Overjustification effect

Two mechanisms are offered for it. The attributional one is self-perception: asked why they drew, or solved, or posted, the person now has a sufficient external reason available, and infers that the reward was the cause. The reward overjustifies an action that needed no justification, and the original interest is no longer the explanation they reach for.

The second mechanism, cognitive evaluation theory, is the one that survives contact with the moderators, because it predicts both directions. Every reward carries two aspects at once. A controlling aspect shifts the perceived locus of causality outward — I am doing this because the contingency requires it — and that costs autonomy. An informational aspect conveys competence — I did this well. Which aspect is salient decides the sign of the effect. That is why praise, which is almost all information, enhances intrinsic motivation, while a promised token for engaging in a task, which is almost all control, reduces it; and why the same praise, delivered in controlling language, flips to undermining. It also explains the measurement asymmetry that runs through the whole literature: the undermining effect appears in free-choice behaviour — whether people pick the activity up again when nobody is watching or paying — and barely appears in what people say about how much they enjoyed it. If you only ask, you will not see it.

The anchor is Deci, Koestner and Ryan's meta-analysis of 128 studies. Across 101 studies with a free-choice behavioural measure, all rewards together gave d = -0.24 (CI = -0.29, -0.19); across 84 studies with a self-report measure, d = 0.04 (CI = -0.02, 0.09), not significant. The structure sits in the moderators. All 92 tangible-reward studies gave d = -0.34 (CI = -0.39, -0.28) on free choice. Expected tangible rewards gave d = -0.36 (CI = -0.42, -0.30); the 9 studies of unexpected tangible rewards gave d = 0.01 (CI = -0.20, 0.22), and the 7 studies of task-noncontingent rewards gave a non-significant d = -0.14 (CI = -0.39, 0.11). By contingency: engagement-contingent d = -0.40 (CI = -0.48, -0.32, k = 55), completion-contingent d = -0.36 (CI = -0.50, -0.22), performance-contingent d = -0.28 (CI = -0.38, -0.18, k = 32). Verbal rewards ran the other way: positive feedback gave d = 0.33 (CI = 0.18, 0.43) on free choice across 21 studies and d = 0.31 (CI = 0.19, 0.44) on self-reported interest after three outliers were removed.

This was a contested literature, and the dispute is the most useful part of the record. Cameron and Pierce's 1994 meta-analysis reached the opposite conclusion and called for abandoning cognitive evaluation theory. As Deci, Koestner and Ryan report their figures, Cameron and Pierce found no overall reward effect on free-choice behaviour, enhancement on self-reported attitudes (d = 0.14), verbal-reward enhancement on both measures (d = 0.38 and 0.39), tangible-reward undermining on free choice (d = -0.21) and expected-tangible undermining on free choice (d = -0.25) — in other words, the same signs on the measures where the later analysis found them, read as a null because the aggregate was reported first. Deci and colleagues level four charges: collapsing across cells with significant interactions on theoretically meaningful variables, including initial task interest and informational versus controlling administration, without running those as moderators; using inappropriate control groups, including one study where every group received money and there was no no-reward control; misclassifying studies, including one whose method section states that rewards were given for working with a puzzle for at least eight minutes but which was coded as task-noncontingent; and omitting nearly 20% of studies as outliers rather than isolating the cause of the variance.

Cameron, Banko and Pierce then re-analysed the literature explicitly to resolve the difference. Their abstract still says that in general rewards are not harmful to motivation to perform a task — and then states the conditions: rewards given for low-interest tasks enhance free-choice intrinsic motivation; on high-interest tasks, verbal rewards produce positive effects on free-choice motivation and self-reported task interest; negative effects are found on high-interest tasks when the rewards are tangible, expected (offered beforehand), and loosely tied to level of performance. The two camps disagree about how to summarise that sentence. They do not disagree about the sentence. Design the interface from the conditions and the dispute is irrelevant to you.

The replication record is unusually clean for a 1970s effect. Lehtivuori's 2023 doctoral meta-analysis ran a fresh systematic search of seven databases, 158 qualifying articles, and synthesised 142 randomised controlled laboratory experiments from 125 published articles. The free-choice numbers land on top of the 1999 ones: all rewards d = -0.28, tangible d = -0.39, expected d = -0.41, engagement-contingent d = -0.42, completion-contingent d = -0.48, performance-contingent d = -0.24, unexpected d = -0.04 and not significant, rewards not tied to doing a specific task d = 0.10 and not significant, positive feedback d = 0.33. The self-report analysis again showed weaker effects, with engagement-contingent rewards the only undermining contingency (d = -0.16). A third meta-analysis in the same thesis, of 42 independent samples from 35 articles in work settings, found no significant overall association between extrinsic rewards and intrinsic work motivation; base salary and positive feedback were positively associated (r = 0.19 each), performance-based rewards were null and extremely heterogeneous (r = 0.05), and performance-based rewards perceived as more controlling than informational were weakly negative (r = -0.10).

Two studies take it out of the one-session lab. Greene, Sternberg and Lepper ran a token economy in two classrooms: 44 children, 11 per condition, 13 days of baseline, 13 of treatment and 13 of withdrawal, with time spent on target maths activities recorded unobtrusively. The reinforcement worked — every experimental group increased time on target during treatment, all within-group differences significant at p < .01. After the tokens were withdrawn, all three experimental groups spent less time on their target activities than they had at baseline, significantly so for the choice group, t(10) = 4.14, p < .01, and the high-interest group, t(10) = 2.38, p < .05. The low-interest group's decrement was not significant. Murayama, Matsumoto, Izuma and Matsumoto induced the effect in adults with a performance-based monetary reward and tracked it with fMRI: reward undermined intrinsic motivation as measured by the number of voluntary engagements in the task, and activity in the anterior striatum and prefrontal areas decreased along with the behavioural effect.

The counterweight matters as much as the effect. Cerasoli, Nicklin and Ford's 40-year meta-analysis (k = 183, N = 212,468) found intrinsic motivation to be a medium to strong predictor of performance (ρ = .21–.45) and that the importance of intrinsic motivation to performance remained in place whether or not incentives were presented. Incentive salience changed its predictive weight in a crowding-out pattern — intrinsic motivation mattered less where incentives were directly tied to performance — but their conclusion is that incentives and intrinsic motivation are not necessarily antagonistic and are best considered simultaneously. They also found the division of labour that should govern your metric choice: intrinsic motivation predicted more unique variance in quality of performance, incentives in quantity.

The one piece of directly relevant interface evidence is a null on motivation and a hit on volume. Mekler, Brühlmann, Tuch and Opwis ran a 2 × 4 online experiment (N = 273) on an image annotation task, comparing points, levels and a leaderboard against a control. Game elements significantly increased the number of tags generated, F(3, 265) = 10.09, p < .001, partial eta squared = .103, with points beating control (t(141) = 2.613, p = .01, d = .44) and leaderboard and levels in turn beating points (d = .39 and d = .35). Tag quality did not differ, F(3, 265) = .727, p = .537. Intrinsic motivation did not differ either (p = .702), nor did competence need satisfaction. Their reading is that the elements functioned as extrinsic incentives, effective only for promoting performance quantity. So the evidence for undermining in digital interfaces specifically is one experiment that found no effect on intrinsic motivation in either direction — the strong evidence here is for the lab effect and its conditions, not for a measured collapse of motivation inside a product.