Perverse Incentives

ChatGPT response · editorial edition

Perverse incentives are rewards or penalties that lead to unintended and undesirable outcomes by encouraging harmful behavior contrary to the interests of the incentive creators.

Gemini response · editorial edition

A perverse incentive makes behaviour that undermines a policy’s goal attractive to participants. For example, a bounty intended to eliminate mice could make breeding mice for the bounty profitable; this is a hypothetical illustration.

ChatGPT response · editorial edition

1. Cobra Effect in Colonial India — unverified anecdote: The familiar story says a bounty encouraged cobra breeding, followed by releases when payment stopped. No reliable documentary basis for that particular episode is established here. It illustrates a possible mechanism—rewarding a supplied pest rather than a reduction in pests—but should not be presented as verified colonial history.
2. Rat Bounty in Hanoi, 1902 — documented case: Historian Michael Vann’s archival research describes payments for rat tails, reports of tailless living rats and rat farming. A tail was rewarded as a proxy for eradication, creating a reason to preserve the supply. This establishes the incentive problem without requiring an exact estimate of the resulting rat population.
3. The Great Sparrow Campaign in China, 1958 — campaign and ecological harm: Sparrow killing pursued a grain-protection goal while disregarding the birds’ role in controlling insects. A 2025 working paper finds evidence of crop losses associated with eradication. Its estimates are research findings, not a claim that this campaign alone caused the famine. The relevant target was killing birds rather than improving harvests; this is harmful compliance with a mistaken policy as well as a warning about incentives.
4. Australian Cane-Toad Introduction, 1935 — failed biological control: Toads introduced to control cane beetles became invasive and harmed native wildlife. The National Museum of Australia documents the introduction and its consequences. The original explanation based solely on jumping height is inadequate. This is principally an ecological policy failure: an incentive example would additionally need to identify whose reward encouraged which harmful behaviour.
5. Mexico City’s “Hoy No Circula,” introduced in 1989 — disputed avoidance mechanism: Restricting weekday driving by licence-plate number gives households a possible reason to acquire another vehicle. Davis’s 2008 study found no air-quality improvement and evidence of more vehicles and a shift toward higher-emitting vehicles. A later travel-survey study also found little reduction in travel but rejected second-car purchasing as the explanation in its setting. The policy’s disappointing results are better supported than a universal account of how households evaded it.
6. The Streisand Effect, 2003 — publicity backfire: A lawsuit seeking to restrict access to a coastal photograph became a subject of public attention; the July 2003 court transcript records the dispute and arguments about that attention. The episode illustrates how attempted suppression can publicize information. It is not automatically a perverse reward scheme: that further diagnosis requires identifying the audience’s or publisher’s incentive to spread it.
7. Dead Sea Scrolls publication — restricted access and delay: The Library of Congress exhibition documents slow publication and the 1991 opening of photographic access. Exclusive control can reduce pressure to share, but that is a mechanism to investigate, not a complete causal explanation. The original claims about a Jordanian copyright grant and consequent black-market sales are not established by this evidence and are withdrawn.
8. South African apartheid job reservation — a different kind of harm: Racial job restrictions deliberately excluded Black workers; their discriminatory purpose should not be redescribed as an innocent unintended effect. The original additional story—that these rules caused mechanization and reduced overall employment, including protected jobs—requires evidence for particular industries and periods and is not established here. A policy can be unjust without also fitting this specific incentive-backfire narrative.
9. Naval suppression and Nassau’s pirates — causal story unestablished: The original asserts that stronger suppression created a stronger “Pirate Republic.” It does not identify a policy, comparison or reliable source supporting that sequence. Suppression might provoke coordination in some settings, but that possibility does not establish this historical explanation. Retain it as a question for research, not as one of the verified cases.
10. The Window Tax in England, introduced in 1696 — documented avoidance: Tax liability tied to windows encouraged owners to block openings or provide fewer of them. Parliament’s account explains the resulting loss of light and ventilation, especially in tenements. The tax made a reduction in the taxable feature attractive even when the reduction worsened living conditions. It does not follow that every historic blind window was created to avoid tax.

Gemini response · editorial edition

1. “The Sparrow of Paris” — unsupported location and title: The original describes rat breeding under an eighteenth-century Paris bounty but supplies no identifiable episode. Do not count it as history. The documented Hanoi case above provides the relevant rat-tail mechanism; changing its city and century would create a second story without a second source.
2. “The London Fog” — unsupported tax narrative: The original says an anti-smoke coal tax caused substitution toward dirtier fuels. That particular policy and outcome are not established here. The 1889 parliamentary debate on coal duties discusses revenue and public works, which is not evidence for the claimed anti-pollution purpose. Substitution toward untaxed pollution is a possible design risk, not proof that this episode occurred.
3. The Great Sparrow Campaign, China — same case, different emphasis: Encouraging the destruction of a grain-eating bird ignored its insect-control role. The historical campaign illustrates why visible compliance can be a poor measure of agricultural success. It is the same episode as in the first list, not an additional independent case, and should not be used as a single-cause explanation of the famine.
4. “The Cobra Effect (1960s India)” — unsupported dating: Neither the 1960s date nor this version of the bounty story is established here. The mechanism can be taught as a hypothetical: if breeding a pest earns more than it costs, a bounty can reward increasing its supply. That does not turn the illustration into a documented Indian government programme.
5. “The Boeing Bonus (1970s)” — unsupported allegation: No identified bonus scheme, accident investigation or source is supplied for the alleged chain from engineers’ production bonuses to fatal crashes. This edition does not assert that story. In a hypothetical production scheme, paying only for quantity could discourage time spent on quality checks; applying that explanation to a named company requires specific evidence.
6. “The Texas Two-Step (1990s)” — unsupported programme: The named doctor-incentive scheme and its alleged unnecessary surgeries are not established here. The general mechanism is understandable: payment per procedure can reward volume rather than patient benefit. That incentive alone proves neither that a procedure was unnecessary nor that this particular Texas story occurred.
7. Zero-tolerance drug policies — an underspecified category: The original spans many jurisdictions and decades, so it is not one testable historical story. An arrest-count target could encourage low-level arrests over work that better reduces harm; this is a conditional incentive mechanism. Assess an actual policy using its rules, enforcement choices and outcomes rather than inferring the mechanism from imprisonment alone.
8. Subprime mortgages and the financial crisis — documented incentive concerns: The Financial Crisis Inquiry Commission examined compensation linked to loan volume and short-term gains while loan performance and losses unfolded later. Such arrangements could reward originating loans without adequately bearing their longer-term risks. This was one mechanism among several investigated causes; the report also contains dissenting interpretations.
9. Clickbait — a general mechanism rather than one historical incident: If advertising revenue or promotion depends heavily on clicks, a publisher may profit from an enticing headline even when the article disappoints or misleads. The diagnosis requires the actual payment or ranking rule and evidence of the response. A striking headline is not itself proof of deception, and no named platform’s current policy is asserted here.
10. Teacher evaluation and testing — distinguish instruction from manipulation: Rewards attached to scores can encourage narrow preparation or alteration of answers, but teaching tested material can also be legitimate. Jacob and Levitt’s research used follow-up testing to distinguish suspicious gains from gains consistent with effective teaching. This is a concrete research example, replacing the original vague date label; it does not imply that every high-scoring teacher cheated.

ChatGPT response · editorial edition

To reduce the chances of introducing perverse incentives when creating policies, consider employing the following principles and heuristics:

1. Impact Assessment: Map the intended benefit, the rewarded action and who bears the cost. For each participant, ask for the cheapest way to earn the reward while defeating the goal. Historical examples help only when their facts and mechanisms survive checking.
2. Incremental Implementation: Pilot where feasible, with comparison groups or baseline measures suited to the policy. Predetermine warning signs and stopping conditions. A small pilot may miss responses that become profitable only at larger scale.
3. Feedback Loops and Flexibility: Establish mechanisms for continuous feedback from those affected by the policy. Be prepared to make adjustments and amendments to the policy as unintended consequences are identified.
4. Align Incentives with Outcomes: Suppose a support team is paid for closing tickets quickly. Counting repeat contacts and whether the issue was resolved can reveal closures that merely shift work to the customer. Check whether the revised measures create new ways to hide unresolved cases.
5. Broad Stakeholder Engagement: Engage a diverse group of stakeholders in the policy design process, including those who will be directly affected. This can uncover potential perverse incentives that policymakers might not anticipate.
6. Avoid Reliance on a Single Target: Combine a small number of interpretable measures with independent quality checks. Broader measures are not automatically better: a complicated score can hide failures and make responsibility harder to locate.
7. Transparent Communication: Explain the objective, the rules and how people can challenge errors. Clarity reduces misunderstanding, but intentional gaming can occur precisely because the rule is understood. Pair communication with checks on whether the objective is being met.
8. Simplicity and Understandability: Make the rule easy to explain and its exceptions reviewable. Simplicity helps participation but does not ensure alignment: the very simple rule “pay per tail” is vulnerable because its proxy can be supplied without eradication.
9. Monitoring and Evaluation: Compare the rewarded metric with outcomes measured through a partly independent channel. Rising ticket closures alongside rising repeat complaints warrant investigation, not immediate praise. Specify who can revise the policy and when review occurs.
10. Ethical Considerations and Equity: Ask who can profit from the rule and who absorbs the harm. Test whether rewarding easy cases causes participants to avoid people with more difficult needs. A good average outcome can conceal that exclusion.

By applying these principles thoughtfully, policymakers can better anticipate and mitigate the risks of perverse incentives, leading to more effective and beneficial outcomes.

Gemini response · editorial edition

Here are some principles and heuristics to reduce perverse incentives when crafting policies:

Consider All Actors and Their Motivations:

• Think Like a Gamer: Rehearse how someone could maximize the reward while complying with the letter of the rule. This is a stress test, not an assumption that everyone is malicious. Also ask how an honest participant with limited time or resources would respond.
• Multi-Stakeholder Analysis: Identify all parties affected by the policy (citizens, businesses, government agencies) and consider their potential responses.

Focus on Long-Term Goals:

• Don’t Just Track Outputs: Distinguish completed activity from achieved benefit. Arrest counts do not directly measure safety, just as test scores do not capture all learning. Use the output as one signal and look for costs displaced elsewhere.
• Test Changed Conditions: Consider new technology, higher participation and changes in relative costs. A loophole that is unprofitable today may become attractive tomorrow. Record assumptions that would trigger another review.

Design for Transparency and Accountability:

• Clear Goals and Metrics: Define success before choosing a measure, then state what the measure misses. Include an independent check that would reveal an improving score alongside a worsening outcome.
• Monitoring and Evaluation: Track both suspected gaming and the burdens created by enforcement. If investigators are rewarded solely for detected violations, they too acquire an incentive to optimize a count rather than accuracy.

Embrace Flexibility and Iteration:

• Pilot Programs: Test policies on a small scale before full implementation to identify and address perverse incentives early.
• Adaptive Management: Be prepared to adjust the policy based on real-world data and feedback from those affected.

Remember: There’s no foolproof method, but these principles can help design policies that achieve their intended goals and avoid unintended pitfalls.

ChatGPT response · editorial edition

Begin by distinguishing deliberate deception, an honest error and lawful adaptation to a poorly designed rule. Here “quarantine” means a temporary, proportionate restriction on a risky transaction or account function while evidence is checked. Removing one participant will not repair an incentive that keeps producing the same behaviour.

1. Data Analysis and Monitoring: Use unusual patterns to select cases for investigation. A spike in claims may reflect manipulation, a disaster or a delayed reporting batch. Seek corroboration tied to the alleged action before treating an anomaly as proof of misconduct.
2. Machine Learning and AI: A model can prioritize review if its performance has been evaluated on relevant cases. Measure missed abuse and false accusations, including across affected groups. Past enforcement labels may encode past mistakes; automatic retraining does not guarantee improvement.
3. Verification and Authentication: Choose checks suited to the suspected abuse. A second authentication factor can reduce account takeover but does not establish that the account owner’s claim is honest. Avoid burdensome identity checks that do not address the actual failure mode.
4. Community Reports: Provide a way to submit specific evidence and correct mistakes. Reports are leads, not verdicts; coordinated false reporting can itself game the system. Do not reward the number of accusations without checking their quality.
5. Transparent Rules and Consequences: Clearly communicate the rules of the system and the consequences for attempting to game it. Knowing that there are stringent checks in place and that penalties will be enforced can deter bad actors.
6. Audits and Reviews: Combine targeted investigation with a genuinely random sample. Targeted checks find suspicious cases; random sampling can reveal abuse missed by the suspicion rules and help estimate how common it is. Track overturned decisions too.
7. Information Sharing: Share relevant, substantiated incident information through authorized channels, with its uncertainty and correction history attached. An unreviewed blacklist can multiply a single false accusation across organizations.
8. Dynamic Responses: When the risk warrants intervention before investigation finishes, limit the vulnerable action—for example, hold an unusual payout—rather than automatically closing every service. Specify an expiry or review deadline and a route to restore access.
9. Policy and Authority: State the prohibited conduct, what evidence supports each response and who can authorize it. Separate internal rule enforcement from referral of evidence of possible crime. Disagreement with a policy is not itself abuse.
10. Reviewable Automated Decisions: Keep a record of the evidence and reason for a restriction, allow correction, and provide human review able to reverse the outcome. An explanation that merely repeats a risk score does not show why this person’s action justified the restriction.

Employing these strategies requires a balance between being effective in deterring and catching bad actors and ensuring that legitimate users are not unduly burdened or falsely accused. Continuous evaluation and adaptation of these strategies are crucial as tactics to game the system evolve.

Gemini response · editorial edition

Use detection to identify conduct for investigation, and use temporary restrictions to contain a specified risk. Calling someone a “bad actor” should follow evidence of deliberate abuse, not substitute for it:

Data Analysis and Monitoring:

• Track Anomalies: Compare like cases and ask what else could explain the pattern. Geographic clustering might reflect a local need or a shared reporting channel. Check evidence of coordination before inferring organized manipulation.
• Risk Scoring: Use relevant, tested indicators to prioritize checks rather than automatically punish. Record the error rates and investigate whether association-based indicators merely penalize people who share an address, workplace or device.

Enhanced Verification:

• Additional Verification: Match the check to the problem: secure account access, verify a duplicate claim or confirm a disputed transaction. Multi-factor authentication uses distinct kinds of credentials; a phone call, biometric check and document request are not automatically interchangeable or sufficient.
• Random and Targeted Audits: Distinguish random selection from selection because a case looks suspicious. Use each for its purpose, with a recorded sampling method, so the results do not exaggerate the prevalence of abuse.

Human Expertise and Investigation:

• Trained Analysts: Invest in personnel trained to identify red flags and investigate potential fraud.
• Collaboration: Involve appropriately authorized investigators when evidence and seriousness justify it. Share the relevant facts and their limits. Ordinary policy disagreement or a correctable mistake does not require a criminal response.

Quarantine Strategies:

• Restrict Access When Warranted: Start with the least disruptive restriction that contains the demonstrated risk. More serious or lasting exclusion needs stronger grounds and review; a suspicion score alone should not make a temporary hold permanent.
• Limit the Vulnerable Function: Restrict a particular action while preserving unrelated legitimate use where feasible. State the reason, what information could resolve the concern and when a decision will be reviewed.
• Escalation and Alerting: Distinguish a pending concern from a confirmed finding. Attach corrections and reversals to shared records so the same mistaken flag does not repeatedly exclude a person.

Remember:

• Balancing Security and Access: Striking a balance between security and user experience is crucial. Overly stringent measures can discourage legitimate participation.
• Adapting to Evolving Tactics: Bad actors constantly develop new methods. Regularly review and update detection techniques to stay ahead.
• Transparency and Communication: Explain the conduct rules and review process while protecting information that would enable evasion. Evaluate investigators on accurate, timely resolution rather than raw numbers of accounts blocked.

By combining data analysis, enhanced verification, human expertise, and appropriate quarantine strategies, you can create a more robust system that discourages manipulation and protects its integrity.

ChatGPT response · editorial edition

Quiz

1. What is a perverse incentive?
2. Which documented rat-bounty case illustrates rewarding a proxy that can be supplied while the pest survives?
3. What ecological role did the Great Sparrow Campaign overlook, and what limit should accompany its use as a famine explanation?
4. How did cane-toad introduction backfire, and why is it not automatically an incentive example?
5. What did the Mexico City studies find, and why is the second-car explanation qualified?
6. What does the term “The Streisand Effect” refer to?
7. What is established about restricted access to the Dead Sea Scrolls, and which original causal claims remain unsupported?
8. Why should apartheid’s intended racial exclusion be distinguished from the original employment-backfire claim?
9. Why is the original story that naval suppression strengthened Nassau’s pirates not counted as verified?
10. What was a negative consequence of the Window Tax in England?

Answers

1. Rewards or penalties that lead to unintended and undesirable outcomes by encouraging harmful behavior.
2. Hanoi in 1902: payments for rat tails encouraged practices inconsistent with eradication. No exact population increase is established here.
3. Sparrows also controlled crop-eating insects. Research supports an ecological contribution to crop losses; the campaign is not a complete explanation of the famine.
4. The toads became invasive and harmed native wildlife. A failed ecological intervention needs a separate account of rewards and behavioural responses to establish a perverse incentive.
5. Davis found no air-quality improvement and evidence of vehicle-fleet changes. A later study found little reduction in travel but rejected second-car purchasing as the mechanism in its setting. Different evidence and policy settings must be distinguished.
6. Efforts to hide, remove, or censor information can result in increased publicity.
7. Publication was slow and photographic access expanded in 1991. The asserted copyright-grant and black-market causal story is not established by the evidence cited.
8. Racial exclusion was deliberate. The further claim that job reservation caused mechanization and overall employment losses requires specific causal evidence not established here.
9. It lacks a sourced policy sequence and evidence that suppression caused the asserted strengthening; a possible coordination mechanism is not historical proof.
10. People bricked up windows, leading to poor ventilation and health problems.

Gemini response · editorial edition

Perverse Incentives and Policy Design Quiz

1. The term “perverse incentives” refers to situations where a reward or punishment unintentionally:
   a) Increases costs.
   b) Encourages undesired behavior.
   c) Creates confusion.
   d) Limits participation.
2. Which documented case illustrates paying for rat tails as an unreliable proxy for eradication?
   a) The Great Sparrow Campaign
   b) The unverified Boeing bonus story
   c) Hanoi’s rat bounty in 1902
   d) Clickbait
3. Which response to score-only teacher evaluation would illustrate a perverse incentive?
   a) Improving understanding of the tested material
   b) Strengthening useful reasoning skills
   c) Drilling only predictable test items while neglecting broader learning
   d) Giving students helpful feedback
4. When designing policies to reduce the chance of perverse incentives, it’s important to:
   a) Set clear and measurable goals.
   b) Only consider easily tracked outputs.
   c) Focus solely on short-term benefits.
   d) Ignore the motivations of those involved.
5. The “Think Like a Gamer” approach to policy design involves considering:
   a) How complex the policy will be to implement.
   b) How the policy might be misinterpreted.
   c) How people might exploit the system for personal gain.
   d) How the policy will affect public perception.
6. Pilot programs are a valuable tool for identifying perverse incentives because they allow for:
   a) Faster implementation on a larger scale.
   b) Early detection of unintended consequences.
   c) Reduced costs associated with policy changes.
   d) Increased public support for the policy.
7. What can a sudden, unexplained spike in claims provide?
   a) Proof that critics are dishonest
   b) A reason for investigation, requiring corroboration
   c) Automatic grounds for permanent exclusion
   d) Evidence that every claimant is cheating
8. What does multi-factor authentication most directly help with?
   a) Proving every claim is honest
   b) Establishing the user’s motives
   c) Protecting access against compromised credentials
   d) Proving a transaction is beneficial
9. What should guide a temporary restriction while suspected abuse is investigated?
   a) Public shaming
   b) Proportionality, a review deadline and a route to correct errors
   c) Indefinite exclusion based on a score
   d) Restricting every service regardless of the risk
10. When might referral to law enforcement be appropriate?
    a) A minor good-faith mistake
    b) Substantiated evidence of serious possible criminal conduct, through authorized channels
    c) An unclear policy interpretation
    d) Public disagreement with the policy

Answer Key:

1. (b)
2. (c) Hanoi, 1902; the Paris story is not established.
3. (c) The problem is sacrificing the educational goal for the score, not teaching tested material as such.
4. (a)
5. (c)
6. (b)
7. (b) An anomaly is a lead, not a finding of guilt.
8. (c) Account security is distinct from the truth of a claim.
9. (b) Contain the risk while making an erroneous restriction reversible.
10. (b) Serious evidence and proper authority, not ordinary disagreement.

ChatGPT response · editorial edition

1. How do perverse incentives reveal the complexity of human behavior in response to policy measures?
2. Discuss the ecological consequences of introducing non-native species as a solution to pest control, using the example of cane toads in Australia.
3. How did the Great Sparrow Campaign make compliance with a target diverge from agricultural success? Why is the campaign only part of a famine explanation?
4. What design risk does the cobra-bounty anecdote illustrate, and how should you teach it when its historical basis is not established?
5. How do policies like Mexico City’s “Hoy No Circula” illustrate the challenge of addressing environmental issues through regulatory actions?
6. The Streisand Effect shows the impact of attempting to suppress information. How does this phenomenon play out in today’s digital age, especially on social media?
7. How might exclusive research access discourage sharing, and what evidence would distinguish that incentive from other causes of publication delay?
8. How would you separate apartheid’s intended racial exclusion from any claimed unintended economic response? What evidence would each claim require?
9. What evidence would be needed to test the claim that suppressing piracy strengthened pirate communities, and what rival explanations should be considered?
10. What does the Window Tax in England teach us about the use of proxies for taxation and its potential pitfalls?
11. How can policymakers better anticipate and mitigate the risks of perverse incentives when designing policies?
12. What role does ethical consideration play in preventing perverse incentives in policy-making?
13. How can the use of data analytics and AI help in identifying and mitigating the effects of perverse incentives in various systems?
14. Discuss the importance of stakeholder engagement in identifying potential perverse incentives in policy design.
15. Reflect on the balance between using technology to monitor and control bad actors and the risk of infringing on privacy and personal freedoms. How can systems be designed to maintain this balance?

Gemini response · editorial edition

Perverse Incentives and Policy Design Discussion Questions

1. Can you think of any personal anecdotes where a reward or punishment backfired and produced an unintended consequence?
2. Beyond the examples provided, what are some potential dangers of perverse incentives in the technological age (e.g., social media algorithms)?
3. Can a policy with some harmful incentives still produce net benefits? What would justify tolerating a particular harm instead of redesigning the reward?
4. Is it realistic to expect policy designers to anticipate all potential perverse incentives? Why or why not?
5. How can we balance the need for clear and measurable goals in policy with the potential for unforeseen consequences?
6. Should “thinking like a gamer” be the only approach to policy design, or are there other valuable perspectives to consider?
7. What are the ethical implications of using data analysis to track and potentially restrict the actions of individuals?
8. How can we strike a balance between robust verification methods and a user-friendly experience when designing online systems?
9. When quarantining bad actors, how do we ensure fairness and avoid unfairly penalizing honest mistakes?
10. Is collaborating with law enforcement always the best solution for identifying and addressing bad actors exploiting a system? Why or why not?
11. How can we encourage a culture of ethical participation in online systems and discourage attempts to manipulate them for personal gain?
12. What role can education play in raising awareness about perverse incentives and their impact on policy design?
13. How can we ensure that policy evaluation goes beyond simply measuring outputs and considers the broader impact on society?
14. Are there situations where short-term perverse incentives might be acceptable if they lead to long-term positive outcomes? Discuss the potential risks and benefits.
15. In an ideal world, how would we design and implement policies to maximize their intended benefits and minimize the risk of unintended consequences?