你奖励什么

#1. Executive Summary

This chapter’s central thesis is simple to state and hard to live by: incentives are the rules that shape behavior, and people respond not to what you intend but to what you actually reward — which is why well-meant incentive designs so often produce perverse, gamed, or demotivating results. An incentive is not a lever that reliably pushes behavior in the direction you want; it is a signal that changes the meaning of an act, redistributes effort across measured and unmeasured tasks, and invites strategic counter-adaptation from everyone it touches.

The robust empirical core is this. Monetary incentives work well — often dramatically — on tasks that are simple, individually attributable, and well-measured (Edward Lazear’s 2000 Safelite study found a 44% productivity gain from switching to piece rates). But the same instrument reliably backfires under identifiable conditions: when it crowds out pre-existing intrinsic motivation (Deci 1971; Gneezy & Rustichini’s 2000 “A Fine Is a Price” daycare study), when it corrupts a social or moral signal (Titmuss on blood; Frey & Oberholzer-Gee 1997 on nuclear-waste siting), when it targets one measurable task among many and starves the unmeasured ones (Holmström & Milgrom 1991’s multitask model), and when the measure becomes a target and is gamed (Bevan & Hood 2006 on the NHS; Jacob & Levitt 2003 on teacher cheating). These are not random failures; they are predictable from the structure of the situation.

The practical payoff is a design discipline. Before deploying any incentive — on an employee, a citizen, a child, or yourself — you can ask a short sequence of questions (named behavior, task type, cheapest way to game it, crowding-out risk, second-order reactions, feedback loop) that catches most failures before they happen. This chapter turns the previous chapters’ knowledge of how people think and err into an engineering practice of how to design the environments that shape what they do.

#2. Why This Topic Matters

Every organization, market, and public policy is, at bottom, a bundle of incentives. When institutions fail, the instinct is to blame people — lazy workers, greedy bankers, uncaring bureaucrats. This instinct is usually wrong. Most institutional failures are incentive-design failures, not people-failures: put ordinary, well-meaning people inside a badly structured reward system and you will reliably get bad outcomes, and swapping in “better people” will not fix it because the next set of people faces the same structure. Charlie Munger’s dictum — “Show me the incentive and I will show you the outcome” — is not cynicism; it is the first thing a diagnostician should check.

For the individual, understanding incentives is the difference between fighting your own psychology and designing around it. Willpower is a weak and exhaustible force; a well-built incentive architecture — commitment devices, identity framing, feedback loops — does the work that willpower cannot sustain. The person who understands incentives stops asking “why can’t I make myself do this?” and starts asking “what environment would make this the path of least resistance?”

This is the chapter where the curriculum turns a corner. Chapters 1–7 were about understanding cognition: how we decide, how we err, how we reason under uncertainty, how second-order consequences propagate. This chapter is about engineering environments: taking that knowledge and using it to design the rules that govern behavior — our own and others’. The previous chapter (Second-order Thinking) established that “how will others react?” is the most common engine of second-order consequences. Incentives are that engine’s underlying logic, studied head-on. We handle the empirical consequences of incentives — the real behavioral responses and the unintended ones. The formal framework of strategic interaction — what happens when everyone is optimizing against everyone else’s optimization — is game theory, and it is the next chapter. Here we stay empirical: what actually happens when you change the rules.

#3. Foundations

Core concepts.

• Incentive: any feature of the environment that changes the expected costs or benefits of an action, and thereby the likelihood it is taken. Crucially, incentives operate not only through material payoff but through meaning — what the incentive signals about the act, the actor, and the relationship.
• Extrinsic vs. intrinsic motivation: extrinsic motivation comes from separable consequences (money, grades, fines, praise); intrinsic motivation comes from the activity itself (interest, mastery, meaning, identity). This distinction, formalized in Edward Deci and Richard Ryan’s Self-Determination Theory, is the conceptual spine of the whole chapter.
• Principal-agent problem: one party (the principal) needs another (the agent) to act on their behalf, but the agent has different interests and better information. All incentive design is downstream of this asymmetry.
• Moral hazard (hidden action): once shielded from consequences, agents take actions the principal cannot observe (the insured drive less carefully; the salaried employee shirks).
• Adverse selection (hidden type): the principal cannot observe the agent’s type before contracting (the sick disproportionately buy insurance; lemons drive out good cars).
• Multi-tasking: when a job has several dimensions and only some are measurable, paying for the measurable ones diverts effort away from the unmeasured ones (Holmström & Milgrom 1991).
• Crowding-out: an extrinsic reward can reduce total motivation by displacing intrinsic motivation, especially when it is experienced as controlling.
• Pay-for-performance: compensation tied to measured output; powerful for simple tasks, treacherous for complex ones.
• Goodhart’s Law (revisited from Ch. 6): “When a measure becomes a target, it ceases to be a good measure.” The moment a metric carries stakes, people optimize the metric rather than the thing it was meant to track.
• Skin in the game (Taleb): the alignment achieved when the decision-maker personally bears the downside of their decisions — the deepest and most gaming-resistant incentive.

Historical development. The idea that self-interest can be harnessed for social benefit is Adam Smith’s (1776) — the butcher and baker serve us not from benevolence but from regard to their own interest. But Smith also warned (in The Theory of Moral Sentiments) that humans are moved by sympathy and propriety, not money alone — a duality the field has been re-discovering ever since. Chester Barnard’s The Functions of the Executive (1938) argued that material incentives are weak organizational glue compared to purpose and persuasion. The principal-agent revolution of the 1970s (James Mirrlees, Stephen Ross, Bengt Holmström) gave the field its mathematical spine, formalizing how to write contracts under hidden action and information asymmetry. In parallel and in tension, Edward Deci’s (1971) Soma-cube experiments launched the psychological research program showing that money can undermine intrinsic motivation. The 1990s–2000s brought a field-experiment wave (Lazear on piece rates, Gneezy and colleagues on fines and crowding-out) that tested theory against real behavior. The 21st century has produced a backlash against metric-driven management — Bevan & Hood, Jerry Muller’s The Tyranny of Metrics, Michael Sandel’s moral critique — arguing that the measurement-and-reward machine corrodes the very activities it tries to improve.

#4. Current Scientific Understanding

The robust core (well-established).

1. Intrinsic motivation is real and measurable. Deci’s (1971) free-choice paradigm — leaving subjects alone with a task and seeing whether they keep doing it unpaid and unobserved — operationalized “doing it for its own sake,” and hundreds of studies have used it since.
2. Monetary incentives improve performance on simple, well-measured tasks. Lazear’s (2000) Safelite windshield study is the canonical demonstration: a 44% output gain. Piece-rate evidence across contexts (tree-planting, fruit-picking, factory work) corroborates.
3. Incentives can crowd out intrinsic motivation under specific, predictable conditions. The 1999 meta-analysis by Deci, Koestner & Ryan (128 experiments, Psychological Bulletin) concluded that tangible, expected, performance-contingent rewards significantly undermine free-choice intrinsic motivation — while verbal rewards (praise, competence feedback) enhance it. This is not folklore; it is one of the more replicated findings in motivation science, though its magnitude is contested (see below).
4. Multi-tasking explains why pay-for-performance fails on complex work. Holmström & Milgrom (1991) proved formally that when tasks vary in measurability, strong incentives on the measurable task distort effort away from the unmeasurable — sometimes so severely that the optimal contract pays a flat wage with no performance component.
5. Gaming is a structural, predictable response to measurement, not a moral aberration. Wherever a metric carries stakes, a fraction of agents will optimize the metric directly. This is Goodhart’s Law with data behind it (Bevan & Hood 2006; Jacob & Levitt 2003).

The live debates (plausible theory / unsettled).

• How general is crowding-out? Judy Cameron and W. David Pierce’s meta-analyses (1994, 2001) argued the undermining effect is small, fragile, and largely confined to specific reward procedures; Deci, Koestner & Ryan (1999, 2001) countered that Cameron & Pierce’s analysis was methodologically flawed and that the effect is real and substantial. Lepper, Henderlong & Gingras (1999) sided largely with Deci et al. The honest summary: crowding-out is real but conditional, and reasonable scientists still dispute how wide those conditions are.
• Do financial incentives corrupt or merely activate markets? When money enters a domain, does it destroy a social signal (the economist Bruno Frey’s “motivation crowding”) or simply switch on ordinary market behavior? The answer appears to be “it depends on framing and prior meaning” — which is a real finding but frustratingly hard to predict ex ante.
• Is CEO pay a genuine incentive or a rent-extraction ritual? The “optimal contracting” view says executive pay aligns managers with shareholders; the “managerial power” view (Bebchuk & Fried) says executives capture their own boards and extract pay subject only to an “outrage constraint.” The empirical pay-performance sensitivity appears low relative to what naïve theory predicts: Michael Jensen and Kevin Murphy (1990, Journal of Political Economy 98:225–264, “Performance Pay and Top-Management Incentives”) found that CEO wealth changes by only $3.25 for every $1,000 change in shareholder wealth — “hardly enough,” they wrote, to get CEOs to reliably pursue value-increasing projects and avoid value-destroying ones. This “Jensen–Murphy puzzle” (CEOs then held a median of only ~0.25% of their firm’s stock) is precisely the datapoint the two camps read in opposite ways.
• What is optimal design when multiple tasks matter? Still open. The theory says “blend measured incentives with fixed pay and subjective judgment,” but calibrating that blend remains an art.

#5. Interdisciplinary Perspectives

The four disciplines that own this topic disagree productively.

Where they complement: Psychology explains the motivational channels that economics abstracts away — the economist’s “utility” is the psychologist’s autonomy-competence-relatedness. Sociology explains the meaning-layers that money can destroy — why a blood donation and a blood sale are different acts even at the same quantity. Behavioral economics is the bridge: it keeps the economist’s optimization framework but populates it with the psychologist’s real human.

Where they conflict: The sharpest clash is over meaning. To the economist, “a price is a price”: adding money to an act simply adds an incentive, and more incentive means more of the behavior. To the psychologist and sociologist, “a price changes the meaning of the act”: introducing money can transform a gift into a transaction, a civic duty into a purchase, a moral choice into a cost-benefit calculation — and the behavioral response can reverse sign. The daycare fine study is the empirical detonation of “a price is a price.” A second clash: the economist’s faith in optimization-by-metrics versus the sociologist’s warning that metric cultures (“targets and terror,” in Bevan & Hood’s phrase) hollow out the professions they measure.

#6. Mental Models

1. “You get what you pay for” — the fundamental theorem of incentive design. Not “you get what you want” — you get what you measure and reward. Corollary: be careful what you pay for, because you will get exactly that, including all the cheapest ways of producing it. When it works: as a diagnostic — trace any perverse outcome back to what was actually rewarded. Where it fails: it can breed paralysis if taken to mean “never incentivize anything”; the point is precision, not abstinence.

2. The intrinsic/extrinsic crowding-out map. Ask two questions: (a) Does the person already do this for its own sake? (b) Will the reward be experienced as controlling or as informational (signaling competence)? If there is little pre-existing intrinsic motivation (installing windshields), monetary incentives are relatively safe and often powerful. If there is strong intrinsic or moral motivation (donating blood, teaching, hobby art), controlling monetary rewards risk crowding it out; prefer recognition, competence feedback, or restructured rewards. Where it fails: the map tells you the direction of risk, not the exact magnitude, which is context-dependent.

3. The multitask trap (visible vs. invisible tasks). Whenever a role has a measurable dimension and an unmeasurable one, strong incentives on the measurable dimension will cannibalize the unmeasurable. Teachers paid for test scores under-teach curiosity; surgeons ranked on mortality avoid the sickest patients; call-center staff timed on call length hang up on hard problems. Design response: either measure more dimensions, or deliberately keep incentives muted and rely on professional norms and subjective evaluation.

4. The fine-as-price reframe (sanctions as permission). A penalty attached to a behavior can convert a moral prohibition into a purchasable option. “Don’t be late” becomes “lateness costs $3, and I’ll decide if it’s worth it.” Before imposing a fine, ask: am I forbidding this, or pricing it? If you set the price too low, you have legalized the behavior.

5. Skin in the game (asymmetry of consequences). The most robust incentive is one where the decision-maker eats their own cooking — bears the downside, not just the upside. Bankers who keep bonuses from loans that later default have upside without downside; the fix is clawbacks, deferred compensation, co-investment. Where it fails: some valuable roles (research, exploration) require shielding people from downside so they take socially useful risks; skin in the game can induce excessive caution.

6. Incentive-compatibility thinking (“would this survive being gamed?”). Before deploying a rule, become its adversary: what is the cheapest action that maximizes my reward without producing the outcome you actually want? If a cheap gamed response exists, redesign before launch. This is Goodhart pre-mortem.

7. The identity incentive (who you are as the deepest incentive of all). The most durable motivator is not payoff but self-concept: “I am a runner,” “I am the kind of person who keeps their word.” Identity-based motivation is self-sustaining (it needs no external payer) and gaming-resistant (you cannot cheat your way to being someone). Its limit: identities are slow to build and can become rigid.

#7. Common Misconceptions

• “People are lazy and only respond to money.” False and self-fulfilling. Humans are massively motivated by status, meaning, curiosity, reciprocity, and identity; design as if only money matters and you will build systems that crowd out everything else, producing the money-only behavior you assumed. Intelligent people fall for this because money is the easiest motive to measure and model.
• “More money always motivates more.” Only on tasks where effort maps cleanly to measured output. Above a fairness/adequacy threshold, added money often does little for complex work, and on tasks requiring intrinsic engagement it can reduce performance. (Large stakes can even hurt performance on cognitively demanding tasks — a finding from Ariely and colleagues, flagged with caution below.)
• “Performance pay always works.” It works spectacularly on the Safelite quadrant (simple, measurable, individual) and fails predictably off it. Generalizing from the success cases is the single most common management error in this domain.
• “If you want to stop a behavior, fine it.” Sometimes a fine licenses the behavior by pricing it and stripping its moral charge (the daycare result). Ask whether you are forbidding or pricing.
• “Incentives are manipulation, therefore unethical.” Every environment already contains incentives — the default is not “no incentive” but “unexamined incentive.” The ethical question is not whether to shape behavior (you always do) but which incentives, chosen how transparently, in whose interest.
• “Gamification always boosts motivation.” Points, badges, and streaks are extrinsic rewards; layered onto an already-loved activity they can crowd out intrinsic interest, and once withdrawn, engagement can fall below baseline. Gamification is an incentive intervention subject to all the same failure modes.

#8. Real-World Applications

Personal life. Use extrinsic incentives to start habits (crossing the activation threshold) and intrinsic/identity motivation to sustain them. Never monetize a hobby you want to keep loving without expecting some erosion of the joy. Build feedback loops (streaks, logs) that supply the competence signal money would otherwise provide.

Parenting and education. Paying kids to read raises reading now and can lower love-of-reading later (the exact worry Gneezy, Meier & Rey-Biel 2011 raise). Prefer rewards that signal competence (“you’ve become a strong reader”) over rewards that signal control (“read and I’ll pay you”). Grades are an incentive system with a massive multitask problem: they measure test performance, not learning, curiosity, or integrity.

Management. Match the compensation instrument to the task quadrant. Reserve strong individual pay-for-performance for simple, measurable, individual work; for complex collaborative knowledge work, blend modest incentives with fixed pay, team-based measures, and retained subjective judgment. Treat OKRs/KPIs as incentive systems, not neutral thermometers — the moment a KPI carries a bonus, expect it to be gamed.

Product design. Loyalty programs, referral bonuses, and streaks are incentive engineering. Streaks work because they recruit loss aversion (breaking a 100-day streak hurts) and identity (you become “a streak person”). But extrinsic reward structures can crowd out intrinsic use and collapse when the reward stops.

Public policy. Incentive-based policy is now run at civilizational scale, and its empirical record is instructive. Carbon pricing is the flagship. British Columbia’s revenue-neutral carbon tax (introduced 2008 at C$10/tonne, rising to C$30/tonne by 2012, covering ~three-quarters of provincial emissions) is estimated by Murray & Rivers (2015, Energy Policy) to have reduced BC greenhouse-gas emissions by 5–15% with negligible aggregate economic effect — though that headline figure is a synthesis of models, and Pretis (2022, Environmental and Resource Economics) argues the aggregate CO₂ reduction is not yet statistically detectable even as transport emissions clearly fell. For cap-and-trade, Bayer & Aklin (2020, PNAS) find the EU Emissions Trading System saved more than 1 billion tonnes of CO₂ between 2008 and 2016 (~3.8% of EU emissions) despite persistently low permit prices. Sin taxes show a robust pattern — high price pass-through and reliable reductions in purchases of the taxed good, but weaker and contested effects on actual consumption/health: Mexico’s 1-peso/litre soda tax cut taxed-beverage purchases ~6% in year one and ~7.6% over two years, with the largest declines among low-income households (Colchero et al. 2016, BMJ; 2017, Health Affairs); Berkeley’s penny-per-ounce tax cut SSB consumption in low-income neighborhoods ~21% (Falbe et al. 2016, AJPH); Philadelphia’s tax cut in-city taxed-beverage volume ~38% net of cross-border shopping (Roberto et al. 2019, JAMA). Soda taxes are nominally regressive but the burden is small (0.1–1.0% of income for low-income households) and partly offset because low-income consumers are more price-responsive and reap more of the health benefit (Allcott, Lockwood & Taubinsky 2019, QJE/JEP). Finally, conditional cash transfers are an RCT-backed incentive success: Mexico’s Progresa/Oportunidades — averaging ~22% of beneficiary-household income, conditioned on school attendance and clinic visits — raised enrollment (Schultz 2004 and the program’s randomized evaluation report gains such as ~10% for boys and ~20% for girls at some margins), was scaled to millions of households, and became the template for 60+ countries.

AI. Reward modeling and reinforcement learning from human feedback (RLHF) are literally the principal-agent problem in machine form: you specify a reward proxy, and a powerful optimizer finds the cheapest way to maximize the proxy — “specification gaming” / “reward hacking” — which is Goodhart’s Law with a gradient. The catalog of AI systems gaming their reward functions (compiled by researcher Victoria Krakovna and collaborators) is the same phenomenon as Soviet nail factories, and it is a preview of later chapters.

#9. Case Studies

#Failure (crowding-out): the daycare fine

Uri Gneezy and Aldo Rustichini’s “A Fine Is a Price” (Journal of Legal Studies, 2000) is the field’s most-cited crowding-out result. In ten private daycare centers in Israel, parents sometimes arrived late to collect their children, forcing teachers to stay. In six centers (randomly assigned; four were controls), the researchers introduced a small fine (NIS 10) for pickups more than ten minutes late, after four weeks of baseline observation; the fine ran for weeks 5–16 and was removed for weeks 17–20. The deterrence prediction: late pickups fall. What happened: late pickups roughly doubled after the fine, and — the truly damning part — did not return to baseline after the fine was removed.

The mechanism: before the fine, arriving late meant imposing on a teacher who was doing you a favor — a social contract with moral weight. The fine converted this into a market transaction: lateness now had a posted price, and parents felt entitled to buy it. Removing the fine did not restore the moral contract, because the interaction had already been reframed as commercial — you cannot un-ring the bell.

Claim-evaluation caveat (applying Ch. 7’s protocol): this is a single-context, small-sample field experiment — ten centers, one country, one behavior. It is vivid and directional but statistically thin, and a later attempted replication in a survey/vignette setting (Homonoff and colleagues, “Is a Fine Still a Price?”, on MTurk) did not reproduce the original result cleanly. Treat the daycare study as a compelling illustration of a real mechanism (crowding-out via reframing), not as precise quantitative proof. The mechanism is corroborated by other evidence; this one study is not load-bearing on its own.

#Failure (crowding-out at civic scale): nuclear-waste siting

Bruno Frey and Felix Oberholzer-Gee (“The Cost of Price Incentives,” American Economic Review, 1997) surveyed Swiss residents of communities proposed as nuclear-waste repository sites. When asked whether they would accept the repository as a civic duty, 50.8% said yes. When the same respondents were then offered substantial annual compensation to accept it, acceptance fell to about 24.6% — roughly halved. Raising the offered sum barely moved anyone: only a single respondent who had declined the first offer accepted a higher one. Money converted a question of civic responsibility (“will you bear a burden for your country?”) into a market question (“what’s your price for accepting a hazard?”) — and bribery for danger is repugnant in a way that duty is not. This is crowding-out of public spirit, and it generalizes the daycare mechanism to policy.

#Failure (crowding-out of the gift): blood donation

Richard Titmuss’s The Gift Relationship (1970) argued that paying for blood would reduce and degrade the supply by destroying the altruistic gift. For decades this was a plausible-but-untested argument. Carl Mellström and Magnus Johannesson (“Was Titmuss Right?”, Journal of the European Economic Association, 2008) tested it with a field experiment: offering to pay (SEK 50, ≈$7) for becoming a blood donor. The result was conditional: for men, payment had no significant effect; for women, the supply of donors roughly halved when payment was introduced — but the effect vanished when donors could donate the payment to charity (restoring the gift frame). The nuance matters enormously: crowding-out is real but heterogeneous across people and framings, and can be neutralized by design. Titmuss was partly right, and specifically right about how framing carries the effect.

#Success (simple task): Safelite piece rates

Edward Lazear’s “Performance Pay and Productivity” (American Economic Review, 2000) analyzed Safelite Glass Corporation’s switch from hourly wages to piece rates (about $20 per windshield installed, with a guaranteed minimum) across ~2,700 installers over 19 months. Output rose 44%. Critically, Lazear decomposed the gain: about half came from incentive effects (existing workers installing more) and about half from selection/sorting (the piece rate attracted and retained more productive workers, while slow workers left). This is a genuine, well-identified success — and it is instructive precisely because Safelite sits squarely in the quadrant where pay-for-performance should work: output is simple, individually attributable, easily and cheaply measured, and quality is separately verifiable (a badly installed windshield comes back). Move any of those conditions and the same design fails — which is why you cannot copy Safelite’s scheme into a hospital, a school, or a research lab.

#Failure (gaming): Soviet quotas and Wells Fargo

The Soviet planned economy is the museum of Goodhart’s Law. Factories judged on the number of nails produced made vast quantities of tiny, useless nails; judged on weight, they made a few gigantic ones. Glass plants rewarded by the ton produced glass so thick it was opaque; rewarded by square meter, so thin it shattered. Shoe factories hitting unit quotas produced only the most common small size. Each quota was met; each underlying purpose was defeated. The structure is exact: a measure became a target, and effort flowed to the measure.

The same structure appears in market economies. The Wells Fargo scandal (from Ch. 6) is the Soviet nail factory in a bank. The bank’s “Eight is Great” cross-selling goal — CEO John Stumpf’s mantra that the bank should sell at least eight products to every household — drove employees to open unauthorized accounts to hit quotas. Per the CFPB’s September 8, 2016 enforcement action (a $185M fine, and ~5,300 employees fired), the initial estimate was ~2.1 million unauthorized accounts, later revised upward to about 3.5 million in August 2017. Nobody in the Soviet ministry wanted useless nails, and nobody at Wells Fargo headquarters wrote “commit fraud” — both got exactly what they measured and rewarded. The lesson: gaming is not a property of communism or of banking; it is a property of high-stakes measurement without gaming-resistant design.

#Failure (multi-tasking): test-based teacher accountability

The No Child Left Behind era (and the value-added-measurement movement) attached high stakes to standardized test scores. The multitask model predicts the result exactly: effort flows to the measured task (tested subjects, tested item formats) and away from the unmeasured (curiosity, untested subjects, deep understanding) — “teaching to the test” and curricular narrowing. At the extreme, the incentive corrupts the measure directly through cheating. Brian Jacob and Steven Levitt (“Rotten Apples,” Quarterly Journal of Economics, 2003) built an algorithm detecting suspicious erasure/answer patterns in Chicago Public Schools and estimated serious teacher/administrator cheating in at least 4–5% of classrooms annually — and found the cheating rate responded sharply to changes in the stakes. The Atlanta Public Schools scandal was the industrial-scale version: the July 2011 Georgia governor’s special-investigators’ report confirmed cheating in 44 schools implicating 178 educators (including 38 principals), 82 of whom confessed; 35 educators were indicted in 2013 on racketeering and related charges. Each is a predictable multitask/Goodhart response, not a story of unusually bad teachers.

#Personal scale: the monetized hobby

Consider the common experience of turning a loved hobby into a business — the baker who opens a bakery, the photographer who goes pro — and finding the joy quietly drained. This is the overjustification effect (Deci 1971; Lepper, Greene & Nisbett 1973) at life scale: once the activity is done for money, the mind re-attributes the motive from “I do this because I love it” to “I do this because I’m paid,” and when the intrinsic reason is crowded out, the activity feels like work. Contrast a fitness routine that only sticks when reframed as identity (“I’m a runner”) rather than reward (“I’ll treat myself if I run”). Charness and Gneezy (“Incentives to Exercise,” Econometrica, 2009) showed the constructive flip side: paying previously-sedentary people to attend a gym during a one-month intervention produced “an attendance level that is twice as high as the level when people have not been paid, even well after the end of the intervention” — but this was “driven primarily by the impact on non-users,” while “regular users are essentially unaffected.” Money is a good catalyst to cross the activation threshold and a poor fuel for the long haul.

#10. Practical Framework — The Incentive Design Checklist

A six-step protocol for designing any incentive — for an employee, a citizen, a child, or yourself.

Step 1 — Name the desired behavior precisely (not the proxy). Write down the actual behavior you want, not the metric that stands in for it. “Students who understand mathematics,” not “higher test scores.” “Careful, honest sales,” not “accounts opened.” The gap between the behavior and the proxy is where all gaming lives.

• Reflect: What is the real outcome I care about? What metric am I tempted to use instead? What does the metric miss?

Step 2 — Classify the task. Is it simple / measurable / individual → monetary pay-for-performance is safe and often powerful (the Safelite quadrant). Is it complex / hard-to-measure / collaborative → strong incentives will distort; blend modest incentives with fixed pay, subjective judgment, and intrinsic design.

• Reflect: Can output be cleanly attributed to this person? Is quality separately verifiable? How many dimensions does the job really have?

Step 3 — Ask the Goodhart question. What is the cheapest way to game this measure without producing the real outcome? Become the adversary. If a cheap gamed response exists, redesign before deploying.

• Reflect: If I were lazy and self-interested, how would I hit this target? What would I stop doing? Who would I harm?

Step 4 — Audit for crowding-out. Does this reward attach to something people already do for its own sake or out of duty? If yes, prefer recognition and competence feedback over controlling payment; if payment is necessary, frame it as informational (“you’re excellent at this”) rather than controlling (“do this and you’ll be paid”).

• Reflect: Is there pre-existing intrinsic or moral motivation here? Will this reward feel like control or like recognition? What happens to motivation when the reward stops?

Step 5 — Map second-order reactions (link to Ch. 6’s Consequence Depth Protocol). For each affected party ask “and then what?”: Who adapts? Who games? Who exits? Who is affected who wasn’t targeted?

• Reflect: Who are all the agents this touches? What is each one’s best response? What happens at the boundary (cross-border shopping, arbitrage, selection)?

Step 6 — Build feedback and sunset. Treat every incentive as a hypothesis, not a solution. Measure whether it produced the real behavior (Step 1), watch for gaming and crowding-out, and retire measures that have become targets.

• Reflect: How will I know if this is being gamed? What would make me remove it? When will I review it?

#Printable checklist

INCENTIVE DESIGN CHECKLIST
[ ] 1. Named the actual behavior (not the proxy metric)
[ ] 2. Classified the task (simple/measurable → pay; complex → blend)
[ ] 3. Ran the Goodhart pre-mortem (found & closed the cheapest game)
[ ] 4. Audited crowding-out (recognition vs. control; what happens when reward stops)
[ ] 5. Mapped second-order reactions (who adapts / games / exits)
[ ] 6. Set a feedback + sunset rule (incentive = hypothesis)

#Two-week personal experiment

Pick one habit you want to build (e.g., 20 minutes of reading, a short daily walk). Run a controlled comparison of framings:

• Week 1 — Reward framing: give yourself a concrete external reward each day you complete it (a small treat, a token, a checkmark you “cash in”).
• Week 2 — Identity framing: drop the reward; instead, each completion is evidence of who you are (“I’m a reader”). Keep only a streak log.

Record daily completion and, each evening, rate your felt motivation (1–5). At the end, compare: Which framing produced more completions? Which felt more sustainable? Which would survive if you stopped tracking? Most people find identity framing sustains behavior with less depletion — but the point is to generate your own evidence rather than trust the average. (Set realistic expectations: Lally et al. (2010, European Journal of Social Psychology) found that habit automaticity took a median of 66 days to form, with an enormous individual range of 18–254 days — so two weeks tests initiation, not automaticity.)

#11. Criticisms and Limitations

• Contested effect sizes. Crowding-out is real but its magnitude and breadth are genuinely disputed (Deci-Koestner-Ryan vs. Cameron-Pierce). Several flagship results rest on thin or single-context evidence: the daycare study (ten centers, one country, imperfect replication), the generality of crowding-out beyond the specific reward procedures that produce it, and the real-world magnitude of CEO pay incentives.
• The rational-agent assumption. Classical principal-agent theory assumes agents are rational payoff-maximizers. Behavioral economics complicates this at every turn: present bias means future bonuses are steeply discounted, loss aversion means penalties and losses are weighted more than equivalent gains, and social preferences mean agents care about fairness and reciprocity, not just payoff. A contract optimal for a rational agent can be badly wrong for a real one. (Fehr & Gächter’s 2000 public-goods experiments show, for instance, that people will pay out of their own pocket to punish free-riders even with no material benefit — a “social preference” no purely selfish model predicts.)
• “Incentive success” is itself Goodhart-vulnerable. How do we measure whether an incentive worked? Usually by a metric — which can itself be gamed, so evaluations of incentive schemes inherit the very pathology they study.
• The Ariely problem — evidence hygiene. Some widely-repeated “money changes meaning” and dishonesty findings are associated with Dan Ariely, whose 2012 PNAS honesty-pledge paper was retracted in 2021 after the blog Data Colada showed its data were fabricated; independent investigators concluded fabrication “beyond any shadow of a doubt,” and Ariely’s broader body of work has been called into question. The ideas he popularized (market vs. social norms) have independent support from other labs, but any specific Ariely-associated datapoint should be treated with suspicion and cited, if at all, only via independent replication. This is the Ch. 7 Claim Evaluation Protocol applied to a celebrity source.
• The ethics of incentive design. Who is entitled to design other people’s incentives? This connects to Ch. 5’s nudge-ethics debate. Incentive design shades into manipulation when it exploits biases against the target’s own interest, hides the influence, or removes meaningful choice. The defensible line: transparency, alignment with the target’s own goals, and preservation of autonomy. But the line is contested and there is no formula.
• The unresolved deep question. Can intrinsic motivation be deliberately manufactured at scale? SDT gives conditions that support it (autonomy, competence, relatedness), but supporting an existing seed is not the same as creating one where none exists. Whether large institutions can engineer intrinsic motivation, rather than merely avoid destroying it, is genuinely unknown.

#12. Future Directions

• AI and algorithmic incentive design. Two frontiers, opposite in direction. First, AI as designer: personalized motivation systems and AI coaches that tailor incentives to the individual — promising but ethically fraught (hyper-personalized incentives are hyper-effective manipulation). Second, AI as agent: reward modeling and RLHF are the principal-agent problem in silicon, and “reward hacking” / specification gaming is Goodhart’s Law executed by a tireless optimizer. The alignment problem is, in large part, an incentive-design problem — and everything in this chapter about gaming, proxies, and multitasking transfers directly. This is flagged for a later chapter.
• Data-driven sunsetting of bad metrics. The most actionable near-term advance: institutionalizing the retirement of metrics once they show signs of becoming targets — building Goodhart-detection into management systems rather than treating gaming as a surprise.
• Behavioral public-policy innovations. Income-based (day) fines that make sanctions bite equally across the wealth distribution — Finland’s system, in which a former Nokia executive was fined €116,000 ($103,600) for speeding in 2002, is the vivid example — plus advance/refundable tax credits that fight present bias by front-loading rewards, and better-designed conditional transfers.
• Intrinsic motivation as the scarce resource of the AI age. Speculation (author’s extrapolation): as AI automates measurable, incentivizable tasks, the human comparative advantage may concentrate in exactly the domains where extrinsic incentives work worst — creativity, care, judgment, meaning. If so, the capacity to cultivate and protect intrinsic motivation becomes an economic and civilizational priority, not a soft nicety.
• The moral-limits debate. Michael Sandel (What Money Can’t Buy, 2012) argues some goods are corrupted by being priced (friendship, honors, civic duties, bodily integrity). Which domains should be market-priced at all is a live and intensifying social argument — and the empirical crowding-out literature is one input into a question that is ultimately about values.

#13. Recommended Resources

Beginner.

• Drive — Daniel Pink. A readable popularization of Self-Determination Theory (autonomy, mastery, purpose). Read alongside its critiques: Pink overstates how broadly extrinsic rewards backfire; pair with the Cameron-Pierce side of the debate.
• Freakonomics — Levitt & Dubner. Vivid on incentives and gaming (the teacher-cheating and daycare-fine stories originate here for many readers). Read with skepticism: entertaining but sometimes overclaims from single studies.
• The Tyranny of Metrics — Jerry Z. Muller. The essential short book on how measurement-and-reward cultures corrode education, medicine, policing, and the military. Muller’s own summary: “measurement is not an alternative to judgment; measurement demands judgment.”

Intermediate.

• Misbehaving — Richard Thaler. The behavioral-economics backstory, including why real humans respond to incentives differently than theory predicts.
• What Money Can’t Buy — Michael Sandel. The moral-limits-of-markets argument; the philosophical complement to the empirical crowding-out literature.
• Skin in the Game — Nassim Taleb. On the asymmetry of consequences as the deepest incentive. Read critically: rhetorically forceful, empirically loose; take the core model, discount the polemics.
• Why Nations Fail — Acemoglu & Robinson. Incentives at civilizational scale: inclusive vs. extractive institutions as macro-incentive structures.
• Bullshit Jobs — David Graeber. A provocative sociological take on work, meaning, and mis-incentivized labor. Read as provocation, not data.

Advanced (landmark papers).

• Deci (1971), “Effects of externally mediated rewards on intrinsic motivation,” JPSP — the origin of crowding-out.
• Deci, Koestner & Ryan (1999), Psychological Bulletin — the meta-analysis; read with Cameron & Pierce (2001) for the dispute.
• Gneezy & Rustichini (2000), “A Fine Is a Price,” Journal of Legal Studies.
• Frey & Oberholzer-Gee (1997), “The Cost of Price Incentives,” AER.
• Holmström & Milgrom (1991), “Multitask Principal-Agent Analyses,” JLEO — the theoretical keystone.
• Lazear (2000), “Performance Pay and Productivity,” AER.
• Fehr & Gächter (2000), “Cooperation and Punishment in Public Goods Experiments,” AER.
• Gneezy, Meier & Rey-Biel (2011), “When and Why Incentives (Don’t) Work to Modify Behavior,” JEP — the best single synthesis; start here if you read only one.

Influential researchers: Bengt Holmström, Edward Lazear, Uri Gneezy, Bruno Frey, Edward Deci & Richard Ryan, Sendhil Mullainathan, Marianne Bertrand, Daron Acemoglu.

Courses/lectures: any personnel-economics or contract-theory course sequence; the Gneezy-Meier-Rey-Biel JEP article functions as a self-contained course on when incentives fail.

#14. Self-Check

Attempt these from memory before reviewing the chapter.

1. Explain the daycare fine experiment: what happened, and what does it reveal about the difference between a price and a sanction? What are the study’s sample and replication limits?
2. State the multitasking problem in your own words, and give a real example not used in this report.
3. When does performance pay work, and when does it predictably fail? Name the quadrant and the conditions.
4. What is the difference between crowding-out and gaming? Give one clean example of each.
5. Design an incentive for one of your own habits — then apply the Goodhart question to your own design. What is the cheapest way you could game it?
6. Why did offering compensation reduce Swiss acceptance of a nuclear-waste site, and what does this share with the blood-donation result?
7. Explain why the same Jensen–Murphy finding — that CEO wealth changes by only $3.25 per $1,000 of shareholder wealth — can be read as either evidence of an incentive or of a rent-extraction ritual.
8. When money enters a domain, give two conditions under which it activates market behavior and two under which it corrupts a social/moral signal.

Synthesis (not an answer key). If your answers keep returning to a few structural ideas — that incentives change the meaning of an act and not just its payoff; that you reliably get what you measure and reward rather than what you want; that strong incentives on one measurable dimension starve the unmeasured ones; that pre-existing intrinsic or moral motivation is the thing most at risk from control-flavored rewards; and that every incentive invites strategic counter-adaptation you must anticipate — then you have the load-bearing structure. If instead you are recalling isolated study results without the mechanism underneath, re-read §§4, 6, and 9: the studies are illustrations, and the mechanisms are the point.

#15. Knowledge Card

## Knowledge Card — Incentives
- Core terms:
  - Intrinsic vs. extrinsic motivation: doing something for its own sake vs. for a separable reward/punishment.
  - Crowding-out: an extrinsic reward reducing total motivation by displacing intrinsic/moral motivation.
  - Principal-agent problem: aligning an agent's behavior with a principal's goals under information asymmetry.
  - Moral hazard / adverse selection: hidden action after contracting / hidden type before contracting.
  - Multitasking problem: incentivizing the measurable task diverts effort from the unmeasurable ones.
  - Goodhart's Law: when a measure becomes a target, it stops being a good measure (people game it).
  - Skin in the game: alignment achieved when the decision-maker bears their own downside.
  - Pay-for-performance: output-linked pay; powerful on simple/measurable/individual tasks, treacherous off that quadrant.
- Core mental models:
  - "You get what you pay for" — you get the measured-and-rewarded behavior, including its cheapest gamed form.
  - The crowding-out map — money is safe where there's no prior intrinsic motive, risky where there is; prefer recognition to control.
  - The multitask trap — visible tasks crowd out invisible ones under strong incentives.
  - Incentive-compatibility thinking — before deploying a rule, ask "how would I cheaply game this?"
- Connections to prior chapters:
  - Ch. 6 (Second-order Thinking): incentives are the mechanism by which second-order consequences propagate; extends Goodhart, moral hazard, cobra effect, and the Wells Fargo case.
  - Ch. 5 (Behavioral Economics): present bias, loss aversion, and social preferences predict when incentives work or backfire; deepens commitment devices.
  - Ch. 7 (Critical Thinking): the Claim Evaluation Protocol applied to this chapter's own thin/contested/retracted evidence (daycare study, Ariely).
  - Ch. 1 (Deciding Well): confidence tiering and process-over-outcome carried throughout.
- Recommended next chapter: Game Theory — the formal framework of strategic interaction, which supplies the equilibrium logic behind the empirical incentive responses studied here.
- One habit to keep: Before deploying any incentive (on others or yourself), run the Goodhart pre-mortem — "what is the cheapest way to hit this target without producing the real outcome?" — and redesign before launch.