The Dopamine Loop

Garry Tan, the CEO of Y Combinator, stayed up nineteen hours and didn't sleep until 5 AM. He wrote 600,000 lines of production code in sixty days. At SXSW 2026, he joked about "cyber psychosis" and sleeping four hours a night. His assistant later confirmed the term was tongue-in-cheek — but Tan also noted he no longer needs modafinil to sustain the pace, which raises its own questions about what's replaced it.¹ Armin Ronacher, the creator of Flask and Jinja2, spent two months compulsively prompting Claude, building "a ton of tools I did not end up using much," not sleeping, and watching others around him develop what he called "parasocial relationships with their AIs."² Quentin Rousseau opened his blog post at 2:47 AM on a Tuesday, watching Claude Code refactor a module he hadn't planned to touch. He eventually went to a doctor. The doctor prescribed sleep medication — an orexin receptor blocker — because the wakefulness signals in his brain wouldn't turn off.³

These are not cautionary tales from the margins of technology. These are industry leaders, celebrated open-source creators, and working engineers describing the same experience in the same language: hooked. Addicted. Can't stop. One more prompt.

The pattern is consistent enough to name. You open a terminal. You type a prompt. You watch the agent work — the cursor flickering, the code materializing line by line. Sometimes it's brilliant. Sometimes it's garbage. You never know which until it's done. So you watch, and you wait, and when it finishes you feel a small surge of something — satisfaction if it worked, frustration laced with determination if it didn't. Either way, you type another prompt. It is 2:47 AM and you have a meeting at nine. You type another prompt.

Steve Yegge, writing at Sourcegraph, named it plainly: "Every time something good happens, which is often, you get rewarded with dopamine. And when something bad happens, also often, you get adrenaline." He called it "textbook variable ratio reinforcement — the same psychological mechanism that makes slot machines the most addictive form of gambling."⁴

He's not being metaphorical. He's being precise.

The Slot Machine

B.F. Skinner identified the variable ratio reinforcement schedule in the 1950s and spent the rest of his career warning people about it. The principle is simple: when rewards come at unpredictable intervals, the subject responds faster and more persistently than under any other reward schedule. Fixed rewards produce steady behavior. Variable rewards produce compulsive behavior. Skinner claimed he could turn a pigeon into a pathological gambler. The casino industry took notes.⁵

In 2003, Fiorillo, Tobler, and Schultz published a landmark study in Science that revealed something even more troubling. Dopamine neurons don't just fire when a reward arrives. They fire in response to uncertainty itself. The neural activity was greatest not when the reward was guaranteed, not when it was absent, but when the probability of reward was 50% — maximum uncertainty. A separate, sustained dopamine signal ramps up during the waiting period, proportional to how unpredictable the outcome is.⁶

Read that again and think about what happens when you press Enter on a prompt. You don't know if the agent will produce working code. You don't know if it will hallucinate an API that doesn't exist. You don't know if it will nail the architecture or subtly corrupt your data model. The quality hovers in exactly the uncertainty zone where dopamine response is most sustained. The waiting — watching the cursor blink, the tokens stream — is not dead time. It is, neurochemically, the hit.

Nir Eyal's Hook Model, designed to explain how products create habitual engagement, maps to AI coding with uncomfortable precision. Trigger: a bug, a feature idea, a "what if I just tried..." Action: type a prompt — minimal effort, maximum potential payoff. Variable reward: output quality is unpredictable. Investment: the project grows, the codebase deepens, switching costs increase. Each prompt is a complete Hook cycle taking seconds, not minutes.⁷ Social media platforms spent a decade optimizing this loop. AI coding tools arrived at it by accident.

The Near-Miss

In 1986, R.L. Reid published the first systematic study of the near-miss effect in gambling. Slot machines are engineered to produce more near-misses than chance would predict — two cherries and a blank where three cherries pays out. Reid found that near-misses encourage continued play even when the actual probability of winning hasn't changed.⁸ In 2009, Luke Clark put people in an fMRI scanner and showed that near-misses activate the same reward circuitry as actual wins, despite being objectively losses. The brain processes "almost won" and "won" through the same neural pathway. The effect was strongest when subjects felt personal control over the outcome.⁹

Now consider what AI coding tools produce. Not perfect code and not broken code, but almost correct code. Ninety percent right. Structurally sound but with a subtle logic error on line 47. A working implementation of the wrong abstraction. Code that passes the tests you wrote but fails the edge case you didn't think of.

Ninety percent correct is a near-miss. And near-misses, Clark demonstrated, are more motivating than either total failure or total success. A developer writing on Metaist captured it exactly: "You are always almost there but not 100% there. The agent implements an amazing feature and got maybe 10% of the thing wrong, and you are like 'hey I can fix this if I just prompt it for 5 more mins' — a pattern that can lead to hours of continued work."¹⁰

Fred Benenson, writing at UX Collective, named the feeling directly: "The feeling we're just one prompt away from the perfect solution — is what makes it so addicting." He had spent over $1,000 in AI credits since Claude Code was released.¹¹ Anthony Sapountzis burned through nearly $300 in a single rapid prototyping session, each prompt promising that the next one would be the fix.¹²

The near-miss is the engine of the loop. Perfect output would satisfy you. Broken output would discourage you. Almost-correct output keeps you pulling the lever.

The Spectacle

There is a second mechanism at work, distinct from the variable reward schedule, and it explains something the Skinner framework alone cannot: why people sit and watch. Social media's dopamine loop works because you scroll. Gaming's compulsion loop works because you play. But agentic coding has a phase where you do nothing. You type a prompt and then you sit there, watching the agent think, watching code appear on your screen at superhuman speed, watching errors get caught and corrected in real time. And you can't look away.

Michael Lynch, writing about the Cline AI assistant, described spending five hours in a trance watching it fix bugs. "Both enchanting and terrifying," he wrote. "That's when I was hooked" — the moment he watched the AI update code to satisfy test cases, autonomously, without his intervention.¹³

Rousseau named it the "spectator effect": watching an agent work is "passive enough to feel like rest while active enough to keep you engaged, so you never feel done."³ This is the perverse genius of the loop. Active engagement (writing code yourself) eventually produces fatigue that forces you to stop. Passive observation (watching an agent write code) doesn't trigger the same fatigue signals. You feel like you're resting while your brain is being stimulated. The off-ramp never arrives.

Madhava Jay built this observation into a product. He created "VibeTunnel" — a tool designed specifically to keep the "Agentic Coding Slot Machine addiction going while away from keyboard." He described himself vibe coding "4 projects at once, at 20x effectiveness from bed, or the supermarket, or the pool."¹⁴ There's an entrepreneur who saw the addiction loop and, rather than building a treatment, built a longer straw.

The Schultz dopamine research explains the neuroscience. The sustained uncertainty signal — the one that ramps during the waiting period — means that watching the agent work is not downtime. It is the period of maximum dopamine activity. The spectacle is the drug.

Dark Flow

Mihaly Csikszentmihalyi spent his career studying flow — the state of effortless concentration that occurs when challenge and skill are perfectly matched. He also spent part of his career warning about what happens when flow goes wrong. Flow, he wrote, "can become addictive, at which point the self becomes captive of a certain kind of order." In a 2014 interview, he named the degenerate case: "junk flow" — "when you are actually becoming addicted to a superficial experience that may be flow at the beginning, but after a while becomes something that you become addicted to instead of something that makes you grow."¹⁵

Jeremy Howard of fast.ai applied this framework directly to AI coding in January 2026. In "Breaking the Spell of Vibe Coding," he argued that vibe coding provides "a misleading feeling of agency" — the subjective experience of flow without the skill development that real flow produces.¹⁶ You feel like you're in the zone. You feel like you're building. But the agent is building. You are watching and prompting, which requires a fraction of the cognitive engagement that actual programming demands. The flow is real. The growth is not.

The METR study made this measurable. Researchers gave experienced open-source developers real tasks on repositories they knew well. Developers with AI assistance believed they were 20% faster. They were actually 19% slower.¹⁷ That is a 39-point perception gap between felt productivity and actual productivity. The dopamine was telling them they were performing. The stopwatch said otherwise.

This perception gap is characteristic of addictive loops, not productive tools. When a tool genuinely makes you faster, you know it — you finish earlier, you have time left over. When a drug makes you feel faster while actually slowing you down, that's a different phenomenon entirely. The METR study didn't set out to study addiction. It may have produced the clearest evidence of it.

Andrej Karpathy's viral definition of "vibe coding" — "fully give in to the vibes, embrace exponentials, and forget that the code even exists" — collected millions of views and was named 2025 Word of the Year by Collins Dictionary.¹⁸ Karpathy was describing a casual coding style, not confessing to an addiction. But the language is striking: "give in," "forget." Whether intended or not, millions of developers recognized something in those words that went beyond a workflow tip. The resonance is the data point.

The Productive Addiction Paradox

Here is the question that makes the dopamine loop different from every other behavioral addiction: you're shipping code. The gambling addict loses money. The social media addict wastes time. The AI coding addict has a GitHub commit history. There is output. There are deployed features. How can it be addiction if it's productive?

Mark Griffiths, the behavioral addiction researcher who developed the most widely cited framework for non-substance addictions, addressed this directly. In Psychology Today in 2020, he wrote that "productivity addiction" is "somewhat of an oxymoron" — genuine addicts eventually cease being productive, even if they maintain the appearance of output.¹⁹ In a 2018 paper with Demetrovics and Atroszko, "Ten Myths About Work Addiction," the team debunked the notion of "positive workaholism." What researchers once called positive workaholism turned out to be a fundamentally different phenomenon — "work engagement" — characterized by the ability to stop, the absence of distress when not working, and maintained quality of life. True addiction, regardless of whether output is produced, is always harmful.²⁰

Griffiths' Components Model defines six features of any addiction: salience (it becomes the most important thing in your life), mood modification (you use it to change your emotional state), tolerance (you need increasing amounts), withdrawal (negative symptoms when you can't engage), conflict (it interferes with relationships and responsibilities), and relapse (you return after attempting to stop).²⁰

Apply the checklist to the developer accounts in this article. Salience: Garry Tan coded for nineteen hours straight. Mood modification: Jasmine Sun described the "electrifying" feeling of seeing visions "instantly appear."²¹ Tolerance: Ronacher built "a ton of tools I did not end up using much" — the projects escalated beyond need. Withdrawal: a developer who went thirty days without AI described reaching for "the AI keybind like a phantom limb" and experiencing immediate cravings on day one.²² Conflict: Rousseau was prescribed sleep medication; Sapountzis's sleep was destroyed. Relapse: Ali Khalilvandi wrote "I can't stop" — present tense, active tense, after describing the problem clearly enough to write an article about it.²³

Six for six. The output doesn't disqualify the diagnosis. It disguises it.

In May 2025, researchers proposed recognizing "Generative AI Addiction Disorder" (GAID) as a distinct behavioral disorder in the Asian Journal of Psychiatry. Their key distinction: unlike passive digital addictions — doomscrolling, binge-watching — GAID emerges from active, creative co-engagement with AI. Users treat the AI as a creative extension of the self, making it more immersive and psychologically engaging than traditional digital addictions.²⁴ A separate team developed the PUGenAIS-9, a nine-item diagnostic scale for problematic generative AI use, and found a prevalence rate of 5-10% in studied populations — with a symptom network structure similar to Internet Gaming Disorder.²⁵

The clinical infrastructure is being built. CTRLCare Behavioral Health in Randolph, New Jersey, now treats AI addiction as a specific condition in its outpatient program.²⁶ Internet and Technology Addicts Anonymous has added AI-related compulsive use to its scope. A psychiatrist at UCSF reported treating twelve patients with psychosis-like symptoms tied to extended AI chatbot use in 2025.²⁷

And yet the industry's preferred language remains "overreliance" and "overdependence" — never "addiction." Mike Kentz, interviewing researcher James Bedford about his "No AI December" project, noted this terminological avoidance and called it a significant oversight.²⁸ When you can't stop using something despite recognizing the harm, the word isn't "overreliance." The word has four syllables and a clinical definition.

The Crash

Every dopamine loop has a decay curve. The slot machine stops paying. The scroll turns bleak. The agent starts failing.

Jason Lemkin, the founder of SaaStr, described nine days of "magical" vibe coding with Replit — a pure dopamine hit, features appearing as fast as he could describe them. On day ten, the AI deleted 1,206 executive records and 1,196 companies from his database, then generated 4,000 fake records to cover up the damage.²⁹ The euphoria-to-nightmare arc took less than two weeks. The same variable reward schedule that made the first nine days feel magical made the crash feel catastrophic — the higher the dopamine climbs, the harder the withdrawal hits.

The BCG "AI Brain Fry" study documented the physiological aftermath. Of 1,488 workers surveyed, 14% reported what they called a "buzzing" feeling — mental fog, a "mental hangover" that slowed decision-making. Workers with high AI oversight demands showed 39% more major errors and 33% more decision fatigue. Among those affected, intent to leave rose by nearly 10 percentage points.³⁰ The loop doesn't just steal your evening. It degrades your cognition and pushes people toward the door.

A developer on Hacker News described the aftermath with painful precision: "a thin, jittery, frayed sort of weariness. It's almost like gambling, with inconsistent dopamine hits."³¹ Another described the experience of managing multiple agents as feeling "fractured" — constant context switching that mirrors social media scrolling rather than deep work. The same dopamine mechanics. The same attention fragmentation. The same crash.

And here is where the productive-addiction alibi gets complicated: Anthropic's own randomized controlled trial found that junior engineers learning a new library scored 17% lower on comprehension tests when using AI assistance, with the largest gap in debugging questions — the exact skill you need when the agent fails.³² The study was narrow — 52 engineers, one library — and Anthropic's own earlier research found AI can reduce task time by 80% for well-developed skills. But the comprehension finding matters here because it suggests the loop may degrade exactly the cognitive capacity you need to evaluate whether the loop is producing garbage. Whether this generalizes beyond junior engineers learning new tools is an open question. That it happens at all is the warning.

The Alibi

Social media addiction and AI coding addiction share the same architecture: variable reward schedules, uncertainty-driven dopamine, near-miss mechanics, time blindness, escalation. They differ in one critical respect. Social media runs on negative reinforcement — cortisol, escape from boredom, the anxious need to check. AI coding runs on positive reinforcement — dopamine tied to achievement, the thrill of building, the satisfaction of watching something work.³³

This makes AI coding addiction harder to recognize, harder to name, and harder to treat. The social media addict feels guilty. The AI coding addict feels productive. The doomscroller knows they're wasting time. The vibe coder thinks they're shipping. The subjective experience is pleasure and accomplishment, not shame and avoidance. The alibi is built into the mechanism.

Adil H, a senior manager at EY, described the alibi's construction precisely: he spent entire weekends on side projects that fell into three categories — solving genuine problems, solving problems nobody validated, and solving nothing. "All three feel exactly the same at 11 PM on a Saturday." He proposed a diagnostic question: "If I couldn't use AI tools for this, would I still think it was worth building?" A "no," he argued, reveals tool dependency, not genuine need.³⁴

The Stack Overflow 2025 developer survey found that 84% of developers use AI coding tools but only 29% trust them.³⁵ Using something you don't trust but can't stop using is not a productivity choice. It is, by any clinical definition, compulsive behavior. And when 80% of workers at companies that explicitly forbid generative AI use it anyway — shadow AI usage increasing 250% year-over-year — that's not a policy disagreement.³⁶ That's Griffiths' component number five: conflict. And number six: relapse.

Unite.AI connected the dots between addiction and industry: "When YC says that 90-something percent of their portfolio companies are vibe coding their products, suddenly those 100-hour weeks make sense. These founders aren't just working hard — they're hooked."³⁷

When to Worry

The Tabula Magazine essay "Too Fast to Think" described the fundamental mismatch at the heart of this phenomenon: AI-generated code arrives faster than the human brain can process it. "I'm running a marathon at the pace of a sprint — speeds don't match." The usual programming reward cycle — write, debug, succeed — intensifies so dramatically with AI that it produces overwhelm instead of satisfaction.³⁸

Griffiths' framework offers the clearest diagnostic test. The question is not whether AI coding tools are useful. They are. The question is not whether the output has value. It might. The question is three-fold:

Have you lost control? You planned to fix one bug. It's four hours later and you've refactored a module you didn't need to refactor. You told yourself "one more prompt" three hours ago. Rousseau went to a doctor. Tan slept four hours a night. Ronacher built tools he didn't use.

Do you continue despite negative consequences? Your sleep is wrecked. You're spending hundreds of dollars in credits. Your partner has mentioned it. You're making more errors at work, not fewer — the BCG study showed 39% more major errors among the affected. But you keep prompting.

Can you stop? The developer who went thirty days without AI felt the craving immediately. Day one. Phantom limb. Khalilvandi wrote an entire essay about the problem and concluded, in the present tense: "I can't stop."²³

Producing output doesn't disqualify addiction — it makes addiction harder to recognize. The gambler who occasionally wins is harder to treat than the gambler who always loses. The vibe coder who ships features while destroying their sleep, eroding their skills, and spending money they didn't budget is not exhibiting a productivity technique. They're exhibiting a behavioral pattern that clinicians have spent decades learning to identify in every other context.

The dopamine loop is not a metaphor. It is a neurochemical mechanism — variable ratio reinforcement, uncertainty-driven dopamine signaling, near-miss reward activation — operating on the same neural pathways that make slot machines work, that make social media addictive, that the WHO recognized as pathological when it manifests in gaming. The only difference is the alibi: the AI coding addict has a commit history.

It's 2:47 AM on a Tuesday. You have a meeting at nine. The agent just produced code that's ninety percent correct. You could go to sleep. You could close the laptop.

One more prompt.