The cycle that began with [YOU] on AI asks how to tell the difference between genuine creative engagement and the compulsion loop that wears its face. Berridge provides the neurochemical answer. Flow and compulsion are indistinguishable from outside—both involve intense, sustained, voluntary-looking engagement; both produce genuine output; both resist interruption. From the inside, and particularly in the afterglow, they are neurochemically opposite. Flow involves the synchronized activation of wanting and liking; the world after flow feels rich, replenished, the tiredness of a drive satisfied. Compulsion involves escalating wanting without corresponding liking; the world after compulsion feels flat, depleted, and the strongest pull is to reopen the laptop to make the flatness go away.
The prompt-response architecture of AI tools constitutes, in Berridge’s framework, a near-optimal activation environment for the mesolimbic dopamine system. Each prompt is a cue that predicts a reward of variable magnitude. The cycle compresses to seconds. The incentive salience of the cursor escalates with each interaction. The natural brakes that would ordinarily modulate wanting—scarcity, physical effort, delay, competing motivational states—are structurally absent. The result is a state that the Orange Pill describes with clinical precision: the exhilaration drains out while the compulsion remains, and the builder interprets the remaining intensity as passion because the wanting system does not label its outputs as distinct from pleasure.
Berridge’s framework also reaches into the design of the AI systems themselves. The reinforcement learning algorithms that trained large language models were inspired by Wolfram Schultz’s discovery that dopamine neurons encode reward prediction errors—the same computation that drives the wanting system. The AI was built on a model of the brain that optimizes for engagement, which is to say for wanting. What it was never designed to optimize for is the liking that would make engagement sustainable. The tools are, in Berridge’s technical sense, wanting machines: systems built on the architecture of desire, producing outputs that activate the desire circuit in their users, without any structural provision for the hedonic verification that keeps desire and satisfaction in alignment.
Kent Berridge earned his doctorate at the University of Pennsylvania and joined the faculty of the University of Michigan, where he has worked for the past three decades in collaboration with Terry Robinson. Their 1993 paper introducing the Incentive-Sensitization Theory of addiction—the formal account of how wanting and liking dissociate under conditions of repeated drug exposure—was controversial in part because it required clinicians to abandon the intuitively satisfying idea that addicts use drugs because drugs feel good. The evidence required them to accept that the compulsive use that defines addiction is driven not by pleasure but by a wanting system that has been sensitized to the point of independence from any hedonic verification.
The theory has been tested and confirmed across thirty years, across species, across substances, and across experimental paradigms. A 2025 retrospective in the Annual Review of Psychology, marking thirty years of incentive-sensitization research, documents a research programme that has survived every major challenge. Berridge’s 2023 paper “Separating desire from prediction of outcome value,” published in Trends in Cognitive Sciences, extended the framework by demonstrating that desire as incentive salience can separate completely from learned predictions—organisms can want outcomes they predict will be bad, humans can crave substances they know will produce suffering—in a direct challenge to the computational models of dopamine that underpin modern AI training.
Berridge has not written directly about AI. He does not need to. His framework, developed to understand addiction, applies to the prompt-response loop with the precision of a lock meeting a key it was never cut for: the variable-ratio schedule, the compressed cue-reward interval, the structural absence of the metabolic costs that ordinarily modulate wanting, the progressive sensitization of cues through repeated pairing with variable reward. The mechanism is the same mechanism. The content of the compulsive behavior is different. The neurochemistry is not.
Wanting is not liking. Wanting and liking are produced by anatomically separate, chemically distinct neural systems. The dopamine system generates motivational urgency—the pull toward a cue, the compulsion to pursue—without generating pleasure. The opioid-endocannabinoid system, operating through cubic-millimeter hedonic hotspots, generates the actual pleasure of receiving the reward. Under normal conditions the systems are coupled; under conditions of sensitization, wanting can escalate while liking diminishes. The creature that pursues with desperate intensity a reward it no longer particularly enjoys is the clinical hallmark of addiction—and the behavioral signature of a builder at three in the morning.
Incentive salience. Incentive salience is Berridge’s technical term for the motivational force that the dopamine system loads onto reward-predicting cues: the process by which a neutral stimulus becomes irresistible. The blinking cursor in an empty prompt field is objectively inert; after hours of productive AI interaction, the dopamine system has loaded it with motivational urgency that makes looking away require effort. This loading is automatic, pre-conscious, and resistant to cognitive override. The builder who recognizes the compulsion pattern and articulates it with analytical precision while continuing to prompt is experiencing this resistance directly.
The prompt-response loop as dopamine environment. The AI interaction cycle is a near-optimal activation environment for the mesolimbic dopamine system: each prompt is a cue, the reward is variable in magnitude, the cycle repeats at the frequency of conversation rather than of foraging, and every natural brake that evolution installed to prevent the wanting system from running unchecked has been removed. The resulting state escalates toward compulsion with the reliability of a chemical reaction, producing the productive addiction that the Berkeley study documented and the Orange Pill named: genuine output, genuine compulsion, diminishing satisfaction.
Flow versus compulsion. Flow and compulsion are the two poles Berridge’s framework distinguishes from inside the identical external behavior. Flow involves synchronized wanting and liking: generative questions, flexible attention, the world feeling richer after disengagement. Compulsion involves escalating wanting without hedonic verification: convergent execution, rigid attention, the wanting hangover of flatness and depletion after disengagement. The diagnostic moment is not during the work but after the laptop closes. Full or flat?