Variable reward

The variable-ratio schedule is how slot machines, gacha pulls and loot boxes deliver prizes, and textbooks describe it as producing fast, steady responding with no pausing. Schlinger, Derenne and Baron's review of fifty years of ratio-schedule research traces that belief to Ferster and Skinner's cumulative records and calls it a procedural artifact: their variable-ratio distributions usually incorporated 1 as the lowest ratio, so a reinforcer could arrive on the very next response. When later work raised the lowest ratio to 4, 7 or 10, pausing reappeared and lengthened until it resembled fixed-ratio pausing in duration. The review's conclusion is that pausing under variable ratios can and does occur, depending on the distribution of the individual ratios. A sparse unpredictable payout does not inherit the behaviour of a dense one, so the schedule is not a citation that justifies thin rewards.

The human mechanism is better described as resolution as reward. Shen, Hsee and Talloen argue that the act of finding out is itself rewarding, which is why an uncertain incentive can out-pull a larger certain one. In Shen, Fishbach and Hsee's four studies with real rewards, participants invested more effort, time and money to qualify for an uncertain reward worth less in expectation — a coin flip between $1 and $2 against a guaranteed $2 — but only when attention was on the process of pursuit; under an outcome focus it did not hold. Four further experiments by Shen, Hsee and Talloen, in lab and field and at small and large magnitudes, added two boundary conditions: the effect arises only if the uncertainty is resolved immediately and not if resolution is delayed, and only after, not before, a person has engaged in repetitions. Their field study ran a 15-day event at a Hong Kong running club on a 400-metre track, where one arm earned five points a lap and the other three or five points a lap at random; the uncertain arm ran more laps despite the worse payout. Nothing in that account says uncertainty makes people want the thing more. It says the reveal is a small, separate reward laid on top of the thing. That is eight experiments from two overlapping author teams, called counterintuitive in the 2019 paper's own abstract, with no multi-lab and no registered-report replication of either — so the uncertain arm is the one that has to earn its place against a stated reward.

Uncertainty also has its own neural signal. Fiorillo, Tobler and Schultz recorded dopamine neurons under cues signalling different reward probabilities and found two distinct responses: a phasic activation that varied monotonically across the full range of probabilities, consistent with coding the discrepancy between predicted and actual reward, and a separate slower response that covaried with uncertainty — a gradual increase in activity until the potential time of the reward, maximal at p = 0.5, where uncertainty is greatest. The authors link the second to attention-based learning and risk-taking behaviour. That is a mechanism for anticipation, not for liking: what an uncertain reveal adds is the wait before it, which pays only if the wait is seconds long and ends on the same screen.

Rarity, not randomness alone, carries the response. Larche and colleagues had Overwatch players open loot boxes under physiological recording, in two experiments with 47 and 40 analysed participants. Rarer tiers were rated more valuable and more arousing, and the urge to open another box rose monotonically with tier: 25.39, 32.77 and 44.24 on a 0–100 scale for rare, epic and legendary in the first experiment, and 32.84, 40.00 and 55.63 in the second. In the second experiment pauses before the next box lengthened with tier — 4.91, 5.40 and 5.70 seconds — and skin conductance responses were largest for legendary (0.77), though rare (0.58) and epic (0.54) did not differ from each other. Arousal rose before the reveal rather than after it: skin conductance level slopes during the anticipatory period were steeper than at baseline, t(39) = −3.88, p ≤ .001, matching the uncertainty signal Fiorillo and colleagues recorded. A tiered reveal is therefore the strongest version of this pattern and the one closest to the harm literature: the pull to try again scales with how badly the top tier is wanted and how rarely it appears.

The one large-scale product dataset points at reward rate rather than unpredictability. Lindström and colleagues fitted reinforcement-learning models to over one million posts from over 4,000 individuals across several social media platforms, plus an online experiment (n = 176) on an Instagram-like platform. Users spaced their posts to maximise the average rate of accrued social rewards, subject to both the effort cost of posting and the opportunity cost of inaction, and the experiment showed social rewards causally influencing behaviour. The paper's own framing is that casting the online world as a Skinner Box for the modern human is a popular characterisation that evidence had not yet supported: what it supplies is evidence for reward-rate maximisation, not for variance in likes. The thing to build is the supply of real content.

The strongest body of work on this principle is about monetised randomness, and all of it is correlational. Zendle and Cairns surveyed 7,422 adult gamers and found loot box spending rose with problem gambling severity, χ²(3) = 284.255, p < 0.001, η² = 0.054, which they describe as small-to-medium; spending on other in-game microtransactions produced η² = 0.004. Garea, Drummond, Sauer, Hall and Williams meta-analysed 15 studies covering 16,229 participants for a pooled r = 0.26, 95% CI [0.15, 0.38], rising to r = 0.37 [0.25, 0.48] under trim and fill, with very high heterogeneity (I² = 98.25) and a Rosenthal fail-safe N of 8,360; the link to excessive gaming was r = 0.25 [0.15, 0.36] across 7 studies and 3,436 participants. Close and Lloyd pooled six open datasets of 7,771 loot box purchasers and obtained the same r = 0.26, p < 0.0001, noting the association is similar to or stronger than those between problem gambling and depression, drug use or current alcohol dependence. None of the three sets of authors can determine the causal direction; Garea and colleagues judge the explanation that people with existing symptoms buy disproportionately to be the more plausible one, since it does not require loot boxes to produce the pathology. The one prospective test is Palmer, Brooks and Clark's six-month study of 19- to 25-year-olds in British Columbia, 83 non-gamblers and 43 gamblers at baseline after cleaning: baseline loot box spending and Risky Loot Box Index scores predicted gambling initiation among the non-gamblers and predicted later gambling expenditure, while direct-purchase microtransaction spending did not predict initiation once loot box spending was controlled. The randomisation, not the spending, tracked the risk.

The revenue is concentrated in the people most at risk, and the pattern reaches children. In Close and Lloyd's pooled data, roughly 5% of loot box purchasers — those spending more than about £70 or $100 a month — generate around half of all loot box revenue and the top 2% around a third; almost a third of that top 5% fall in the problem-gambler band of the Problem Gambling Severity Index, and nothing in the dataset links higher loot box spend to higher personal earnings. Any feature that depends on repeat paid draws is operating inside that distribution. Zendle, Meyer, Cairns, Waters and Ballou found loot boxes in 58.0% of top-grossing Google Play games, 59.0% of top-grossing iPhone games and 36.0% of the most-played Steam games, with 93.1% and 94.9% of the loot-box mobile games rated suitable for ages 12 and over against 38.8% on desktop. The grade here follows the weaker half of the record: the harm association is sizeable and independently reproduced but cannot speak to cause, and the motivational effect a product team would actually build on rests on eight experiments from two overlapping author teams with no independent replication. Mixed, and the burden of proof sits with the uncertain version.