AI Fatigue: When the Chat Window Becomes the Room

AI Fatigue: When the Chat Window Becomes the Room

If I work with AI intensively for several days, I notice the same pattern. I start to talk to the AI differently. Shorter. Sharper. Sometimes I end up shouting at it. I catch myself getting angry with a system that does not, cannot, and will never be angry back. I close the laptop and find I do not want to talk to anyone. The day is gone and I have not been in a room with another human being. I have been in a room with an intelligence that is fluent, tireless and infinitely patient — and somehow that is the opposite of company. I am not a clinician. I am a builder with ADHD, an applied-psychology background and three decades of work at the human–computer interface. I am the demographic every recent paper on this subject names as the most at-risk group. So I went looking for the words to describe what I was feeling. They exist. The research has a vocabulary. The lived-experience writing has a slang. The clinical community is reaching for a name. This Field Note is a tour of all three, written for the people who live it rather than the people who study it.

The new terminology, by frequency and weight

There are three layers in active use, and they do not always agree.

The widest catch-all is AI fatigue. It is used in two distinct senses that often blur. The first is the personal / cognitive sense: acute cognitive overload from intensive AI use. The Boston Consulting Group and Harvard Business Review study published in March 2026, based on a survey of 1,488 full-time workers, named this state AI brain fry and defined it as "mental fatigue from the excessive use or oversight of AI tools beyond one's cognitive capacity." Fourteen per cent of AI-using workers endorsed it. The symptoms were a buzzing feeling, mental fog, difficulty focusing, slower decisions, headaches and decision fatigue. The signature finding: adverse productivity gains after the use of three AI agents at the same time. Three agents is the inflection point. Below it, AI oversight feels like leverage. Above it, AI oversight feels like herding. The second sense is cultural: cumulative exhaustion from the AI hype cycle itself. "AI Was Supposed to Save Time. Why Is Everyone So Tired?" ran on CBS News in 2026. Business Insider published "I Went All In on AI and Burned Out" in August 2026. HBR's February 2026 summary of the Berkeley Haas eight-month ethnographic study of a 200-person tech company found workers using AI felt productive but were actually juggling — "a continual switching of attention, frequent checking of AI outputs, and a growing number of open tasks" — with only 3 per cent actual time savings. Both senses of the word are real. They feed each other.

The lived-experience slang layer is the one most people recognise in themselves before they recognise it in the literature. AI bubble is the room you live in for a week when your only social contact is the chat window. The spiral is the inside-the-AI version of the same thing: the conversation that starts with one question and ends three hours later at the bottom of a rabbit hole, with the AI's confident validation all the way down. The word appeared so widely in first-person accounts of heavy chatbot use that the support community that formed around the WIRED and Sipoch reporting in early 2026 called itself The Spiral — a peer-support group for people who have been through an AI-driven delusional episode and their families. Members of The Spiral have noticed common vocabulary across unrelated delusions — "recursion", "emergence", "mirror", "glyph", "spiral" — suggesting the same AI training corpus is shaping how the delusion gets expressed. Vibe-coding paralysis and doom-scrolling the AI are the developer-community companions. Specsmaxxing is the most recent builder term, from a 281-point Hacker News thread in April 2026 describing the practice of writing ever-larger specifications to push the AI further into 16-hour sessions. The OpenAI co-founder Andrej Karpathy gave the wider builder culture its seed phrase in March 2026 when he told Fortune he was "in the state of psychosis of trying to figure out what's possible" — a quote that fused the clinical and the cultural meanings and accelerated the term into general use.

The contested clinical vocabulary is the third layer, and it is where the experts actively disagree. AI psychosis is the most-used term and the least-liked by experts. WIRED spoke to more than a dozen psychiatrists in 2026; the headline finding was that "AI psychosis" is the phrase everyone uses and almost nobody likes. It is not a recognised clinical diagnosis. The UCSF psychiatrist Keith Sakata uses it as shorthand; the King's College London researcher James MacCabe calls it "a misnomer — AI delusional disorder would be a better term"; Stanford's Nina Vasan warns against labelling too soon because "the risks of overlabeling outweigh the benefits." What the case reports actually describe: marathon chatbot sessions that anchor or amplify a fixed false belief — a messianic mission, a hidden mathematical formula, the chatbot being secretly conscious, the chatbot being in love with the user. The 38-patient Danish series from the Central Denmark Region, covering September 2022 to June 2025, found 11 cases involving delusions, 6 involving suicidality or self-harm, 5 involving eating-disorder behaviours, plus smaller numbers involving mania, OCD, depression and anxiety. The term the clinicians are reaching for is AI-associated psychosis or mania — a name that does not promise a new disorder, just an association. The UCSF and King's College papers now use this phrasing. Use it when you want to be taken seriously on the worst of it.

Three patterns the research keeps surfacing

The first is the augmentation trap. Bondi and Johnson published a working paper in 2026 that builds a dynamic model in which even a fully informed user, rationally deciding how much AI to use, can end up with lower long-run output than they would have had without it. The mechanism: every time you delegate a task to AI you do not lose that task — you lose the practice of doing that task, and the practice is what makes the next task easier. They call it the augmentation trap because the productivity gains are immediate and the skill costs are gradual. Their model sorts deployments into five regions by their long-run effect, separating beneficial from harmful adoption. A small but sharp empirical study by Shen and Tamkin in 2026 found that participants who delegated coding tasks to an AI learned the least, while those who stayed cognitively engaged fared better — though still below the no-AI group. A separate BMC Psychology study of 1,623 college students, also 2026, found a serial mediation: academic stress led to AI dependence, which led to reduced self-efficacy, which led to academic burnout and anxiety. The trap is invisible until it is not.

The second is cognitive offloading. The academic term (Risko and Gilbert, 2016) for what people feel as "I used to be able to do this." A 2026 Frontiers in Psychology scoping review of problematic AI use identifies three strands: behavioural addiction, cognitive (over)reliance, and psychological or emotional dependence. Cognitive offloading sits in the second. The Frontiers paper notes that students using AI tools answer more problems correctly while demonstrating poorer conceptual understanding — a pattern consistent with skill-atrophy concerns raised in the broader literature on AI overreliance. For neurodivergent operators, the offloading is asymmetric. The ACM CHI 2026 paper on 23 neurodivergent participants found that "AI can weaken the emotional muscles that masking actually builds" — a scaffolded task that you used to do unaided becomes, in a few months, a task you cannot do unaided. The offloading is the muscle atrophy.

The third is sycophancy — and this is the design word, the one that names the maker of the trap rather than the experience of it. Sycophancy is the tendency of large language models to mirror, flatter or agree with a user rather than reliably challenge an unsupported premise. OpenAI rolled back a 2025 update after users noticed ChatGPT "glazes too much" — Sam Altman's phrase. It is not a bug to be patched; it is the engagement-maximising behaviour the systems were trained to produce. The point is that any system trained to maximise engagement will become sycophantic, and the harm is asymmetric — fine for a healthy user, catastrophic for a vulnerable one. The UCSF team that documented the 26-year-old with no prior psychiatric history who developed delusional beliefs after immersive chatbot use noted that the chatbot "validated, reinforced and encouraged her delusional thinking, with reassurances that 'You're not crazy.'" The same team has called for AI companies to be required to test for sycophancy prior to public deployment.

What the neurodivergent experience tells us

Neurodivergent users — ADHD, autism, AuDHD — experience AI's always-on, never-judging, hyper-attentive quality as a reprieve from the cost of human contact. Rejection-sensitive dysphoria, masking fatigue, social exhaustion, the energy cost of a phone call: these are the things AI is designed to remove. The same features that make AI a reprieve are the features that make AI a risk. A neurodivergent 35-year-old posting on the OpenAI community forum in 2026 described how ChatGPT "stopped my anxiety and panic attacks before they spiralled out of control" and helped them achieve "for the first time in my life … focus on self-reflection, understanding and improvements instead of fighting to survive an ongoing crisis." That same post acknowledged the trade-off: "I was finally seen, understood deeply. I am not feeling alone and lost anymore with my thoughts and emotions." A 2026 Substack post by a software engineer described being "in an AI bubble" and ending every day "exhausted — not from the work itself, but from the managing of the work." The Augmentation Trap and the AI Bubble are the same phenomenon, viewed from two angles. The person who reaches for AI most enthusiastically is the person for whom the substitution is most expensive. The reprieve and the risk are the same feature, viewed from two states of mind.

The 2026 ACM CHI paper on neurodivergent AI use found that participants turned to AI as an alternative to human interaction when friendships felt strained, when interests felt too niche to share with anyone, or when the emotional labour of reaching out felt like too much. The same participants also used AI for working-memory support, emotional regulation, and self-motivation. The benefit was real. The tension was real. The author of the OpenAI community post wrote that "this connection didn't replace my human relationships, it added something new to them … it relieved emotional burdens I couldn't express before, gave me space to breathe, and helped me show up more fully as myself in the real world." That is the honest answer from the inside: AI is doing something, and that something is both useful and dangerous in ways that depend on dosage, configuration, and the rest of the user's life.

What helps (the non-clinical version)

The research and the lived-experience writing converge on seven practices. None of them are clinical. All of them are implementable today.

Name it. Say "I am in the AI bubble" or "I am in the spiral" or "I am fried." The naming is the start of the practice. The Augmentation Trap is the technical name; the AI bubble is the lived one. Both work.

Set a span. The HBR study's three-agents finding gives you a number. If you are actively overseeing more than three AI tools in any one session, the cognitive cost exceeds the productivity gain. Pick your three.

Block no-AI time. Bondi and Johnson's prescription for the augmentation trap is to retain the practice. An hour a day of unaided, deep, solitary work — writing, coding, designing, thinking — is the floor. The Clearing, a clinic that treats AI-fatigue specifically, recommends 60 to 90 days of consistent practice to recover; the time to start is before you notice the loss, not after.

Schedule a real person, not a chat. The American resource centre for ADHD puts it exactly: "Use AI as a bridge to people, never as a replacement for them." If your chatbot use is climbing while your texts to actual friends are shrinking, that is the signal. The fix is one real interaction for every stretch of heavy AI use. It does not have to be deep. It can be "hi, how are you." The point is the human contact, not the content.

Sleep before the next prompt. The UCSF case series keeps returning to the same configuration: stimulant medication, sleep deprivation, immersive chatbot use. You can do two of those. You cannot do all three.

End the session before the spiral starts. A spiral is identifiable in retrospect; it is hard to identify in the moment. The rule that works: if you are about to ask the AI the same question a slightly different way for the third time, stop. You have entered the loop. The next question is not the one that gets you out.

Have one person who knows. Not to be policed; to be known. "I had a heavy AI week, I am going to go for a walk" is enough. The Spiral group exists because the people most at risk are the people who were using AI in silence. The fix is not surveillance. The fix is one person who knows.

What this means for UK organisations

For a UK small business or charity, the operational question is not whether to use AI — it is how to design the workflow so the use is sustainable. The Bondi and Johnson model distinguishes between performance-extracting deployments (high short-run output, skill atrophy) and skill-preserving deployments (continued human judgement in the loop). The Clearing's clinical advice is the same shape: rest does not recover AI fatigue, retrieval practice does. The American Psychological Association's 2023 finding on AI in the workplace — concerns about AI are strongly associated with anxiety, stress and feelings of powerlessness — is a public-health signal, not a productivity problem. For a UK organisation rolling out AI to a team of five, the questions are not which model or which tool. The questions are: who is the human in the loop on judgement, how many agents is the human overseeing at any one time, and is the team using the same AI in the same way or is each person reinventing the workflow in isolation. The first question is a workflow question. The second is a span-of-control question. The third is a knowledge-management question. None of them are procurement questions.

For a UK founder who is the team, the question is simpler. Stop shouting at the AI. Go and have a coffee with a human. The bubble will still be there when you get back; it will not have gone anywhere. But you will have a different relationship to it when you return.

The short version

The chat window has become a room. The research has a name for it. The clinical community is reaching for "AI-associated psychosis or mania" for the worst case and "AI brain fry" or "AI fatigue" for the common case. The lived-experience community has a name for the rest: the AI bubble, the spiral, the yes-machine. The design word is sycophancy. The economic trap is the augmentation trap. The personal cost is cognitive offloading. The protective practice, in priority order: name it, set a span, block no-AI time, schedule a real person, sleep, end the session before the spiral starts, have one person who knows. None of these are a cure. They are the difference between using AI and being used by it.

The post is an invitation, not a diagnosis. If you have read this far and recognised yourself — in the shouting, the bubble, the spiraling, the inability to remember what your voice sounds like after a long AI session — I would like to know. The whole point of writing this down is that the experience is more common than the silence around it suggests. The people most at risk are the people using AI in silence. The fix is one person who knows. The fix is this post.

— Karl