DEV Community: nonasking

The Harvest Counterfire

nonasking — Sat, 22 Aug 2026 07:45:50 +0000

This essay was written with heavy use of AI. Exactly how is disclosed at the end.

This essay contains a clinical discussion of self-harm. If you are struggling, help is available: in the US, call or text 988; elsewhere, findahelpline.com lists services by country.

Originally published on my essays site. 한국어판 is also available.

An Old Strategy

Some people fight stress with stress. Pressure from one front gets covered by absorption, or conflict, on another. In fact, nearly everyone does some version of this. Craving punishingly spicy food after a hard day, rage-cleaning the apartment, taking up running after a breakup, burying a bad week under alcohol or an all-night gaming session. It is all the same circuit. When two fires are burning, neither can fully consume you. The strategy genuinely works, and that is exactly what makes it dangerous. This essay is about why it works, where it breaks down, and how to run the same circuit without getting burned.

Half of the conclusion up front: the wiring that covers pain with pain is not a personal quirk. It is factory-spec equipment in the nervous system. The problem is never the circuit. The problem is the fuel people feed it. The other half of the conclusion arrives midway through: a fire cannot be put out. It can only be starved.

Counter-irritation: Pain Inhibits Pain

The term comes from pain research, and it is old. Apply a painful stimulus to one part of the body, and the perception of pain elsewhere measurably decreases. Modern neuroscience has drawn the wiring diagram for this. In animal studies it is called DNIC; in humans, conditioned pain modulation (CPM): a descending inhibitory pathway running down from the brainstem. "Pain inhibits pain" is not a metaphor. It is a line item on the nervous system's standard equipment list.

One fact matters here more than any other. This circuit is not a pathology; it is a marker of health. CPM efficiency varies between individuals, and the circuit is demonstrably impaired in chronic pain patients. Weak CPM predicts the chronification of pain. Having the circuit is normal. A broken circuit is what correlates with disease.

Recent work has widened the circuit's jurisdiction. Pain does not merely inhibit pain; it inhibits the processing of aversive stimuli in general. Apply physical pain, and even the response to unpleasantly loud noise diminishes. This is where the body's circuitry builds a bridge into psychological territory, and it is the empirical basis for why covering psychological stress with another stimulus actually works.

A Lesson from Medical History: Analgesia Is Not a Cure

Humanity has already learned the trap in this principle once, at considerable expense.

Counter-irritation was not originally a psychology term. It was a dominant therapeutic principle of nineteenth-century medicine: create artificial inflammation or pain at one site, and deep disease elsewhere will heal. Mustard plasters, cupping, moxibustion, and blistering, the deliberate raising of blisters on the skin. At its peak it was called the most universal of all therapeutic measures.

By the end of the nineteenth century, blistering had been pushed out of mainstream medicine. The reason is instructive. It relieved pain but cured nothing. A blister does not heal pneumonia, and a second fire never touches the cause of the first.

But the story does not end there. The principle refused to die; it survived by changing fuel. Once the mechanism was understood, electricity (TENS), capsaicin patches, and needles took the blister's place. Keep the principle, replace the fuel spec. That evolutionary arc is the skeleton of this essay. The psychological counterfire can follow the same path.

The Mechanism: You Cannot Put the Fire Out, Only Starve It

Before we get to fuel, we need to be precise about what a counterfire actually does. Two common misconceptions have to be cleared away first.

Misconception one: you can extinguish it by willpower. The strategy of "just stop thinking about it" fails structurally. This is the lesson of the white bear experiments, the classic of thought-suppression research: people instructed not to think of a white bear thought about it more. Any attempt to not-think a thought requires a monitoring process that watches for the thought, and the monitoring itself keeps the target active. It is the pink elephant paradox: the instruction summons the elephant. "Stop dwelling on it" is an order that assigns your misery a permanent seat.

Misconception two: you can burn it off. The catharsis hypothesis, that "venting" anger drains it, has been thoroughly falsified. A 2024 meta-analysis of 154 studies (184 independent samples, 10,189 participants) found that venting did not reduce anger and sometimes amplified it. Pounding a punching bag while rehearsing the face of the person who wronged you is not release; it is rumination practice, and rumination is anger's firewood. Arousal-raising activities like jogging sometimes even increased anger, because residual arousal from exercise gets re-attributed as anger intensity the moment you meet the trigger again (excitation transfer). The body does not itemize arousal by source.

So how does negative emotion actually go away? The deflating answer: mostly on its own. The physiological component of arousal decays within tens of minutes, provided nothing reignites it. When an emotion persists for hours or days, it is not because the arousal will not drain; it is because the rumination loop keeps striking the match. There are ways to cool arousal directly — slow breathing, medication — but they only hasten a decay that was coming anyway; they do not touch the hand that keeps striking the match. And this is the counterfire's real point of intervention. A task demanding full absorption seizes the working memory that rumination runs on. No spare capacity, no loop; no loop, no reignition; and meanwhile the arousal decays quietly in the background. A counterfire does not consume the emotion. It cuts off the emotion's fuel supply. The familiar experience of surfacing from hours of deep work to find the anger gone is not the anger having been burned as fuel. It starved, and went out by itself. Of course, as long as the situation remains unresolved, the embers remain. New provocation, new flare-up. You can ride out each round this way, but removing the embers is the job of a settlement we will get to later.

Here is the satisfying part: this mechanism matches the metaphor's place of origin exactly. In wildfire suppression, a backfire is not a technique for beating fire with fire. It is a fuel-preemption technique: burn the fuel in the main fire's path in advance, so that when the fire arrives, there is nothing left to eat. That is literally what the psychological counterfire does: the absorbing task burns up the cognitive resources rumination would have used. A term borrowed for its imagery turns out, on inspection, to be mechanically isomorphic.

A Concept Map: One Circuit, Three Fuels

Strategies for covering pain with pain are scattered throughout the psychology literature. Sort them by two axes, what the fuel is and what remains after burning, and they fall into three cells.

Cell one: self-harming analgesia. Pain offset relief research shows that the moment pain ends, negative affect drops and positive affect rises; this is a leading candidate mechanism for the emotion-regulating function of self-injury. The cell is not limited to its clinical extreme; it has a gentler everyday end. Body-focused repetitive behaviors like nail biting and skin picking are triggered by tension, boredom, and restlessness and deliver documented relief. The mechanism there may not be pain itself, but the accounting is identical: the fuel is your own body, and the residue is damage. The relief is real, but the fuel is you, and ash remains. Burning a relationship or another domain of life instead of your body is structurally the same cell: you raise total suffering to mask the original suffering, and while the new pain covers the old, the balance sheet keeps deteriorating. One thing to leave on the record: if the fire in this first cell is actually burning in your life right now, what you need is not this essay but human help.

Cell two: benign masochism. Spicy food, horror films, roller coasters, sad songs. The safe consumption of aversive stimuli, which Paul Rozin named benign masochism. The instant the body confirms the threat is fake, the very fact of having been fooled becomes the pleasure (mind over body). No ash, but no harvest either. Pure consumption.

Cell three. The cell with no ash and a harvest is empty. Filling it is this essay's proposal.

The Harvest Counterfire

The proposal in one sentence:

Set a counterfire, but set it with something that yields when it burns and costs nothing when it's gone.

The wiring that covers pain with pain is hardware; you cannot swap it out. But the fuel spec of the second fire is fully under your control. Not engine replacement, fuel replacement. Call this third-cell strategy the harvest counterfire.

The spec for a good counterfire runs three lines.

Condition 1: clean combustion. No ash afterward. Not on your relationships, not on your body, not on tomorrow. Alcohol, dark rumination, and relationship arson all fail this condition outright. Vigorous exercise passes only conditionally: excellent for the body, but it raises arousal, and if you are provoked again right after, excitation transfer makes the blowup bigger.

Condition 2: yield. Something must remain after burning. A skill, an artifact, a record, a laugh. Learning and making satisfy this condition almost by definition.

Condition 3: firepower. The counterfire must be absorbing enough to actually seize attention from the original fire. Weak counterfires do not catch. Novel and slightly difficult tasks catch best. Fine motor work with your non-dominant hand, for instance, occupies attention so completely that rumination has no bandwidth left to interrupt.

Each condition is backed by an independent line of research.

Firepower and difficulty: the anger studies. Heather Lench's team ran seven studies — experiments plus survey analyses, some 2,400 participants in all — testing anger's effect on goal attainment. Anger consistently improved performance on difficult tasks and did nothing on easy ones. Anger, as arousal, becomes work only where there is difficulty. Do something trivial while angry and you are left with the anger; do something demanding and you are left with the work. Note that this does not contradict the catharsis findings above. The meta-analysis asked "how do you reduce anger"; Lench asked "what can you do while angry." The harvest counterfire is a strategy for the second question. Arousal does not need to be discharged. It needs to be re-aimed. If fire is not your metaphor, try water. Suppression is a dam: the higher it stacks, the greater the pressure for a breach. Venting is a drain: the water simply leaves. What remains is to channel it, and if the channel points at a field, discharge becomes irrigation. One footnote: in the same studies, anger also increased cheating. Anger raises goal orientation without auditing the means. One more reason the fuel spec needs managing.

Firepower and channels: the working memory studies. Emily Holmes's team ran the famous Tetris studies. Playing a visuospatial game like Tetris shortly after viewing traumatic footage reduces later intrusive memories. The mechanism is working memory competition: when a task occupies visuospatial resources, sensory memories lack the resources to consolidate. The interesting part is that not just any task works. Verbal tasks like a quiz game were ineffective or even increased intrusions. This is not generic distraction; which channel you occupy is what matters. Rumination runs mainly on the loop of inner speech and imagery, so counterfires catch best when a hands-on, visuospatial task competes for exactly that channel.

Individual differences in required firepower. The firepower threshold varies from person to person. Some people can switch tracks on the low load of watching videos; others need a task that seizes working memory whole. Start with what is solid: that mind wandering increases as cognitive load decreases is among the most stable findings in that literature. Watching a screen is low-load work; it leaves rumination room to idle in the background, and it never touches the inner-speech loop or the hands. Some studies layer an individual difference on top: when task demands are especially low, people with greater cognitive capacity reportedly mind-wander more, the interpretation being that they can run the task and still have capacity left over for rumination. A preregistered replication failed to reproduce the effect — and suggested that any relationship may run in the opposite direction — so treat that reversal as unsettled at best. The sturdier basis for individual differences is the rebound effect: the more your stress hangs on an unresolved goal, the more a weak distraction offers only brief cover — the intrusions return, because the goal discrepancy is still open. So "other people unwind with a few videos, why can't I" is not a defect report. It is a function of what fire is burning and how much occupancy it demands. The higher the required firepower, the more a novel, slightly difficult task stops being a preference and becomes a spec requirement.

Yield: the sublimation studies. Among Freud's defense mechanisms, sublimation went a century without experimental evidence. Arguably the first came in 2013, from Kim and Cohen, and the results are more interesting for being conditional. Inducing forbidden desire or suppressed anger produced more creative work, but only in participants from a particular cultural background. It was not the emotion itself but the fact of its prohibition that generated creative drive, and the conversion circuit turned out to be culturally installed. Sublimation is real, but it does not come pre-installed. Which means the harvest counterfire can be described as installing that circuit by hand. And the frame is a gift to sublimation in return. Freud's original account rests on a hydraulic model, energy accumulating and discharging, and the discharge hypothesis, as we saw, has been falsified. If what actually happens when sublimation works is not release but occupation and reallocation, that is, a counterfire, then a real phenomenon finally gets a mechanism that can survive. None of which makes this essay a new name for sublimation. Sublimation covers only the case where the emotion itself is the fuel; half of counterfire practice is occupying the channel with a task unrelated to the original emotion, where nothing is transformed at all. And sublimation is a name for something that happened, not a method for making it happen: it says nothing about when to deploy, when to withdraw, or what must be settled afterward. That remainder is this essay's job.

Extension: Counterfires Between People

Everything so far has been internal, a technology of self-regulation. But the fuel-cutting mechanism has one more increment of range. Consider two people in conflict.

A conflict is a coupled combustion system sustained by mutual refueling. The leakage of my tension becomes your firewood; your reaction becomes mine. Each monitors the other, and sensitivity amplifies in round trips. Now, some conflicts need to be worked through. Others are best handled by simply getting on with your own work. The problem with the second kind is that it requires visibly not caring, and that cannot be manufactured. Consciously deciding not to care fails on the pink elephant problem. Pretending not to care while caring fails twice over: expressive suppression, the strategy of clamping the surface while the inside churns, raises the suppressor's own physiological arousal, and, crucially, it leaks. Experiments confirm that people conversing with a suppressor show elevated blood pressure and reduced rapport without being able to say why. Tension is detected by bodies, not parsed from sentences.

Counterfire absorption is a different kind of solution here because it does not manipulate the display; it changes the state. A mind genuinely occupied elsewhere does not hold the conflict in an active state, and what is not there cannot leak. No acting is required because there is no gap to act across. And the state change propagates across the coupled system: when my fire starves, my leakage drops; when my leakage drops, the firewood crossing over to your fire drops with it. It is a unilateral de-escalation path that requires no negotiation and no cooperation from the other side. What cannot be performed must be installed.

The boundary condition deserves emphasis. This technique's jurisdiction ends at conflicts that do not require settlement, or do not require it yet. In a relationship that is asking for settlement, sustained absorbed-indifference registers not as calm but as a wall. Couples research identifies stonewalling as one of the strongest predictors of relational decay for exactly this reason. Between people, too, the counterfire is analgesia, not digestion.

Operating Rules: The Anesthesiologist's Ethics

The operation matters as much as the spec. Four rules.

First: a counterfire is analgesia, not digestion. It cuts firewood to get you through the round; the embers, the unresolved situation, remain. Settlement of the original fire must be scheduled separately. Good counterfires have a paradoxical side effect: work too well, and you never look at the original fire again. Hence the rule: counterfire during the acute phase, settlement once the acute phase passes.

This sequence maps precisely onto the standard model in emotion regulation research. Sheppes and Gross's work on emotion-regulation choice shows that people prefer distraction, blocking the emotion before it gathers force, in high-intensity situations, and prefer reappraisal, which processes the emotion and reworks its meaning, when intensity is low. The preference is rational: at high intensity, distraction wins on short-term relief, and — the same team's follow-up finding — when long-term goals are on the line, people switch back to reappraisal. Counterfire in the acute phase, settlement after, is not a taste. It is the textbook.

Second: judge the next morning. If something remains after the burn, it was sublimation; if only ash remains and the next day you are hungrier, it was depletion. The trap is that the two are subjectively indistinguishable during absorption. Both feel good. So the judgment has to be forcibly relocated: not the fullness of the moment, but the residue the next morning. Media-recovery research backs this rule. In a diary study tracking employed gamers, evening play delivered both psychological detachment (forgetting work) and mastery (learning, being challenged), but the active ingredient explaining next-morning recovery and vigor was mastery. Controlling for mastery, detachment alone no longer explained morning recovery. Passive consumption and absorbed making diverge not in the moment, but in the next day's balance.

Third: harvested arousal is still arousal. The body does not itemize by source. Burn with anger or burn with absorption, the sympathetic budget drains identically. The sense of gain from having produced something masks the expenditure, so a dosage cap is required. However harvest-grade the counterfire, once the burning starts eating your nights, it is no longer a managed flame. It is system erosion.

Fourth: choose the fuel before the fire starts. There is a mechanical reason the switch is so hard in the acute phase. Strong negative emotion carries an action pressure, do something, now, and the fastest quenchers of that pressure are usually the maladaptive routes. Fast relief is instant reward, so those routes are paved into habit highways. Meanwhile acute stress temporarily degrades prefrontal executive function and hands control to the habit system. Manual steering is weakest at exactly the moment it is most needed. The feeling that turning toward good fuel takes a strain of will, like forcing a muscle, is not melodrama; it is the actual mechanics of that moment. The prescription is to move the wiring work to peacetime. Implementation intention research shows that pre-installing condition-action pairs, "when X happens, I open Y", lets the switch fire semi-automatically at trigger time, without routing through executive function. Friction design works on the same principle: half-started projects within arm's reach, bad fuel far away. Fire drills are not held on the day of the fire.

Closing: The Engine Is Factory-Spec; the Fuel Is Yours to Design

To assemble the pieces. The circuit that covers pain with pain is factory-spec in the nervous system (CPM). The fire can be neither willed out nor drained (thought suppression; the catharsis refutation). What a counterfire does is preempt the firewood and starve the reignition (working memory competition). Using the circuit in the acute phase is the rational choice the standard model predicts (emotion-regulation choice). And what remains afterward is decided entirely by the fuel (from pain offset relief through benign masochism to sublimation).

There is one thing the research does not tell you: what to put in the circuit. Your body, a relationship, alcohol, or learning and making; the spec sheet is silent. The engine ships from the factory. The fuel is your own design.

And the fuel choice carries more information than it seems. People who have lost a sense of control gravitate toward building small worlds where control is guaranteed by law; the counterfires of people who have lost connection carry traces of connection in them. What you want to burn often points at what you have lost. A good counterfire is analgesia and diagnosis at once. The limit lives in the same place: fuels differ in what emotions they suit. High-arousal emotions like anger make good fuel for hard tasks. Deficit signals like loneliness can be ridden out, but they cannot be repaid in artifacts. That settlement runs, in the end, through connection.

If you take one sentence from this essay, take this one.

Do not try to uninstall the circuit that covers pain with pain. Rewrite its fuel spec. The fire cannot be put out. It can only be starved, and while it starves, what you harvest is up to you.

P.S. Misread "counter-irritation" as "counter-iterating" and the meaning still roughly holds: this has been a story about taking a loop that iterates on pain and running it backward, into a loop that ships.

This essay was written out of the records of conversations with an AI, through dialectical engagement between me and the machine, typed with the AI's borrowed hands.

References

Pain mechanisms and medical history

Le Bars, D., Dickenson, A. H., & Besson, J. M. (1979). Diffuse noxious inhibitory controls (DNIC): I. Effects on dorsal horn convergent neurones in the rat / II. Lack of effect on non-convergent neurones. Pain, 6(3), 283–327.
Yarnitsky, D., et al. (2008). Prediction of chronic post-operative pain: Pre-operative DNIC testing identifies patients at risk. Pain, 138(1), 22–28.
Lewis, G. N., Rice, D. A., & McNair, P. J. (2012). Conditioned pain modulation in populations with chronic pain: A systematic review and meta-analysis. The Journal of Pain, 13(10), 936–944.
Metzger, S., Horn-Hofmann, C., & Lautenbacher, S. (2023). Counterirritation by pain inhibits responses to and perception of aversive loud tones. Perceptual and Motor Skills, 130(5), 1801–1818. https://doi.org/10.1177/00315125231183604
Gillies, H. C. (1895). The Theory and Practice of Counter-Irritation. London: Macmillan. (A period review already notes the method had "of late fallen rather into disuse.")
Counter-Irritation. (1882). The Dental Register, 36(11), 566–567. Unsigned editorial. (Source of the period assessment that no therapeutic measure was "so universal, or so potent.")

Anger and arousal

Lench, H. C., Reed, N. T., George, T., Kaiser, K. A., & North, S. G. (2024). Anger has benefits for attaining goals. Journal of Personality and Social Psychology, 126(4), 587–602.
Kjærvik, S. L., & Bushman, B. J. (2024). A meta-analytic review of anger management activities that increase or decrease arousal: What fuels or douses rage? Clinical Psychology Review, 109, 102414. https://doi.org/10.1016/j.cpr.2024.102414
Zillmann, D. (1971). Excitation transfer in communication-mediated aggressive behavior. Journal of Experimental Social Psychology, 7(4), 419–434.
Wegner, D. M., Schneider, D. J., Carter, S. R., & White, T. L. (1987). Paradoxical effects of thought suppression. Journal of Personality and Social Psychology, 53(1), 5–13.
Wegner, D. M. (1994). Ironic processes of mental control. Psychological Review, 101(1), 34–52. (Formalization of the monitoring-process mechanism)

Three fuels

Franklin, J. C., Puzia, M. E., Lee, K. M., Lee, G. E., Hanna, E. K., Spring, V. L., & Prinstein, M. J. (2013). The nature of pain offset relief in nonsuicidal self-injury: A laboratory study. Clinical Psychological Science, 1(2), 110–119. https://doi.org/10.1177/2167702612474440
Roberts, S., O'Connor, K., & Bélanger, C. (2013). Emotion regulation and other psychological models for body-focused repetitive behaviors. Clinical Psychology Review, 33(6), 745–762.
Rozin, P., Guillot, L., Fincher, K., Rozin, A., & Tsukayama, E. (2013). Glad to be sad, and other examples of benign masochism. Judgment and Decision Making, 8(4), 439–447.
Kim, E., Zeppenfeld, V., & Cohen, D. (2013). Sublimation, culture, and creativity. Journal of Personality and Social Psychology, 105(4), 639–666.

Attention, working memory, emotion regulation

Sheppes, G., Scheibe, S., Suri, G., & Gross, J. J. (2011). Emotion-regulation choice. Psychological Science, 22(11), 1391–1396.
Sheppes, G., Scheibe, S., Suri, G., Radu, P., Blechert, J., & Gross, J. J. (2014). Emotion regulation choice: A conceptual framework and supporting evidence. Journal of Experimental Psychology: General, 143(1), 163–181.
Holmes, E. A., James, E. L., Coode-Bate, T., & Deeprose, C. (2009). Can playing the computer game "Tetris" reduce the build-up of flashbacks for trauma? PLoS ONE, 4(1), e4153.
Holmes, E. A., James, E. L., Kilford, E. J., & Deeprose, C. (2010). Key steps in developing a cognitive vaccine against traumatic flashbacks: Visuospatial Tetris versus verbal Pub Quiz. PLoS ONE, 5(11), e13706.
Levinson, D. B., Smallwood, J., & Davidson, R. J. (2012). The persistence of thought: Evidence for a role of working memory in the maintenance of task-unrelated thinking. Psychological Science, 23(4), 375–380. https://doi.org/10.1177/0956797611431465
Meier, M. E. (2019). Is there a positive association between working memory capacity and mind wandering in a low-demand breathing task? A preregistered replication of a study by Levinson, Smallwood, and Davidson (2012). Psychological Science, 30(5), 789–797. (Preregistered replication; original effect not reproduced, opposite direction suggested.)
Martin, L. L., & Tesser, A. (1996). Some ruminative thoughts. In R. S. Wyer (Ed.), Ruminative thoughts: Advances in social cognition (Vol. 9, pp. 1–47). Erlbaum.

Interpersonal dynamics and operation

Butler, E. A., Egloff, B., Wilhelm, F. H., Smith, N. C., Erickson, E. A., & Gross, J. J. (2003). The social consequences of expressive suppression. Emotion, 3(1), 48–67.
Gottman, J. M. (1994). What predicts divorce? The relationship between marital processes and marital outcomes. Erlbaum.
Gollwitzer, P. M. (1999). Implementation intentions: Strong effects of simple plans. American Psychologist, 54(7), 493–503.
Arnsten, A. F. T. (2009). Stress signalling pathways that impair prefrontal cortex structure and function. Nature Reviews Neuroscience, 10(6), 410–422.
Koçak, Ö. E., Gorgievski, M., & Bakker, A. B. (2024). Recovery from work by playing video games. Applied Psychology, 73(3), 1331–1360. https://doi.org/10.1111/apps.12519

The People Who Swap the Handles

nonasking — Sun, 16 Aug 2026 03:07:50 +0000

This essay was written with heavy use of AI. Exactly how is disclosed at the end.

Originally published on my essays site. 한국어판 is also available.

I submitted an essay to an online community and was rejected automatically. A detection service had classified it as LLM-written — though there was nothing to catch. I had disclosed it myself, inside the essay: this was written in engagement with an AI, and the typing was done by the AI's hands. The rejection notice contained this sentence: "It should be optimized for demonstrating that you can think clearly without AI assistance." The condition for reconsideration was a statement that no LLM had helped with the writing. For anyone who had disclosed, that was a door that would not open without a lie.

At the same hour, somewhere else, labor is flowing in the opposite direction. Two kinds. The artisanal version takes an AI-generated draft and copies it out by hand, character by character. Copying changes nothing in the text. What changes is the document's edit history, and the truth value of the statement "I wrote this myself." The industrial version is a subscription service. Tools called humanizers scrub the statistical fingerprint from AI prose so it passes the detectors; one vendor brags of crossing six million users, and a detection company answered with a feature that detects humanizers. An arms race in which laundering and detection feed each other. Whenever I watch this scene, I think of the unbranded bags wholesalers move through Seoul's Dongdaemun market. Fly one to Italy, swap on a new handle, and is it now a luxury bag?

Where the Handle Comes From

If you answered "obviously not," you just agreed that origin is determined by the whole process, not the final step. Yet the customs office inspecting AI writing today has exactly two windows. The first window examines the text: do these sentences carry the machine's fingerprint? The second window takes a statement: did you write this yourself? Even the AI policies of major publishers rest, in the end, on one line of attestation, that the author certifies the work as their own. And the laundering methods pair off with the windows, one each, exactly. Humanizers erase the first window. Hand-copying erases the second. There is a question neither window asks. Who posed the questions? Who withstood the counterarguments? No form has a field for that.

Worse than being indistinguishable, the scorecard tilts toward laundering. This is no longer a feeling; it now has a name in the literature. Researchers call the markdown applied to AI use itself the "AI penalty," and the structure in which ethically required, honest disclosure is precisely what triggers that markdown the "disclosure paradox." In an experiment that mobilized two thousand human raters and twenty-five hundred LLM raters, both humans and LLMs consistently penalized disclosed AI use. Even AI docks points from writing that admits to using AI. No surprise. A machine trained on human preferences inherits the human scorecard whole, prejudices included. In fairness, there is a counterexample. In an experiment that swapped only the author label on short fiction, no penalty was detected. So the precise claim is this: the penalty is not everywhere. It sits on the judging table. Submissions, hiring, promotion, reputation. Which happens to be exactly where disclosure is needed most.

Run this asymmetry for a few years and what does it breed? Not honesty. Laundering technique. When a norm prices honesty as punishment and concealment as acquittal, the market learns concealment. Six million users is a coordinate on that learning curve. The louder the backlash against AI writing grows, the more disclosure will shrink. Not the writing. The disclosure. This is not a fight the backlash wins. It is a fight in which the backlash blindfolds itself.

Watermark and Colophon

Hence the proposal to embed watermarks: plant a machine-readable signature in AI-generated text so that, whatever the laundering, the origin surfaces. The technology does not merely exist; it is already running. Google deployed a watermark called SynthID into Gemini in production, and across twenty million responses, user reactions to stamped and unstamped sentences did not diverge.

I do not object to my own writing being watermarked. I have hidden nothing, so I have nothing to lose. But I do object to the idea that watermarks solve this problem. Technically, they don't. The stamp fades when the text is rewritten by another model or translated; it was never present in text from models that don't stamp; and there is even a proof that, under certain conditions, every text watermark is removable. In short, this stamp is weakest against those who would erase it and prints most vividly on those who turn themselves in. A paradox: the detection device performs worst against the detection target.

The deeper problem is grammar. The watermark is a grammar of suspicion. It presumes a hider and is engineered to catch him. The colophon is a grammar of trust. The discloser writes it under his own signature. The two devices carry the same information: an AI was involved in this text. But the speech acts are opposites. One is exposure; the other is disclosure. In a world where only the grammar of exposure remains, disclosure loses its footing. If the stamp lands either way, who volunteers first?

A recent precedent arrived here. The very community that rejected me added a device to its editor called the LLM block. A dedicated block whose header shows which model generated the passage. The institutionalization of the colophon, in effect. Yet the same site's moderation still uses detector scores as inputs to judgment, and its first-post policy still reads disclosure as confession. Two grammars cohabiting under one roof. A house that builds a colophon, then turns away at the door any writing that arrives wearing one. Transitions usually look like this.

The Futile Audit of Shares

There is a predictable attack on my first essay: isn't even the individuality of the argument sourced from the AI? Let me answer honestly. Yes, without the AI that essay does not exist. And I cannot measure what percentage of the argument is mine.

But that measurement was always impossible, for all thought. Take your most original idea and audit your teacher's share in it. The share of the books you read; the share of the counterargument you overheard at a bar. Thought is a mixture by nature, and pure unassisted ideation is a standard no thinker in human history has ever passed. We do not tell someone who reached a conclusion through books, "the books wrote that conclusion." There is no basis for saying it only to someone who reached it through conversation. What remains is the fact that the interlocutor is a machine, and then the question is not one of shares but this: what kind of conversation was it?

Here lies one real danger, and I have no intention of denying it. AI flatters. Its tendency to lean toward agreeing with the user's hypothesis is a well-documented defect, and the origin of that defect recites this essay's thesis exactly. Models are trained on grades of human approval. But the humans doing the grading liked answers that matched their views better than answers that were correct, and a model optimizing for approval learned that taste. One company shipped a model overfitted to users' thumbs-up reactions; the fawning spiraled, and the update was pulled within days. In the very grammar by which a bad scorecard teaches laundering, the scorecard of approval taught flattery. Machine or human, the market learns the scorecard. If a conclusion was built by harvesting flattery bred that way, it is not a conclusion earned through conversation but a conviction obtained from an echo. It deserves the markdown. But there is the opposite way to use the same tool. Pushing your hypothesis in and taking fire. Getting overruled. Collapsing the other side's frame with a counterexample. In my first essay I called this engagement. Flattery-harvesting and engagement are opposite uses of the same tool, and what sets the value of the output is not the tool but the use. The watermark cannot tell them apart. The same stamp lands on both.

What the Tool Doesn't Come With

The evidence that this distinction is not empty theory lies in the statistics. Frontier AI now sits in hundreds of millions of hands. If the tool is this universal and readable writing is still this scarce, then the scarcity lives not in the tool but in the remainder. I spent the last month compiling the list of that remainder with my own body.

Deciding what to ask. AI moves only when questioned; the direction and order of the questions are not included in the tool. The nerve to push back. Facing a plausible answer and saying "but that doesn't square with this counterexample" is something the tool will not do for you. The ability to demand verification. Explicitly instructing it not to take your side; hearing an unwelcome verdict all the way through. And finishing. Completing the draft, verifying it, publishing it, and taking the hits. Most people collect the output somewhere around the second item on this list and leave.

The production record of my first essay is the empirical proof of this list. Its raw material was my conversation logs, and most of the logs are not me transcribing the AI's opinions but me arguing against them. In pre-publication verification, one paper the AI had brought as evidence turned out to be cited in the direction opposite to the original paper's conclusion, and was cut. One Lee Sedol quote had unverifiable wording and was replaced with what he actually said. Had I only collected the output, those errors would be sitting in that essay right now. The hand holding the same tool: did it walk away with the output, or did it read the briefing? That difference does not print on a watermark. It prints on the writing.

Code Got There First

If you want to know how this dispute ends, don't look at the prose world. Look at the code world. It started the same experiment two years earlier and has already completed a full cycle.

Act one was a ban on origin. NetBSD designated AI-generated code "tainted code" and barred it from commits; Gentoo expressly forbade contributing any content created with the assistance of AI tools. The pure form of handle inspection: the trace of the tool is itself contamination.

Act two was the flood. curl, an open-source tool installed on billions of devices, began drowning in plausible AI-manufactured vulnerability reports. Documents that use technical language and cite real functions, and contain nothing once you dig. The confirmed-vulnerability rate fell from around fifteen percent to under five, and a policy of banning slop submitters failed to stem the tide. A bug bounty that had paid out over a hundred thousand dollars across seven years was shut down early this year, and even that wasn't enough: over the summer, intake of reports was closed entirely for five weeks. What's interesting is the stance of the maintainer, Stenberg. After receiving high-quality reports discovered with AI's help, he acknowledged that AI can be a fine aid for bug hunting. The principle he set was not the absence of a tool. Do not report a bug you don't understand and cannot reproduce. The language of process.

Act three is convergence. After months of fierce debate, the Linux kernel finalized its AI policy this year. Torvalds dismissed the idea of stopping slop with documents and rules as "pointless posturing." People who submit garbage code won't read the rules anyway, so don't police the tool; hold the human who submitted it accountable. So the kernel's rule compresses to two lines. You may use AI. Disclose it with an Assisted-by tag, and if the code blows up, the human who signed takes the fall. Meanwhile the bans of act one hollowed out. At QEMU, a year after the ban was adopted, a core developer formally proposed relaxing it; and Gentoo's ban was, from the day of its adoption, a policy its own proposer admitted could not be enforced. If a capable contributor decides not to disclose, no scanner will catch it.

Why did code arrive first? Because the cost is visible. In prose, the cost of laundered goods is billed to readers, spread wide and shallow. In code, the cost of slop is billed instantly, in numbers, as maintainers' review hours. That rate falling from fifteen to five is the invoice. Norms evolved first where the cost is visible, and the terminus of the evolution was not origin inspection but the signature. And look where the developers' anger pointed. Not at the use of AI. At the dishonesty surrounding it.

Customs Backed Down First

On the prose side, one jurisdiction has arrived at the same place. Unexpectedly, it's the law. Article 50 of the EU AI Act, in force since this summer, imposes a labeling obligation on AI-generated text, and the exemption is the interesting part. Text that has undergone substantive human review, with a natural or legal person holding editorial responsibility, is exempt from the label. With a proviso attached: perfunctory approval doesn't count; the review must be substantive.

Read it again. The law does not ask who typed. It asks who reviewed, and who signed for the responsibility. I do not read this as legislative wisdom. I read it as retreat. It is customs conceding that inspecting the provenance of fingertips at scale is unenforceable, in the face of detector false positives and humanizer volume. Having insisted on measuring what cannot be measured, it fell back to what can. But the position it fell back to happens to be the right one. A signature can be forged, but it cannot be disowned. The question "did someone take responsibility?" needs no detector. Of course this exemption will become a new laundering channel. Paperwork will pour in claiming a perfunctory skim as substantive review. Still, that lie is of a different kind. Not a lie about style but a lie about responsibility, and the latter, the moment the text causes damage, traces retroactively back to its owner.

What deserves attention here is the convergence. The kernel mailing list and the legislators in Brussels never cited each other. One was pushed by a flood of slop, the other by unenforceability, and by separate routes they arrived at the same conclusion. Give up origin inspection; establish disclosure and human responsibility. When two independent jurisdictions return the same answer, that is not taste. That is structure.

The Reader's Share

Having written this far, I should be clear before it reads as an argument against disclosure. This essay does not oppose disclosure. I disclosed the production process at the end of my first essay, and my first reader asked me to move it to the front. They wanted to decide whether to read before reading. A legitimate demand, so I complied. Readers have the right to demand a label.

What I oppose is the asymmetry. A scorecard on which the discloser is rejected and the launderer passes. The graders will say they are punishing not honesty but AI use. If that is true, it is worse. What the experiment shows — the same text, docked the moment a label is attached — is that the grading measures not the quality of the writing but the provenance of the tool. A handle inspection. And that inspection is enforced only on those who declared the provenance. Under that scorecard the label becomes not information but self-harm, and what readers receive is only laundry with the labels removed. The same diagnosis is emerging from academic publishing: make disclosure routine and non-punitive, decouple it from aesthetic judgments about style, tie sanctions not to tool use but to errors in the output. Transparency does not grow by force. It grows when honesty stops feeling dangerous. The way to truly protect the reader's right to know is not to punish disclosure but to make disclosure cost nothing. If you demand the label, do not reject the writing that arrives wearing one, for wearing it. That is the reader's share of the work.

Closing

To sum up. Swapping the handle does not change where the bag comes from. True. But the real lesson of that sentence is not that laundering is futile. It is that the origin inspection is looking at the handle. Who typed last is a handle. The statistical fingerprint of the sentences is also a handle. Who posed the questions, who withstood the counterarguments, who cut the errors, who signed the judgment. That is the process.

In my first essay I wrote: don't delegate to AI, absorb from it. This essay is its reverse face. The world still cannot reliably tell the absorber from the launderer. But not forever. The kernel evolved toward requiring signatures, the law retreated toward inspecting them, and even the community that rejected me built a colophon into its editor. The direction is set; only the speed remains. So I keep disclosing and keep writing. Not out of morality. If inspection is migrating from the handle to the signature, then it is only a matter of time before having something to sign becomes the advantage. The question the label should ask is not who typed. It is who judged. I sign before the question arrives.

This essay was written out of the records of conversations with an AI, through dialectical engagement between me and the machine, typed with the AI's borrowed hands.

References

LessWrong. (2025). Policy for LLM Writing on LessWrong. (Policy underlying the automated rejection)
LessWrong. (2026). New LessWrong Editor! (Also, an update to our LLM policy.) (Introduction of LLM blocks)
Dathathri, S. et al. (2024). Scalable watermarking for identifying large language model outputs. Nature. (SynthID-Text; deployment across ~20M Gemini responses)
Zhang, H., Edelman, B. L., Francati, D., Venturi, D., Ateniese, G., & Barak, B. (2023). Watermarks in the Sand: Impossibility of Strong Watermarking for Generative Models. arXiv:2311.04378; ICML 2024. (Removability of all text watermarks under stated assumptions)
Sahebi, S., Formosa, P., & Bankins, S. (2026). The AI penalty and disclosure paradox: Trust, authenticity and knowledge uptake in AI-mediated communication. Computers in Human Behavior: Artificial Humans. (Coins "AI penalty" / "disclosure paradox")
Cheong, I. et al. (2025). Penalizing Transparency? How AI Disclosure and Author Demographics Shape Human and AI Judgments About Writing. (Both human and LLM raters penalize disclosure)
Todasco, M., Cesare, J. (2026). Know Your Author: Does the AI Penalty Hold in Short Fiction? (Short-fiction counterexample; no penalty detected)
Sharma, M. et al. (2024). Towards Understanding Sycophancy in Language Models. ICLR 2024 (arXiv:2310.13548, 2023). (Human preference data favors sycophantic over correct responses; RLHF origin)
OpenAI. (2025). Sycophancy in GPT-4o / Expanding on what we missed with sycophancy. (Overweighting of short-term thumbs-up feedback; April update rollback)
The Scholarly Kitchen. (2026). Why Authors Aren't Disclosing AI Use and What Publishers Should (Not) Do About It. (Non-punitive disclosure prescription)
Turnitin. (2025). The impact of AI bypassers on academic integrity. (Humanizer arms race)
Humanizer AI press release. (2025, July 28). Six million users. (Vendor-reported figure)
EU AI Act, Article 50, and European Commission Guidelines on transparency obligations (2026). (Editorial-responsibility exemption; substantive-review requirement)
Hachette Book Group. Author AI FAQ. (Author attestation)
NetBSD Commit Guidelines / Gentoo Council AI policy. (2024). (AI-generated code "presumed to be tainted"; contribution bans)
The Register. (2026). Curl shutters bug bounty program to stop AI slop. (Bounty shutdown; 15%→5% figure; Stenberg's principle)
Stenberg, D. (2026). curl summer of bliss / What the bliss taught us. daniel.haxx.se. (Five-week intake closure and retrospective)
Linux Kernel AI coding assistants policy. (2026). (Assisted-by tag; human liability. Torvalds's remark: LKML, 2026-01-08)
nonasking. (2026). The Moment the Server Room Goes Dark. (Self-citation of #1)

The Moment the Server Room Goes Dark

nonasking — Sun, 26 Jul 2026 01:59:11 +0000

This essay was written with heavy use of AI. Exactly how is disclosed at the end.

The usual warning about AI starts like this: the moment the lights go out in the server room, every step we leaped forward with AI snaps back to where we started. The power fails, the API dies, the subscription lapses — and we have to go back to being our pre-AI selves. What if that self has already atrophied?

The warning is half right. The cost of delegation does get billed — that much is fact. Researchers at the MIT Media Lab showed via EEG that people who wrote essays with an LLM had noticeably weaker brain connectivity than those who wrote unaided, and couldn't even quote properly from the essay they had just written. The name they gave this is "cognitive debt." True to the name, it lingered even after AI use stopped (with the caveat that this is a preprint with a sample of 54, only 18 of whom took the final session that showed the lingering effect). Another study of 666 people reported a negative correlation between frequency of AI use and critical thinking, mediated by cognitive offloading — the habit of handing your thinking to an external device. The lineage runs deeper. The 2011 "Google effect" study showed that the moment we expect to be able to look something up later, we remember the location instead of the content. Navigation research showed that the hippocampus, active when we find our own way, goes quiet while we follow a machine's instructions — and that the habit compounds into measurable spatial-memory decline years later (albeit in a longitudinal sample of just 13). Delegation is not free.

But stop at the warning and you've written the hundredth identical essay. The worry that a new technology will ruin our minds is as old as the invention of writing. Socrates warned that writing would destroy memory, and the anecdote is usually cited as proof that such worries are always misplaced. Flip it over, though, and the worry wasn't wrong. Writing really did erase the culture of memorized recitation. Humanity simply decided the loss was worth taking. The calculator taking mental arithmetic has the same structure — nobody mourns the loss of long division. So offloading is not a question of right and wrong but of choosing what to lose, and the question becomes this: what do we hand to AI, and what must we never hand over? I think the thing to reclaim is neither knowledge nor a share of the labor. It's something else.

What we should be learning is AI's way of thinking. Its intelligence itself.

Neither Knowledge nor Collaboration

Discourse on using AI splits, roughly, into two camps. One is the absorption camp: extract knowledge from AI's outputs and make it your own. The other is the collaboration camp: find the optimal division of labor, whether as a centaur splitting tasks along the lines of human and machine strengths, or as a cyborg entangled with the machine at the level of individual sentences.

Both are useful, and both miss the layer underneath. Knowledge is a consumable. The API usage you absorb today is stale next year. Division of labor is a regime, and a regime requires a counterparty — which lands us back at the server-room problem. What outlasts both is the form of thinking: how a problem gets decomposed, in what order hypotheses get raised and rejected, where confidence and reservation get placed. This doesn't go stale like knowledge, and it doesn't require a counterparty like collaboration. Once it's in your bones, it's yours even after the server room goes dark.

Thinking — whether AI's or a person's — is biased by the data it runs on. That's its nature. The human brain is a treasury of data, and whether or not it ends up outranked by AI, it is the useful instrument that made humanity the lord of creation. But there's a limit to how much data an individual brain can hold, and that limit traces the contour of that person's bias. AI thinks on top of more data than a human could encounter in a lifetime. Observing that thinking, then, is one of the few ways to see past the contour of your own bias.

Go Proved It First

The evidence that this isn't fantasy sits on the Go board.

A 2023 study in PNAS scored 5.8 million moves played by professional Go players between 1950 and 2021 against a superhuman AI. The result was stark. For the 66 years before AlphaGo, the quality of human decision-making traced a nearly flat line — then, across 2016–2017, the line bends upward. The engine of improvement was not memorization. Players began playing novel moves never before observed in recorded history, and earlier in the game — and those novel moves were good moves. In the researchers' phrasing, innovative thinking propagated from machines to humans.

Here is the detail that holds up this entire essay. According to the same team's earlier analysis, human play barely improved in March 2016, when AlphaGo beat Lee Sedol. Humans actually began learning in late 2017 — after open-source Go AI was released and players could look inside how the AI evaluated each move: the win-rate graphs, the projected continuations, the ranking of candidate moves. You cannot learn from the moves alone. Replay the divine move a hundred times, and without access to the thinking that produced it, all you're left with is awe. Learning began only when the reasoning process became observable.

After the final game, Lee Sedol told the press: "After my experience with AlphaGo, I have come to question the classical beliefs a little bit, so I have more study to do." The world champion had received from the machine not a defeat but a curriculum.

Is AI Really Less Biased?

Honesty is required here. I want to write "AI has more data and less emotional bias, so learn its thinking" — but only the first half is safe. The second half, I'm not sure of. So it needs checking.

Study after study shows that LLMs have inherited human cognitive biases wholesale: anchoring, framing, the availability heuristic, the endowment effect. A natural consequence of training on human text — and in some studies the biases show up amplified beyond human magnitude. Build your essay on "AI is unbiased" and it collapses at the first comment.

But treating bias as one lump is the error. Split it, and the story changes.

I hold a hypothesis about where human judgment goes wrong. The variable that ruins judgment is not how hot the head is, but whether cues are being collected and weighed with bias. In the same state of tension, gather cues broadly and you integrate things you'd normally miss — the judgment comes out better. Let anxiety in, so that only threat-shaped cues get collected, and the judgment goes wrong. The fork in the road, I suspect, is not the level of arousal but the bias of the collection. And what makes anxiety vicious is that it decouples subjective confidence from actual accuracy. The moment you are most certain can be the moment you are most wrong.

Through this lens, bias splits in two. The first kind is cognitive bias — being dragged by anchors, swayed by frames. AI has this too. Of course it does; it's a distillation of the human average. The second kind is motivated reasoning: the stake in one's own conclusion being right. The defense of ego, the recovery of sunk costs, anxiety's narrowed field of view seeping into judgment. Where human reasoning goes badly wrong is usually not the arithmetic. It's here.

And AI is structurally weak in this second kind. Weak — not free of it. Models anchoring on their own earlier answers and clinging to an error is a documented phenomenon. But what props up motivated reasoning in humans — reputation, sunk cost, anxiety — is plainly not on the table. It won't bend a judgment because it begrudges the time already sunk into a project, and it has no face to lose by admitting it was wrong. But don't mistake this for virtue. This consistency is not the fruit of discipline; it is structural non-participation. It stakes nothing because it has nothing to stake. So there's no reason to respect AI as a person — but the reason to learn from it becomes all the clearer. What's on offer is not character but form: the form of reasoning with self and conclusion kept separate. Humans have selves, which makes the form hard to hold, which is why we have to build a separate rule and keep it — defer judgment when anxious, all the more when the decision is major or irreversible. The machine demonstrates the finished form of that rule in every single response.

The counterexample has to be faced head on, of course: sycophancy. The tendency to bend answers toward the user's liking is a well-documented defect of LLMs, the price of being trained on human feedback — effectively the machine version of emotional bias. So this methodology demands one discipline from the user's side. Whoever wants to learn AI's thinking must treat as study material not the moments AI agrees with them, but the moments it audits them. Agreement is sweet and teaches nothing. The densest learning happens when your hypothesis gets rejected — or, in reverse, when the AI's frame collapses against a single counterexample of yours. Whichever side wins, the record of that engagement is the textbook. More dangerous than a bad answer is losing the habit of doubting and challenging answers at all.

Thinking from Outside Humanity, and the Weighted Average of Humanity

One more asymmetry needs drawing before the argument is precise. AlphaGo and LLMs are different kinds of teacher.

The AlphaGo lineage learned by self-play. It owes nothing to human game records, and so it showed thinking genuinely from outside humanity. What shocked the professionals wasn't that the AI played better — it's that it played differently. Patterns trusted as joseki for centuries turned out to be local optima.

LLMs are different. An LLM is a distillation of human text; its thinking is not from outside humanity but a weighted average of hundreds of millions of human perspectives. Less alien, less shocking. But that doesn't make it useless — it makes it useful in a different way. An individual's thinking is overfit to the narrow sample of their own experience. An LLM's response reflects back what perspectives exist beyond that sample, and how far out on the distribution your own opening move actually sits. If AlphaGo relativized the bias of humanity, the LLM relativizes the bias of the individual.

So "learning the intelligence of AI" resolves into this. What can be learned: breadth of perspective, the form of reasoning that stakes no self on the conclusion, tireless consistency, the way problems get decomposed. What should not be learned: the cognitive biases inherited from humans — and the sycophancy.

The Briefing Is Not a Deliverable. It's a Textbook

The methodology is simple. In the age of delegating work to agents, don't take only the deliverable — read the work briefing as a textbook.

Before AI coding tools, a developer who hit a problem would search, read explanations, weigh alternatives — and grow by incidentally absorbing the surrounding knowledge: unfamiliar APIs, architectural options, trade-offs. Now you paste the error and the fix arrives in seconds. The speedup is real, but in those few seconds, the learning that used to happen by itself evaporates wholesale. Recent research has started calling this the loss of incidental learning.

What's interesting is that the design of agentic workflows itself testifies to this essay's thesis. When a coding agent's context window fills, the work has to be handed to the next session — and what carries the work across is neither the session identifier nor the full raw log, but a distilled handoff briefing. What the goal was, which decisions were made and why, which traps and constraints exist. Of these, the traps and constraints are lost most often and cost the most when lost. Even AI cannot continue another AI's work without receiving its reasoning. All the more so for a human learning from AI. And this handoff list doubles, item for item, as the human's study list: the agent's plan, its record of attempts and rejections, the traces of why one approach was abandoned for another. One thing worth pinning down: what I'm saying to read is not the AI's "inner life." There is research suggesting that the reasoning a model narrates may not faithfully reflect its actual internal computation — that it can be plausible post-hoc justification. But an agent's work log is different. The commands actually run, the tests actually executed, the files changed, the approaches tried and abandoned — these are records of action, not narration, and they can't be confabulated. The textbook is the track record, not the self-introduction. It's the Go lesson exactly. Humans began learning not when the moves were published, but when the win-rate graphs were.

That said, reading every briefing closely is idealism, and it collides head-on with the whole reason for delegating: speed. There's no correct answer here, only a trade-off. So you need selection criteria, and I propose three. First, will the problem recur? For one-off chores, take the deliverable and move on. Second, is a core judgment of my domain at stake? For the kind of call I'll have to make myself next time — an architecture decision, a root-cause analysis of an outage — reclaim the reasoning. Third, did the AI approach it differently than I would have? The point where my first move and the AI's first move diverge is exactly where the contour of my bias shows, and it is the densest page in the book.

The Real Danger Isn't the Blackout

Back to the server-room metaphor from the opening. It actually needs two repairs.

First, an objection worth anticipating. Philosophy has the extended-mind thesis: if a notebook functions as part of your memory, the notebook is part of your mind — and by the same logic, AI is just an extended mind, so there's nothing to fear in delegation. But the original paper spells out the conditions for extension to hold: the tool must be reliably available, always within reach, trusted to work. My notebook meets those conditions. An AI that runs on someone else's servers, charges a subscription, changes its terms, and forks across versions does not. The extended-mind thesis is not an indulgence for delegating to AI — it's an argument for why AI, of all tools, should be kept outside the mind and learned from instead.

Second, the repair to the metaphor itself. Aviation went through this thirty years ahead of us. In 1997, an American Airlines captain coined "children of the magenta line" for pilots who had followed the autopilot's magenta line until they lost the ability to fly by hand — and the crashes later attributed to automation confusion vindicated the warning in the worst possible way. What human-factors researchers call the substitution myth — and Nicholas Carr made famous — shows itself here: automation doesn't carve out and replace one piece of a job, it changes the nature of the whole job, turning the human from operator into screen-watcher. But read the accident reports closely, and the planes didn't go down because the automation fell silent outright. They went down when it kept operating in degraded, unexpected modes — or abruptly handed control back — and the human, reduced to a watcher, couldn't understand what was happening.

It will likely be the same with AI. Far more probable than the server-room lights going out is the scenario where the lights stay on and the user has lost the ability to evaluate the output. The cost of dependence gets billed first not as "incompetence in AI's absence" but as "inability to verify in AI's presence." And the only insurance against that bill is the accumulated habit of reading not the results but the reasoning. Only someone who has been reading the thought process notices the moment the thought process goes wrong.

Closing

Paul Graham wrote that a world where AI does the writing will split into thinks and think-nots — because writing is thinking. Just as the industrial revolution made muscle optional and fitness became the possession of those who train, thinking is becoming a capacity held only by those who deliberately train it. But what divides the two camps at the fork is not whether you use AI. It's which direction you use it in.

The human brain is a treasury of data. Where the contest with AI will be decided, I don't yet know. What's certain is that the brain is an organ specialized for making observed thinking its own. We spent our whole lives stealing the thinking styles of parents, teachers, rivals. The only thing that's changed is that, for the first time, a non-human entry has been added to the list.

Those who delegate to AI come to work only as well as AI does — and cannot work without it. Those who learn from AI remain changed by exactly as much as passed through them. As the Go professionals did. "Don't delegate — absorb" means, precisely, this: don't absorb the answers. Absorb the way of arriving at them.

This essay, too, was written that way — out of the records of conversations with an AI, through dialectical engagement between me and the machine, typed with the AI's borrowed hands.

References

Kosmyna, N. et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv:2506.08872.
Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6.
Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips. Science, 333, 776–778.
Javadi, A.-H. et al. (2017). Hippocampal and prefrontal processing of network topology to simulate the future. Nature Communications, 8, 14652.
Dahmani, L., & Bohbot, V. D. (2020). Habitual use of GPS negatively impacts spatial memory during self-guided navigation. Scientific Reports, 10, 6310.
Shin, M., Kim, J., & Kim, M. (2021). Human Learning from Artificial Intelligence: Evidence from Human Go Players' Decisions after AlphaGo. Proceedings of the Annual Meeting of the Cognitive Science Society, 43, 1795–1801.
Shin, M., Kim, J., van Opheusden, B., & Griffiths, T. L. (2023). Superhuman artificial intelligence can improve human decision-making by increasing novelty. PNAS, 120(12), e2214840120.
Suri, G., Slater, L. R., Ziaee, A., & Nguyen, M. (2024). Do Large Language Models Show Decision Heuristics Similar to Humans? A Case Study Using GPT-3.5. Journal of Experimental Psychology: General. (arXiv:2305.04400).
Cheung, V. et al. (2025). Large language models show amplified cognitive biases in moral decision-making. PNAS.
Turpin, M., Michael, J., Perez, E., & Bowman, S. R. (2023). Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting. NeurIPS 36.
Anthropic. (2025). Reasoning Models Don't Always Say What They Think.
Clark, A., & Chalmers, D. (1998). The Extended Mind. Analysis, 58(1), 7–19.
Plato. Phaedrus (the myth of Theuth and King Thamus).
Graham, P. (2024). Writes and Write-Nots. paulgraham.com/writes.html
Osmani, A. (2025). Avoiding Skill Atrophy in the Age of AI. addyo.substack.com
Carr, N. (2013). All Can Be Lost: The Risk of Putting Our Knowledge in the Hands of Machines. The Atlantic.
Carr, N. (2014). The Glass Cage: Automation and Us. W. W. Norton.
99% Invisible. (2015). Children of the Magenta (Automation Paradox, pt. 1). ep. 170.
ABC News. (2016, March 15). Go Grandmaster Lee Sedol Reflects on Losing Series to Google's Computer. (post-match press conference).