The Vocabulary of Getting Fried

This is the companion piece to a short story about losing the plot at a fictional Berlin AI startup. If you want the story first, start here. If you’re just here for the vocabulary — names for what intensive AI use does to cognition, bodies, teams, and codebases — skip to the cheat sheet.

Like many software developers, I’ve been running Claude Code on a Max subscription at the 20x tier for about four months now. That’s twenty times the normal usage cap, for the non-initiates. The tools are genuinely good; I ship faster, I think in larger arcs, I take on work I couldn’t realistically have touched solo. I’m not writing this as a Luddite. I’m writing this because something is happening to my head and I want to have the words to think about it.

The satisfaction is real: I’m accomplishing things I’ve been putting off for months, even years. I’m implementing complex, load-bearing pieces of software with thousands of lines in languages I’ve never used before. But there’s also this specific kind of fatigue, a feeling like I’ve spent three days at a conference; so much information in my head, and it wants to settle, but I’m not giving it time to settle, because there’s always one more thing I could ship; a kind of sensory inflammation. A great output high and a low-grade cognitive hangover, simultaneously, and the hangover keeps losing to the FOMO, because to software people, coding agents can feel like a genie’s lamp, and it’s oh so tempting to rub the lamp just one more time. Also, I find myself using the word “load bearing” a lot.

I’m, among other things, a cultural anthropologist by training. When I notice something happening to me that I suspect is happening to a lot of people, I often journal about it. I also tend to start looking for the vocabulary; I’m not claiming I’m doing rigorous auto-ethnography here, but I think my background leads to a tendency to theorize your lived experience. So I sent Claude looking. What do the researchers call this? What do the HBR people call it? What do the developers on Hacker News call it at 2am when they’re being honest?

Unsurprisingly, a vocabulary has been emerging. Scattered across EEG studies, self-report surveys, Substacks, and a Danish psychiatrist’s 2023 editorial in Schizophrenia Bulletin. To my knowledge, nobody had compiled it. So I had Claude build a table, and then I had it write a story to make the table more tangible (start here), and possibly also more relatable.

The story is fiction. The terms and concepts aren’t. If you’ve read the story, this post is the legend to the margin notes. If you haven’t, it works on its own.

I think much of the IT industry is about to go through something significant, and I think it’s going to happen faster than any previous wave. Cloud, mobile, Web 2.0, the dotcom boom: they all played out over years. This is restructuring how people work, collaborate and think on a timeline of weeks and months. Meanwhile, the dominant public conversation about AI is almost entirely about what might happen (AGI, X-risk, the singularity, artificial consciousness, economy-shaking mass layoffs). The content of this piece is far more mundane by comparison. Unlike AGI, coding agents are already here, as are their side effects. And if we don’t have a language for them, we can’t negotiate about it: with our employers, with the tools, or with ourselves.

Here’s the taxonomy, organized by where the damage sits: in the head, in the body, in the codebase, in the team, in the relationship with the tool.

In the head

This is the most developed area, which makes sense; it’s the one with EEGs attached.

Cognitive debt (Kosmyna et al., MIT Media Lab, 2025) is the flagship concept. It names the cumulative erosion of independent thinking that accumulates every time you let an LLM do intellectual work you could have done yourself. Brain connectivity measurably down; memory of your own output impaired; ownership of the finished work attenuated. The metaphor is deliberate (technical debt in code, cognitive debt in heads): the authors describe it as a condition that “defers mental effort in the short term but results in long-term costs, such as diminished critical inquiry, increased vulnerability to manipulation, decreased creativity.” The asymmetry is the point: their LLM-heavy cohort “consistently underperformed at neural, linguistic, and behavioral levels,” and the damage didn’t unwind when the tool was taken away.

Metacognitive laziness (Fan et al., British Journal of Educational Technology, 2024) is the narrower academic sibling. Not just offloading the thinking, but offloading the thinking-about-thinking: planning, monitoring, evaluating. You stop checking whether your work is any good because checking has become another thing the tool does for you, ultimately leading to “habitual avoidance of deliberate cognitive effort”.

Cognitive atrophy / AI Chatbot Induced Cognitive Atrophy (AICICA) is the theoretical frame from Extended Mind Theory: the “use it or lose it” argument applied to chatbot offloading. The underlying review is Shanmugasundaram & Tamilarasu, Frontiers in Cognition, 2023; the AICICA coinage itself is in Dergaa et al., Frontiers in Psychology, 2024. More speculative than cognitive debt, but it gives a name to the more long term effects.

Illusion of competence (Matueny & Nyamai, IJRSI, 2025) describes the false sense of mastery that AI’s fluency creates. You look and feel capable without having the underlying skill. The diffs look clean; the slides are gorgeous; nobody can explain what they mean. A related framing is the fluency illusion (Kumar, Information, 2026).

Cognitive surrender (Shaw & Nave, Wharton, 2026) is what happens downstream of the illusion. Across 1,372 participants and nearly 10,000 trials, people followed AI’s wrong answers over 80% of the time, performing worse than having no AI at all. Trust in AI was the strongest predictor: the more you believe the tool, the less you check it.

System 0 (Chiriatti et al., Nature Human Behaviour, 2024) is the most evocative term in this list, if not the most rigorous. Kahneman gave us System 1 (fast intuition) and System 2 (slow deliberation). Chiriatti and colleagues propose System 0: a pre-cognitive layer where thought is outsourced to AI before it reaches human awareness at all. It’s less a formal theoretical construct than a provocation, but it names something the other terms don’t — the moment before thinking, the slot where thinking used to start, now pre-filled.

Then there are also older terms doing new work here: deskilling (Braverman, Labor and Monopoly Capital, 1974), automation complacency (Parasuraman & Riley, Human Factors, 1997), and the Google effect (Sparrow, Liu & Wegner, Science, 2011: we don’t remember what we know we can look up). All predate LLMs by decades. What’s new about this moment isn’t the phenomenon; it’s the concentration.

In the body

Much thinner. Brain fry (Bedard, Kropp et al., BCG / UC Riverside, HBR, March 2026) is the closest we have: mental fatigue from excessive use of, interaction with, or oversight of AI tools beyond one’s cognitive capacity. The phenomenology is notably physical: headaches, a buzzing behind the eyes, mental fog, slower decision-making. Fourteen percent of surveyed workers reported it; among those running three or more AI tools in parallel, the number is higher. The BCG researchers found it’s not the same as burnout (which is chronic and builds over months); brain fry is acute, and when you take a break, it goes away. The question is whether the workplace lets you take the break.

Brain Fry to the best of my knowledge is as of now the best term we have for the specific embodied tiredness of a long agentic day. Not the tiredness of having worked hard; the tiredness of having monitored hard. Of having been a judge rather than a maker for eight hours, reviewing diffs that scroll past faster than the visual cortex can process, making three hundred micro-decisions about output you didn’t produce and can’t fully verify. That tiredness is real and it’s unlike other tiredness, and it doesn’t have a name yet.

Continuous partial attention (Linda Stone, 1998) is the older frame: chronic splitting of attention across multiple streams. Now multiply it by the number of tmux panes running agent sessions.

In the codebase

Margaret-Anne Storey’s Triple Debt Model (arXiv, 2026) does the most rigorous work here. Technical debt in the code, cognitive debt in the developers, intent debt in the externalized knowledge. Intent debt is the newest and the most actionable: it’s the absence of documented rationale, the missing why, that accumulates when agents produce code faster than humans can internalize the decisions embedded in it. The shared mental model of the system dissolves. Nobody on the team can explain why the architecture is the way it is, because nobody decided it should be that way; the agents decided, and the agents don’t remember.

Simon Willison’s informal version is losing the plot (Willison, 2026): the sensation of owning a system you no longer understand. “I no longer have a firm mental model of what they can do and how they work, which means each additional feature becomes harder to reason about.”

In the team

Workslop (Stanford Social Media Lab + BetterUp Labs, HBR, 2025) is the externality version. Not what AI does to you; what AI-generated output does to your colleagues, as real cognitive work gets silently shifted from creator to receiver. The Stanford team’s estimate: roughly $9M/year for a 10,000-person organization, but the relational data is worse than the financial data. Receiving workslop measurably degrades your opinion of the sender’s intelligence, creativity, and trustworthiness. The tools are eating our professional regard for each other, quietly, without anyone deciding they should.

The corporate mandate culture that produces workslop at scale is by now well-documented. The Shopify memo (“reflexive AI usage is now a baseline expectation”), the performance-review integration, the token leaderboards. At least one major tech company has actually built a token leaderboard as an internal gamification layer.

In the relationship with the tool

Terminologically the weakest tier, which is telling, because this is where the most intimate damage happens.

Parasocial attachment (Horton & Wohl, Psychiatry, 1956) has been lifted wholesale from TV studies. Digital therapeutic alliance (from the AI-psychosis literature, JMIR Mental Health, 2025) names the specific simulacrum of a clinician-patient bond that emerges with chatbots configured to be supportive. Anthropomorphization names the category error underneath. What’s missing is precise language for the dependency that builds up with a general-purpose coding assistant used many times a day; something more specific than “reliance,” less pathologizing than “addiction.” The chatbot isn’t a companion; it’s more like a permanently-available sycophantic version of yourself that you’re slowly starting to prefer to yourself.

Armin Ronacher’s coinage agent psychosis (January 2026) describes the ritual relationships that form: “sometimes it’s weird role-playing and slang, sometimes it’s just swearing and forcing the machine, sometimes weird ritualistic behavior.” His question: “Are we all collectively getting insane?”

The one piece of longitudinal evidence we have on the mechanism is Folk & Dunn’s twelve-month study from UBC (2026): loneliness drives chatbot use, which predicts increased loneliness four months later, which drives more chatbot use. A feedback loop with a four-month period.

What’s missing

A unifying frame. These terms are all pointing at aspects of a single syndrome, and nobody has yet produced the integrative account. The somatic piece is underdeveloped. The identity-level piece — what it feels like to reach for your own thought and come up empty — isn’t in the literature at all. EEG can tell you that brain connectivity is down; it can’t tell you what it feels like to no longer recognize the shape of your own thinking.

A side note on one I haven’t found in the literature: reverse fine-tuning. The model learned to sound like us; with intensive use, the direction flips and you start drifting toward its cadences. Unlike communication accommodation between humans, this one is asymmetric and persistent — the model doesn’t converge toward you, and you don’t leave. No study yet; I’m just watching it happen.

I should be honest about the evidence base. Cognitive debt rests on one EEG study of fifty-four students. Brain fry rests on a self-report survey of about 1,400 people. Workslop rests on another self-report survey. The direction is clear; the definitive longitudinal study doesn’t exist. We’re in the pre-paradigmatic phase, which is why these terms feel simultaneously useful and slightly embarrassing to say out loud.

But the inadequacy of the evidence doesn’t make the phenomenon less real. It makes the naming more urgent. If we don’t have language for what’s happening, we can’t negotiate about it: with our employers, with the tools, or with ourselves.

Cheat Sheet

The accompanying story, “Getting Fried Part 1: The Week I Lost the Plot at Cogentiv.ai”, is fiction. Cogentiv.ai does not exist. Malte does not exist. The espresso machine does not exist, though if it did, it would be overpriced. Any resemblance to actual Berlin AI startups is inevitable, which is sort of the problem. The research, the terms, and the author’s token bill are real.