5D Library › Attention Networks, the Working-Memory Bottleneck and Self-Determination Theory · 中文
Chapter 4 Crowding Out by Extrinsic Rewards: From Soma Puzzles to “A Fine Is a Price”
Rewards usually make people more willing to do things; that is one of the most reliable regularities in behavioural science. But from the 1970s a series of experiments found an apparent paradox: under some conditions, rewarding people for something they already liked doing made them less likely to do it of their own accord once the reward was gone. Psychologists call this the overjustification effect; economists speak of motivation crowding-out.
Two classic experiments
Evidence
Deci 1971. Deci had college students solve three-dimensional “Soma” puzzles over three sessions. In the second session the experimental group received one dollar for each puzzle solved and the control group nothing; in the third neither group was paid. In each session the experimenter left the room on a pretext for several minutes, during which participants could freely choose to carry on with the puzzles or do something else (read magazines and so on); the time spent on puzzles in this “free-choice period” served as the measure of intrinsic motivation. Compared with controls, the experimental group spent less free-choice time on the puzzles once payment stopped (Deci 1971). Lepper, Greene and Nisbett 1973. They selected nursery-school children who already liked drawing and assigned them to three groups: told in advance that drawing would earn a “Good Player Award” certificate, and given it (expected reward); not told, but given the certificate afterwards (unexpected reward); or no certificate. Observed one to two weeks later during free play in the classroom, the expected-reward group spent less time drawing spontaneously than the other two groups; the unexpected-reward and no-reward groups did not differ clearly (Lepper et al. 1973).
In both experiments the key lies in the measure: intrinsic motivation is not assessed by asking “do you like it?” but by seeing whether people choose to do the activity when there is no external reason to. Lepper’s experiment also points to an important condition: only the expected reward had an undermining effect.
Meta-analysis and controversy
Evidence
Deci, Koestner and Ryan 1999. This meta-analysis pooled 128 experiments. On free-choice behaviour, the effect of all rewards combined was ; verbal rewards (positive feedback) increased intrinsic motivation, , with for college students and (not significant) for children; tangible rewards overall had , unexpected tangible rewards almost no effect () and expected tangible rewards . By what the reward was contingent on: task-noncontingent (not significant), engagement-contingent , completion-contingent , performance-contingent (Deci et al. 1999). Interesting and dull tasks. In the 13 studies that included both kinds of task (11 of them with a free-choice measure), tangible rewards had an effect on free-choice behaviour of for interesting tasks and (confidence interval including zero) for dull ones; the authors concluded that rewards do not undermine intrinsic motivation for dull tasks, because there is little intrinsic motivation there to undermine (Deci et al. 1999, 2001). Controversy. Cameron and Pierce’s 1994 meta-analysis concluded that rewards do not, overall, reduce intrinsic motivation, and that a slight negative effect on free-choice time appears only when rewards are expected, tangible and unrelated to the quality of performance (Cameron and Pierce 1994); Eisenberger and Cameron in 1996 went further, arguing that detrimental effects occur only under restricted conditions that are easily avoided (Eisenberger and Cameron 1996). Deci and colleagues’ 1999 meta-analysis was largely a reply, arguing that the earlier coding and classification had hidden the differences between kinds of reward.
Calculation
Translate these effect sizes into the more intuitive “common-language” effect size: pick one rewarded participant and one control at random; what is the probability that the rewarded one spends less free-choice time? Assuming normal distributions with equal variance, it is .
# Deci, Koestner & Ryan (1999; interesting vs dull tasks: the 11 free-choice studies of its
# supplemental analysis, also reported in their 2001 review): free-choice behaviour, Cohen's d.
# Common-language effect size: P(rewarded person persists less than a control) = Phi(-d / sqrt(2)),
# assuming normal distributions with equal variance.
from statistics import NormalDist
PHI = NormalDist().cdf
ROWS = [
("all rewards", -0.24),
("verbal rewards (positive feedback)", 0.33),
(" ... children", 0.11),
(" ... college students", 0.43),
("tangible rewards", -0.34),
(" unexpected tangible", 0.01),
(" expected tangible", -0.36),
(" task-noncontingent", -0.14),
(" engagement-contingent", -0.40),
(" completion-contingent", -0.36),
(" performance-contingent", -0.28),
("tangible, interesting tasks (11 studies)", -0.68),
("tangible, dull tasks (11 studies)", 0.18),
]
print("reward type | d | P(rewarded person persists less)")
for name, d in ROWS:
print(f"{name:41s} | {d:+5.2f} | {PHI(-d / 2 ** 0.5):6.0%}")
print("50 % = no difference. Several small-k rows (unexpected, task-noncontingent, children's")
print("verbal rewards, dull tasks) were not statistically significant in the source.")
# reward type | d | P(rewarded person persists less)
# all rewards | -0.24 | 57%
# verbal rewards (positive feedback) | +0.33 | 41%
# ... children | +0.11 | 47%
# ... college students | +0.43 | 38%
# tangible rewards | -0.34 | 59%
# unexpected tangible | +0.01 | 50%
# expected tangible | -0.36 | 60%
# task-noncontingent | -0.14 | 54%
# engagement-contingent | -0.40 | 61%
# completion-contingent | -0.36 | 60%
# performance-contingent | -0.28 | 58%
# tangible, interesting tasks (11 studies) | -0.68 | 68%
# tangible, dull tasks (11 studies) | +0.18 | 45%
# 50 % = no difference. Several small-k rows (unexpected, task-noncontingent, children's
# verbal rewards, dull tasks) were not statistically significant in the source.As probabilities, the effects look more measured: an engagement-contingent tangible reward () means about a 61 % chance that the rewarded person persists less than the control, not “rewards make everyone lose interest”; the effect for interesting tasks () is about 68 %; positive feedback brings the probability down to about 41 %. These are pooled laboratory estimates, with the reward usually given once and a free-choice period of a few minutes.
Work performance and field evidence
Evidence
Intrinsic motivation and performance. Cerasoli, Nicklin and Ford pooled forty years of research and found that intrinsic motivation is a moderate-to-strong predictor of performance, especially its quality, while extrinsic incentives better predict its quantity. When incentives are directly tied to performance (piece-rate bonuses, say) the link between intrinsic motivation and performance weakens; incentives only indirectly tied to performance (such as a fixed salary) do not weaken it. The authors concluded that the two jointly predict performance rather than trading off against each other (Cerasoli et al. 2014). A nuclear-waste repository. Frey and Oberholzer-Gee interviewed 305 residents of two adjacent central Swiss communities under consideration in 1993 as a site for a repository for low- and intermediate-level nuclear waste. Without compensation, 50.8 % said they would accept it; when monetary compensation was added to the question, acceptance fell to 24.6 % (Frey and Oberholzer-Gee 1997). A fine is a price. Gneezy and Rustichini ran a field experiment in ten private day-care centres in Haifa, Israel: six began fining parents who arrived more than ten minutes late to collect their children, and four served as controls. After the fine was introduced, late arrivals rose over a few weeks and settled at about twice the original level; when the fine was removed, they did not fall back (Gneezy and Rustichini 2000).
What these studies share is that money changed not only the “price” of acting but people’s understanding of what they were doing. Compensation turned “accepting a burden for the common good” into a transaction; the fine turned “an apology owed to the teacher” into a service one could buy, and the old norm did not return when the price was withdrawn.
A reward as a signal: a Bayesian model
Why should an expected reward undermine interest when an unexpected one does not? Bénabou and Tirole, in the Review of Economic Studies, offered an explanation from the economics of information: the principal (parent, teacher, manager) often knows more than the agent about the nature of the task; if the principal offers a large reward for doing something, the agent can reasonably infer that “this probably isn’t worth doing for its own sake”, or that “the principal doesn’t think I’ll do it well” (Bénabou and Tirole 2003).
Calculation
The program below is a heavily simplified version of this idea (not the paper’s full model). The task’s intrinsic value to the agent is either high (1.0) or low (0.2); the agent does not know which, prior half and half; the principal knows, and tends to offer more for a low-value task. The agent infers the type from the reward it sees: today it acts if “inferred value + reward” exceeds the cost; tomorrow the reward is withdrawn and it carries on if “inferred value” exceeds the cost.
# A reward as a signal, a stripped-down reading of Benabou & Tirole (2003); values illustrative.
# Task value HIGH (1.0) or LOW (0.2), prior 0.5; the principal knows the type and sets the reward,
# reward ~ Normal(mu_type, 0.15). The agent acts today if E[v | reward] + reward > cost and,
# with the reward withdrawn, tomorrow if E[v | reward] > cost (cost 0.55).
import random
from statistics import NormalDist
random.seed(3)
V = {"HIGH": 1.0, "LOW": 0.2}; COST = 0.55; SD = 0.15
def posterior_high(r, mu):
lh = NormalDist(mu["HIGH"], SD).pdf(r) * 0.5
ll = NormalDist(mu["LOW"], SD).pdf(r) * 0.5
return lh / (lh + ll)
def world(gen, inf=None, announced=True, n=20000):
# gen: how rewards are actually set; inf: how the agent believes they are set (None = ignores)
today = tomorrow = highs = 0
for _ in range(n):
t = "HIGH" if random.random() < 0.5 else "LOW"
r = max(0.0, random.gauss(gen[t], SD)) if gen else 0.0
p = posterior_high(r, inf) if inf else 0.5
ev = p * V["HIGH"] + (1 - p) * V["LOW"]
today += ev + (r if announced else 0.0) > COST
highs += t == "HIGH"
tomorrow += t == "HIGH" and ev > COST
return today / n, tomorrow / highs
TRACK = {"HIGH": 0.1, "LOW": 0.5}; FLAT = {"HIGH": 0.3, "LOW": 0.3}
RICH = {"HIGH": 0.5, "LOW": 0.5}
print("tomorrow's free choice is counted for tasks that are in fact HIGH-value")
print("condition | engage today | HIGH tasks tomorrow")
for label, gen, inf, ann in (
("no reward (prior only: E[v] = 0.60)", None, None, False),
("reward tracks the principal's knowledge", TRACK, TRACK, True),
("same reward whatever the type, agent knows it", FLAT, FLAT, True),
("surprise reward to everyone, after the choice", FLAT, None, False),
("generous pay for every task, agent reads it as", RICH, TRACK, True),
(" a signal that the task needs inducement", None, None, None)):
if ann is None:
print(f"{label:50s} |"); continue
a, b = world(gen, inf, ann)
print(f"{label:50s} | {a:12.0%} | {b:19.0%}")
p = posterior_high(0.5, TRACK)
print(f"a reward of 0.5 moves P(task is worth doing for its own sake) from 50 % to {p:.0%}")
# tomorrow's free choice is counted for tasks that are in fact HIGH-value
# condition | engage today | HIGH tasks tomorrow
# no reward (prior only: E[v] = 0.60) | 100% | 100%
# reward tracks the principal's knowledge | 100% | 92%
# same reward whatever the type, agent knows it | 100% | 100%
# surprise reward to everyone, after the choice | 100% | 100%
# generous pay for every task, agent reads it as | 100% | 11%
# a signal that the task needs inducement |
# a reward of 0.5 moves P(task is worth doing for its own sake) from 50 % to 3%In this model, with no reward the agent relies on the prior (expected value 0.60, above the cost of 0.55): everyone acts, and carries on tomorrow. When the reward follows the principal’s knowledge of the task, the reward becomes a signal of the task’s value: high-value tasks usually get only a small reward, which confirms to the agent that they are worth doing; only the few high-value tasks that happen to get a larger reward are misjudged and abandoned tomorrow (continuation falls to 92 %). If every task gets the same reward, or if a surprise reward is given to everyone after the choice, the reward carries no information and crowding out disappears, in line with Deci and colleagues’ finding that unexpected rewards have no effect. The last line is the most telling: when the principal pays generously for every task and the agent reads payment as a sign that “inducement is needed”, only 11 % carry on with genuinely valuable tasks tomorrow; a reward of 0.5 moves the posterior probability that “this is worth doing for its own sake” from 50 % to 3 %.
Note
The model’s limits. This is only one mechanism, written in Bayesian terms, showing how a reward can be a signal. Bénabou and Tirole’s original model also includes ability, effort and the principal’s strategic choices; psychology’s cognitive evaluation theory explains the same phenomena in terms of the sense of being controlled and the sense of competence (see the next chapter). The two explanations are not mutually exclusive, but their predictions differ: the signalling account predicts that crowding out should weaken when people are sure the reward has nothing to do with the task’s value (for example, when everyone is paid the same).
Hypothesis · not established
Signal-strength hypothesis (not yet established). An expected reward should undermine free-choice behaviour less if participants are told that “this reward goes to everyone who takes part and has nothing to do with the task itself”. Testable prediction: in a three-arm randomised experiment (no reward; reward without explanation; reward with that explanation), free-choice time in the third arm should be significantly higher than in the second. If they do not differ, the signalling account is not supported in this setting.
Chapter summary
Expected tangible rewards contingent on engagement, completion or performance reduce the time people freely spend on an interesting activity once the reward is gone; unexpected rewards do not; positive feedback that affirms competence increases intrinsic motivation (clearly in college students); and for tasks that were dull to begin with, crowding out does not hold. The effect sizes, to , amount to roughly a sixty–forty difference, not “rewards destroy interest”. A Bayesian signalling model shows how crowding out can arise when a reward is read as information about the task’s value. The next chapter places these findings within self-determination theory.
References
Fan, Y. (2026). Attention Networks, the Working-Memory Bottleneck and Self-Determination Theory: From Posner's Tripartite Network and Motivation Crowding-Out to the Information Theory of Flow, Read Beside Neo-Confucian Reverence (Jìng) [注意力网络、工作记忆瓶颈与内在动机的自决定理论] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23761341 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).