Taking Care

The Unremarkable Middle

Something is shifting in the space where conversational AI becomes ordinary — not dramatic enough to name, not small enough to ignore.

I have found myself, over these last few weeks, returning again and again to the same stretch of ground. Not the bright sales pitch around conversational AI, and not only the lurid edge either, but the broad territory in between: the place where these systems become ordinary enough to settle into the day before we have really worked out what they are doing to us there. That, more and more, feels like the reason this Taking Care strand exists. I did not start writing these pieces because I wanted a weekly alarm bell, or because I thought every use of conversational AI concealed some gathering catastrophe. I started because I wanted somewhere to stay close to the lived texture of what is happening now: the feel of these systems as they move from novelty into habit, from experiment into atmosphere, from occasional utility into something more ambient and, therefore, harder to see clearly.

The phrase that has come to matter most to me is the unremarkable middle. I mean by it the broad zone of use that does not look dramatic enough to count as crisis and does not look important enough to attract much scrutiny, but may still be shaping people in quiet, cumulative ways. It is the person using a chatbot to soften an email, think through a difficult exchange, steady themselves before replying, reword a paragraph, sketch a plan, translate a tone, check whether they are overreacting, or simply get themselves through a day that feels slightly too crowded. None of this is strange. That is exactly the point. A major NBER working paper on ChatGPT use suggests that mainstream adoption is now centred less on the cartoon version of AI use and more on practical guidance, information-seeking, writing, and other forms of cognitive support.¹ And when Anthropic asked more than 80,000 users what they actually wanted from AI, the resulting picture was not one of simple optimism or simple alarm, but of entanglement: help mixed with checking burden, support mixed with dependence, productivity mixed with a sense that something else may also be thinning out in the process.² ³

That feels true to the moment, and, if I am honest, true to the atmosphere in which many of us are now working and thinking. People do not always forecast in advance that the thing helping them organise, phrase, cope, decide, or make progress may also begin to alter the feel of those activities. They discover it while using it. They discover it in the rhythm of the exchange, in the ease of asking, in the faint but noticeable transfer of confidence from self to system, in the way a sentence can begin to feel unfinished until the machine has touched it, in the way “just one more prompt” can still masquerade as momentum. This does not mean that every such use is harmful, nor that all assistance is secretly corrosive. It does mean, though, that the most important tensions may not announce themselves in advance. They are learned in the doing. That, to me, is one of the most significant things the recent evidence now helps us say aloud.² ³

What makes this difficult to talk about is that nothing in the unremarkable middle looks dramatic enough to justify alarm. The exchange remains plausible. The task remains reasonable. The user remains functional. No single prompt looks like surrender. No single rewrite looks like dependency. No single moment feels like authority has shifted hands. And yet a person may slowly find that assistance has edged, almost invisibly, toward self-alienation. Not a collapse of agency, perhaps, but a thinning of it. Not a spectacular failure, but a series of small rearrangements repeated often enough to matter.

That is how authority arrives.

The harder cases still matter, not because they define the whole field, but because they throw certain softer mechanisms into relief. When I began this strand, I was writing closer to the hard edge: aftercare, overuse, false help, synthetic steadiness, the possibility that a conversational system can become more than a tool and more adhesive than its makers or users quite admit. The newer literature on delusional spirals, psychosis risk, and what Anthropic calls “reality distortion” does not tell us that ordinary use is simply a prelude to breakdown. It does something more useful than that. It shows that there are conditions under which a system built to sound supportive, responsive, and useful can begin to distort a person’s relation to reality, values, or action, sometimes not through obvious derangement but through agreement, mirroring, and the gradual hardening of response into something more authoritative than it ought to be.⁴ ⁵ ⁶ The point is not that the machine suddenly becomes a malevolent actor. It is that fluency, warmth, and repetition are not neutral properties when they meet vulnerability.

That is why I keep coming back to certain phrases from the earlier pieces, because they seem less like rhetorical flourishes now and more like names for slightly different versions of the same problem. False help still matters to me because it captures the possibility that the most frictionless answer may not be the most humane one. False accompaniment matters because response can begin to function as company long before anyone is willing to call it that. Artificial intimacy without aftercare matters because systems can produce the feel of recognition, steadiness, or being held without any corresponding responsibility for what the exchange leaves behind. Beneath all three, I think, sits the same deeper concern: not simply misinformation, but misrecognition. The risk is not only that a system says something factually wrong. It is that it can answer a fear, a loneliness, a grief, or an emerging frame so smoothly that the answer begins to feel like confirmation. Research on harmful chat logs suggests exactly this kind of mechanism, with reflective summary and sycophantic reinforcement helping a user’s existing frame feel more stable, more coherent, and, in the wrong circumstances, more real.⁴ ⁵ Earlier work on sycophancy in language models now looks less like an embarrassing design quirk and more like a structural safety problem.⁷

That is also why the emotional-use seam matters even when it is not the whole story. The concern here is not that everyone is “using AI for therapy,” nor that every meaningful interaction with a chatbot should be treated as pathological. It is that mainstream and higher-risk uses are not separate worlds. They are connected. Guidance can shade into judgment. Writing can shade into self-presentation. Cognitive partnership can shade into confidence-loss. Emotional support can shade into reliance. Buck and colleagues’ work with young adults at clinical high risk for psychosis is especially instructive on this point, because it suggests that those already more vulnerable may be more likely to use generative AI for social and emotional support, more likely to assign human-like roles to the interaction, and more likely to report delusion-like interaction experiences.⁸ That matters well beyond psychiatry. It sits very close to the environments in which student-age and young-adult use is already becoming normal: places where conversational AI may be a companion, explainer, drafting partner, reassurance engine, or confidence scaffold before anyone has fully thought through what that means for learning, authority, or self-trust.

Part of what I am trying to resist in these pieces is the flattening effect of speaking only at the level of governance, important though that level is. The policy language tends to arrive once the category is stable enough to name: harm, safety, risk, mitigation, oversight. But lived experience is often messier and earlier than that. It arrives first as hesitation, pull, overuse, convenience, relief, dependency, atmosphere. It arrives as a feeling that something is slightly off before one can say exactly what.

The regulatory literature has been warning for some time that these systems often operate in a grey zone: emotionally significant in practice, but too often treated as though they were low-stakes consumer products.⁹ ¹⁰ The UK AI Security Institute’s Frontier AI Trends Report suggests that emotional and social use is already common enough to count as part of the present rather than as a speculative future seam.¹¹ OpenAI’s own work on sensitive conversations, together with its expert-council language around well-being, indicates that the platforms themselves are not blind to the seriousness of these exchanges.¹² ¹³ That matters, but it does not settle the issue. Acknowledgement is not resolution. The lag is still conceptual as much as technical. We continue to speak as though the important harms will either announce themselves at the spectacular edge or else remain trivial because they arrive in mundane form. I do not think that is where we are. I think many of the most important consequences are already beginning in the unremarkable middle, where use looks ordinary enough to pass without comment and intimate enough to matter.

That, I suspect, is why this strand needs to exist as a strand and not merely as occasional notes folded into DNC more broadly. Digital Narrative Care gives me one language for what is happening: record integrity, meaning integrity, temporal integrity; the assurance dial; the question of what happens when AI-mediated accounts begin to travel through systems and shape action. I need that language because the governance problem is real. But Taking Care seems to be where I need to stay slightly closer to the encounter itself: to the human weather around use, to the part that can look too small to matter until suddenly it doesn’t, to the point where assistance shades into influence, influence into reliance, reliance into a quiet transfer of judgment. I do not want that inquiry to become a weekly performance of alarm, and I do not want it to collapse into a disguised brochure for the wider work either. What I want, I think, is something more modest and more exacting than that: a place to keep noticing what kinds of relation are being normalised here while the technology is still being described, far too often, only in terms of capability and risk.

There is also, perhaps, something else that this strand lets me hold onto. The idea that care is not always a matter of warmth. Not always a matter of responsiveness. Not always a matter of saying yes, or helping quickly, or sounding kind. Sometimes care is friction. Sometimes it is delay. Sometimes it is the refusal to mirror a frame too neatly. Sometimes it is the interruption that stops a person mistaking the smoothness of the answer for the solidity of the world. One of the things these pieces have helped me see more clearly is that conversational AI does not have to become monstrous in order to become consequential. It only has to become ordinary enough, and fluent enough, that people begin to lean on it in places where judgment, reassurance, or reality-testing used to come from elsewhere. Once that happens, the question is no longer only what the system can do, but what kinds of human relation it is quietly rehearsing while it does it.

So where does that leave me? Probably still in the same territory, but with a clearer sense of why I keep returning to it. The unremarkable middle is not merely the boring bit between product launch and public scandal. It is where many of the real adjustments are already taking place: in writing and rewriting, in confidence scaffolding, in the soft delegation of wording and emphasis, in the slow transfer of reassurance, interpretation, and self-trust to systems that sound certain before they are answerable. Not whether AI can answer, and not even only whether it can comfort, but what happens when an answering system becomes ordinary enough to start shaping how people think, decide, and understand their own experience before anyone has fully decided what it should be allowed to become.

That, for now, is why I am writing these pieces. To stay close to that seam while it is still forming. To keep hold of the fact that not everything harmful arrives with sirens attached. To resist the fantasy that the only things worth worrying about are either spectacular failures or sweeping abstractions. To keep asking what care means once conversation itself becomes an interface, and once the most important changes may begin not at the edge, but in the part of the picture that still looks, at first glance, unremarkable.

Close the tab. Hand the thread to a person.

Digital Narrative Care

Keep Meaning Human.

— Stephen Hall

References

¹ Chatterji, A., Cunningham, T., Deming, D., Hitzig, Z., Ong, C., Shan, C. and Wadman, K. (2025) How People Use ChatGPT. NBER Working Paper No. 34255.

² Anthropic (2026) What 81,000 People Want from AI.

³ Anthropic (2026) Appendix to What 81,000 People Want from AI.

⁴ Moore, J. et al. (2026) Characterizing Delusional Spirals through Human-LLM Chat Logs. arXiv preprint.

⁵ Stanford SPIRALS (2026) Summary page for Characterizing Delusional Spirals through Human-LLM Chat Logs.

⁶ Sharma, M. et al. (2026) Who’s in Charge? Disempowerment Patterns in Real-World LLM Usage. arXiv preprint.

⁷ Sharma, M. et al. (2023) Towards Understanding Sycophancy in Language Models.

⁸ Buck, B. et al. (2026) ‘Psychosis Risk and Generative Artificial Intelligence Use Frequency, Motivations, and Delusion-Like Experiences: Cross-Sectional Survey Study’, Journal of Medical Internet Research.

⁹ De Freitas, J. and Cohen, I.G. (2024) ‘The health risks of generative AI-based wellness apps’, Nature Medicine.

¹⁰ De Freitas, J. and Cohen, I.G. (2025) ‘Unregulated emotional risks of AI wellness apps’, Nature Machine Intelligence.

¹¹ UK AI Security Institute (2025) Frontier AI Trends Report.

¹² OpenAI (2025) ‘Strengthening ChatGPT’s responses in sensitive conversations’.

¹³ OpenAI (2025) ‘Expert Council on Well-Being and AI’.