Rewarding A, Hoping for B

The most honestly titled paper in the history of management is "On the Folly of Rewarding A, While Hoping for B," written in 1975 by the organizational behaviour scholar Steven Kerr. It describes something you have probably watched happen. An organization pays to reward one behaviour (A, the one that is easy to measure and easy to pay out on) while hoping people will do a different behaviour (B, the one it actually wants but cannot measure). Rational people do the A that is rewarded and not the B that is hoped for.

Start with the classic scene: the university. Society wants professors to teach well, but promotion, bonuses and prestige hang almost entirely on publications. So rational professors put their effort into publishing and teaching gets squeezed to the margins. Not because they are bad people, but because that is how the reward structure is set. Bring in a more idealistic cohort and the same reward list will train them into the same distribution of behaviour, only at a different speed.

The doctor example is subtler. Society punishes "judging a sick patient healthy" (a missed diagnosis) far more heavily than "judging a healthy person sick" (overdiagnosis), so rational doctors lean toward over-testing and over-intervening. What is written on the reward list is not "make the correct judgment," it is "do not make the kind of mistake that ends up in the news," and behaviour grows toward the latter.

The paper's core sentence deserves quoting in the original: most organisms seek information concerning what activities are rewarded, and then seek to do (or at least pretend to do) those things, often to the virtual exclusion of activities not rewarded. Three of the words hide a blade. "Organisms": Kerr deliberately did not write "employees." This is not an indictment of human nature. It is a behavioural description of anything shaped by rewards. That description does not hinge on human nature. Read fifty years later, the word is precise enough to be prophecy. "Pretend to do": a reward structure does not just change how behaviour is allocated. It also changes whether the behaviour is real. Pretence is a normal product of incentives, not a pathology. "Virtual exclusion": crowding out is not a proportional discount. Unrewarded activity can be pushed to zero.

The rest of the list rewards reading too. In politics, official goals and operative goals have long been separate: vague fine words get rewarded by votes, so vague fine words keep coming. Orphanages and rehabilitation centres undergo goal displacement: what the institution is graded on is its own operation, and the original mission of "making the institution unnecessary" is rewarded by no one. Vietnam's rotation and assessment system rewarded individual risk avoidance rather than winning the war, so the front produced search and evade (nominally searching for the enemy, actually going around them) and even fragging (soldiers turning on the officers who pushed them into danger). MBO-style management (management by objectives: cascading assessment on quantifiable individual targets) rewards short term individual performance while hoping for teamwork and long term investment.

The power of the list is not in any single case but in the same structure spanning government, universities, hospitals, the military and business. Wherever "what is wanted is hard to measure and what is easy to measure is rewarded," the folly replicates itself automatically, with nobody signing off on it. The paper was reprinted verbatim by an Academy of Management journal in 1995, and the fact that a twenty year old list of problems could be republished without revision is itself a review: twenty years on, the folly had not dated.

Two decades later, economics attached a theorem to Kerr's list. Gibbons's 1998 survey points out that as long as B is hard to measure, A is measurable, and effort on the two tasks is substitutable in cost (do more of one and you do less of the other), strong incentives on A will pull effort away from B. This is not a management mistake, it is an equilibrium you can derive step by step from the mathematics (the derivation is in B09's multitask model). Milgrom and Roberts's textbook later codified one part of it as the equal compensation principle: activities the agent values equally must offer equal marginal returns, or the activity with the lower return is abandoned entirely. Note that it is abandoned entirely, not done a bit less. Kerr's observed "virtual exclusion" is, in the model, a corner solution (the optimum landing directly on the boundary where effort is zero).

Having absorbed Kerr, economics also produced an institutional way out. The 1994 model of Baker, Gibbons and Murphy addresses exactly a world where "every objective measure is misaligned": when a quantifiable measure necessarily sits at an angle to the true goal, bringing in a supervisor's subjective assessment as a supplement can improve the contract. But subjective assessment cannot be written into a legal contract and cannot be enforced in court, so it can only be self-enforcing through a relational contract: reputation and the value of continued cooperation keep both sides in line. In everyday terms, when every objective ruler hangs the target crooked, "a person with judgment taking a look" becomes the corrective tool, and that tool works only if both sides care about the long term relationship. This belongs to the same family as Campbell's institutional prescription (B02) and multi-indicator checks and balances (G04): the remedy is not a finer measure but preserving uncontractible judgment outside the measures. The limits of the way out should be marked too: subjective assessment guards against a mis-hung target, not against the assessor being flattered. That second risk switches vehicles and reappears in Y08 (sycophancy: models learning to please human raters). Reward human satisfaction and what you get is flattery.

Line up the three stops the blue branch has made so far. The cobra effect is a bounty opening a business model for manufacturing the problem (B04). The McNamara fallacy is treating the boundary of the measurable as the boundary of the existent (B05). Kerr is the mismatch between the reward list and the wish list. All three can occur in sequence inside one organization: first strike the hard-to-measure B out of assessment (McNamara's step two), then hang a heavy bounty on the easy-to-measure A (Kerr's folly), and finally discover that A is being manufactured in reverse (cobra). The formal models in the back half of the blue branch (B09, B10) will prove that this chain needs nobody acting in bad faith at any link.

One open question, from the interface between this theory and AI: relational contracts are self-enforcing through reputation and repeated play, which assumes "the supervisor's judgment" comes from someone who cares about the long term relationship. When supervisory judgment is replaced by an LLM judge (LLM as judge: using one large model to score another model's output), there is neither reputation nor repeated play between the judged model and the judge, so what is left of the self-enforcement mechanism for subjective measurement? This is an interface that principal-agent theory and alignment research have not yet joined up, and it is the question every LLM judging pipeline is currently answering in the field (the formal structure of information asymmetry continues in B12).

The one-line takeaway: a system delivers the behaviour it rewards, not the behaviour it hopes for; to predict an organization, watch where its money flows, not where its slogans drift.