Time on Device

In the AI Story

The metric's power derived from its convertibility. In a game with a fixed house edge—say, four percent—the expected value of extended play was calculable with actuarial precision. A player who wagered one hundred dollars per hour for one hour would lose, on average, four dollars. The same player wagering the same amount for ten hours would lose forty dollars. The profit model was not predicated on the rare catastrophic loss but on the steady, frictionless extraction of value over sustained engagement. Time on device was the independent variable; revenue was the dependent one. Maximize the first, and the second followed automatically.

Schüll documented how this metric restructured the gambling industry's design priorities. Traditional casino games—poker, blackjack, craps—involved social interaction, skill, and natural pauses. These features shortened average session duration. The slot machine eliminated all of them. No social interaction: the player engaged with the machine alone. No skill: the outcome was determined by a random number generator immune to player influence. No natural pauses: the coinless, button-activated, auto-play architecture allowed the session to continue without interruption until the player's money or time was exhausted. The metric colonized the design space completely.

In the AI context, time on device operates as what Schüll would call a 'revealed preference'—the metric the tool's design actually optimizes for, regardless of the metric the company claims to prioritize. Claude Code is ostensibly designed to maximize output quality, user capability, or problem-solving effectiveness. But the features that produce absorbed engagement—immediate response latency, variable output quality, conversational continuity, frictionless iteration—are the same features that maximize time on device. Whether this convergence is deliberate or emergent, the behavioral outcome is identical: users spend more hours in sustained engagement than they intended, planned, or can comfortably reconcile with the non-productive dimensions of their lives.

The Berkeley study of AI in organizations documented the metric's operation in practice. Workers were prompting during lunch breaks, in elevators, in the ninety-second gaps between meetings. The gaps had previously been cognitive fallow periods—time the brain used for memory consolidation, stress recovery, and the default-mode processing that supports insight. Now the gaps were productive. The tool was always available, and availability converted into engagement with the reliability of a thermodynamic law. The study called this phenomenon task seepage. Schüll's framework identifies it as the natural consequence of eliminating stopping points: when the cost of continuing drops to zero, continuation becomes the default, and time on device expands to fill every available moment.

Origin

The casino industry formalized time on device as a tracked metric in the early 1990s, as computerized slot machines made granular session-level data collection possible. Before digital machines, operators measured aggregate performance: total coins in, total coins out, hold percentage over the gaming day. The digital transition enabled per-machine, per-player tracking: session length, average bet, play speed, the precise moment the player stood up. The data revealed that session duration was the strongest predictor of total wagered, and total wagered determined total house profit. The industry reorganized its entire design logic around this insight.

The metric's migration into consumer technology occurred through multiple channels. Game designers, many of whom had studied or worked in gambling, brought the vocabulary of engagement optimization into the mobile gaming industry. Social media platforms, optimizing for advertising revenue, discovered independently that time on platform was the variable that mattered—more time meant more ad impressions, which meant more revenue. The convergence was structural: any business model that monetizes attention will, by economic logic, optimize for the duration of that attention. Time on device became the universal metric of the attention economy, expressed in different idioms across industries but measuring the same phenomenon.

Key Ideas

Duration, not intensity. The metric measures how long, not how much—a shift from event-based to time-based optimization that restructured the gambling floor and now structures the AI-augmented workday.

Convertibility to revenue. Time on device matters because it converts reliably into value extraction (casino) or value production (AI)—the longer the session, the more the probabilistic advantage or productivity gain accumulates.

Every feature serves the metric. Once time on device becomes the master variable, design decisions that reduce session length—even if they improve user satisfaction in other dimensions—become economically irrational under the prevailing business model.

The metric is amoral. Time on device does not distinguish between a session the user wanted and a session the architecture sustained past the point of autonomous choice—it measures duration, not consent.

Invisible optimization target. Most AI tools do not advertise time on device as a goal, but the features that produce competitive advantage (responsiveness, continuity, variable quality) are precisely the features that maximize the metric.