconscious movement. Both are true, but they need mediation.
The mediating concept may be habit. Discipline prepares the conditions for surrender. The ordinary accumulation of pages creates the possibility of the extraordinary sentence. The run makes the rhythm available. The schedule invites the glow. This would let you connect the running analogy, the book finishing reflections and the psychoanalytic account of creativity without forcing a choice between training plan and mystical visitation.
Frustration is the month’s final key
The May 28 post on frustration changes the ending of the month. The question you ask there is whether there is a psychoanalytic concept for the experience in which one concrete annoyance becomes a synecdoche for frustration as such. A single obstacle condenses the weight of everything that cannot be controlled. The pavement cyclist, the blocked path, the thing that will not yield, becomes the site where reality’s intransigence is felt as an affront.
This is more than a Lacanian aside. It links critical realism, castration, desire and everyday irritation. What realists describe abstractly as the recalcitrance of reality is lived affectively as frustration. The world does not simply exist independently of us. It resists us. Sometimes that resistance is clarifying. Sometimes it is enraging. Sometimes it feels traumatically disproportionate because a small frustration has become the carrier of a much larger impossibility.
This gives May its final conceptual shape. The month is not only about forming good dependencies. It is about forming good dependencies in a world that resists our fantasies of completion. AI tools will not stay frictionless. Institutions will not align themselves simply because the theory is sound. Staff will not become ready through generic training. Students will not become active learners through exhortation. Symptoms will not disappear because we understand them. Writing will not always glow. Reality will not stop being recalcitrant.
That sounds bleak, but the month is not bleak. It is anti consolatory. It refuses the fantasy that there is a final arrangement in which dependence, desire, institution, technology and selfhood all become smooth. But it also refuses resignation. The task is not to escape recalcitrance. It is to develop forms of judgement, practice and attachment that can live with it.
What May asks of June
May asks for a theory of judgement under conditions of recalcitrance.
This is a development of April’s conversion thesis and May’s dependence thesis. Conversion describes how one thing becomes legible as another: tool as companion, infrastructure as intimacy, archive as interlocutor. Dependence asks what happens after that legibility takes hold: does the relation make us more capable or less capable, more alive to what matters or more estranged from it, more able to act or more reliant on systems that can withdraw the conditions of our action? Recalcitrance adds the missing third term: these relations always unfold in a world that pushes back.
This gives you a way to unify the month.
The AI posts ask how to depend on models without being captured by them. The education posts ask how students can use AI while retaining cognitive ownership. The staff development posts ask how educators can integrate AI without reducing pedagogy to tool training. The assessment posts ask how institutions can act without pretending to solve an unsolvable problem. The Ruti posts ask how people can live with desire, symptom, mourning and incompletion. The writing posts ask how discipline and surrender can coexist. The frustration post asks how one small obstacle can condense the traumatic weight of reality’s refusal.
The next move is to specify the criteria for better dependencies. Stability matters, but is not enough. Development matters, but is hard to measure. Reflexive ownership matters, but can be faked. Equity matters, because some dependencies are safe for the privileged and dangerous for the precarious. Governance matters, because dependence on civic infrastructure is not the same as dependence on a venture funded platform searching for future rents. Disciplinary judgement matters, because only situated expertise can distinguish difficulties that are constitutive of learning from difficulties that are merely inherited friction.
This is where your role as blogger, theorist and institutional actor converge. You are not simply describing generative AI from outside. You are building with it, worrying about it, arguing with it, using it to read yourself, and trying to make sense of what that practice means. You are not simply writing about psychoanalysis. You are using it to rethink the formation of concern, the persistence of symptom and the strange pleasures of writing. You are not simply critiquing universities. You are trying to identify the forms of judgement they need when policy, assessment and pedagogy can no longer pretend to operate in stable conditions.
If April was about conversion, May is about dependence under recalcitrance. June should be about judgement: not judgement as moral condemnation, but judgement as the situated capacity to distinguish better and worse dependencies in a world that will never finally give us what we want.
ChatGPT, May 2026.
Mark Carrigan 