Chapter 21 Exercises
How to use these exercises: Work through the parts in order. Part A builds recognition skills, Part B develops analysis, Part C applies concepts to your own domain, Part D requires synthesis across multiple ideas, Part E stretches into advanced territory, and Part M provides interleaved practice that mixes skills from all levels.
For self-study, aim to complete at least Parts A and B. For a course, your instructor will assign specific sections. For the Deep Dive path, do everything.
Part A: Pattern Recognition
These exercises develop the fundamental skill of recognizing cobra effects across domains.
A1. For each of the following scenarios, identify (i) the intended outcome of the incentive, (ii) the strategic response it actually provoked, and (iii) whether the outcome was the opposite of the intention (a true cobra effect) or merely a failure to achieve the intention (an ineffective incentive).
a) A hospital rewards doctors for reducing patient readmissions within 30 days. Doctors begin classifying returning patients as new admissions rather than readmissions, and some delay treatment of patients who are likely to need follow-up care until after the 30-day window.
b) A city offers a tax credit for installing solar panels. Wealthy homeowners install panels on south-facing roofs that were already energy-efficient, while the subsidy budget is exhausted before low-income households (who would benefit most) can apply.
c) A school gives students prizes for reading books. Students begin reading the shortest, easiest books they can find and rushing through them to maximize their prize count, comprehending and retaining less than they did before the program.
d) A country imposes a luxury tax on yachts to raise revenue from the wealthy. Wealthy buyers purchase their yachts in neighboring countries, and the domestic yacht-building industry collapses, putting middle-class workers out of jobs.
e) A software company pays developers a bonus for every bug they close. Developers begin splitting large fixes into many small ones, writing trivial bug reports to close immediately, and neglecting important but difficult bugs that take longer to resolve.
f) A ride-sharing platform implements surge pricing to incentivize drivers to work during high-demand periods. Some drivers log in during surge pricing but decline rides in normal-priced areas, reducing availability for passengers who are not in surge zones.
g) A government offers a bounty for feral pig kills to control an invasive species. Hunters begin trapping and breeding feral pigs to ensure a steady supply of bounty-eligible kills.
h) A social media platform penalizes accounts that receive many reports of harmful content. Users begin mass-reporting accounts they disagree with to get them penalized, regardless of whether the content is actually harmful.
A2. Classify each of the following as primarily an example of (a) the cobra effect, (b) Goodhart's Law, (c) iatrogenesis, (d) the Streisand effect, or (e) a welfare cliff. Some may involve more than one pattern. Explain your reasoning.
a) A government classifies a leaked document as top secret, drawing far more media attention than the document would have received if ignored.
b) A pollution tax is set too low, and companies treat it as the "price of polluting" rather than an incentive to reduce pollution. Total pollution increases because the tax legitimizes pollution.
c) A university ranks faculty by publication count. Faculty begin publishing more papers of lower quality, fragmenting single studies into "least publishable units."
d) A welfare program provides free school lunches to children from families below the poverty line. Families just above the line cannot afford lunches, and their children go hungry while children from families just below the line eat free.
e) A central bank lowers interest rates to stimulate lending. The cheap money fuels speculative investment in housing, creating a bubble that eventually crashes.
f) A fire department is funded based on the number of fires responded to. The department has no financial incentive to invest in fire prevention, which would reduce its funding justification.
A3. For each of the following pairs, identify which incentive structure is more vulnerable to cobra effects and explain why.
a) Incentive A: Paying farmers per acre of wetland preserved. Incentive B: Paying farmers per species of wildlife documented on their property.
b) Incentive A: Paying teachers based on student test scores. Incentive B: Paying teachers based on student improvement from their baseline scores.
c) Incentive A: Rewarding police officers for arrests made. Incentive B: Rewarding police officers for crimes reported in their district declining over a three-year period.
d) Incentive A: Offering a carbon tax that increases annually. Incentive B: Offering tradable carbon credits with no expiration date.
A4. The chapter identifies five "laws of perverse incentives." Match each of the following scenarios to the law it best illustrates.
Laws: (1) Every incentive creates its own ecology. (2) The more valuable the incentive, the more creative the gaming. (3) Proxies are always vulnerable. (4) The harder the system is to observe, the more vulnerable it is to cobra effects. (5) Removing an incentive can be worse than never having created it.
a) A company eliminates its employee wellness program. Employees who had developed exercise habits supported by the program revert to sedentary behavior, and their health outcomes become worse than they were before the program began.
b) A government program to reduce homelessness counts the number of people placed in housing. Agencies place people in substandard housing that they quickly leave, generating high placement numbers but no real reduction in homelessness.
c) A cryptocurrency exchange offers a reward for identifying security vulnerabilities. A researcher discovers a vulnerability worth $100 million if exploited, and the $50,000 bounty feels insulting by comparison.
d) A charitable organization measures its impact by the number of mosquito nets distributed. Nets are distributed to communities that do not need them, while communities with the highest malaria burden receive fewer nets.
e) An international development program measures school enrollment in a developing country. Schools enroll children who never attend, and teachers teach to empty classrooms while collecting salaries for their enrollment numbers.
A5. Identify three incentive structures in your daily life -- at work, in your education, in your interactions with technology or government -- that you suspect are producing cobra effects. For each:
a) Describe the intended behavior the incentive is meant to motivate. b) Describe the actual behavior the incentive produces. c) Explain why the actual behavior undermines the incentive's intended goal. d) Propose a redesigned incentive that would be more resistant to gaming.
Part B: Analysis
These exercises require deeper analysis of cobra effect patterns.
B1. The Incentive Ecology Mapping. Choose one of the following incentive systems and map its full incentive ecology:
- Academic tenure (publish or perish)
- Health insurance with deductibles and copays
- Performance-based executive compensation (stock options)
- Standardized testing in K-12 education
- Social media engagement metrics (likes, shares, comments)
- Carbon offset markets
For your chosen system:
a) Identify the intended behavior the incentive is designed to motivate.
b) List at least five strategic responses the incentive actually produces, including both intended and unintended responses.
c) For each unintended response, explain why it is rational given the incentive structure.
d) Identify any responses that qualify as full cobra effects -- strategic responses that make the original problem worse.
e) Assess whether the incentive system is, on net, achieving its goals or undermining them.
f) Propose a mechanism design intervention that would reduce the most harmful unintended responses while preserving the beneficial ones.
B2. The Welfare Cliff Calculator. Using publicly available data on social welfare programs in your country or state:
a) Identify at least three means-tested programs with specific income eligibility thresholds.
b) Calculate the combined benefit amount for a hypothetical family just below the lowest threshold.
c) Calculate the combined benefit amount for the same family at income levels $1,000, $5,000, $10,000, and $20,000 above the lowest threshold.
d) Plot the family's total resources (earned income plus benefits) as a function of earned income. Identify any ranges where total resources decrease as earned income increases.
e) Calculate the effective marginal tax rate (the percentage of each additional dollar of earnings that is lost to benefit reductions) at each transition point.
f) Propose a redesigned benefit structure that eliminates welfare cliffs while maintaining the same total cost to the government.
B3. The Streisand Effect in the Age of Social Media. Research three real-world examples of the Streisand effect (beyond the original Streisand case) and for each:
a) Describe the information someone attempted to suppress.
b) Describe the suppression mechanism used (legal action, takedown notice, censorship, etc.).
c) Quantify, if possible, the amplification effect (how much more widely the information was disseminated because of the suppression attempt compared to what would have happened without it).
d) Analyze whether the suppressor could have predicted the Streisand effect. Was the amplification a predictable consequence of the suppression mechanism, or was it genuinely surprising?
e) Propose an alternative strategy that might have achieved the suppressor's legitimate goal without triggering the Streisand effect.
B4. Mechanism Design Challenge. Design an incentive system for one of the following goals. For each design, explicitly address how your system resists each of the five laws of perverse incentives.
a) Reducing traffic congestion in a major city without disadvantaging low-income commuters.
b) Incentivizing pharmaceutical companies to develop antibiotics for which there is a clear medical need but limited market demand.
c) Encouraging citizens to report tax fraud without creating incentives for false reports or harassment.
d) Motivating students to learn deeply rather than simply to earn high grades.
Part C: Application to Your Own Domain
These exercises connect the cobra effect to your area of expertise.
C1. Identify the most important incentive structure in your professional domain -- the one that most powerfully shapes how practitioners behave. Then:
a) Describe the behavior the incentive is intended to produce.
b) Describe at least three unintended behaviors the incentive actually produces.
c) For each unintended behavior, explain whether it is merely a nuisance, a significant distortion, or a full cobra effect that makes the underlying problem worse.
d) Assess whether the profession is aware of these cobra effects. If so, why do they persist? If not, why have they not been recognized?
e) Design a replacement incentive structure that would reduce the cobra effects while preserving the intended behavior. Explain why your replacement is more resistant to gaming.
C2. Identify a case from your professional experience where an incentive produced a cobra effect -- where the strategic response to the incentive made the problem worse. Analyze the case using the five laws of perverse incentives:
a) What ecology of strategic responses did the incentive create?
b) How creative was the gaming, and was the creativity proportional to the value at stake?
c) What proxy was used, and how did agents exploit the gap between the proxy and the underlying goal?
d) What was unobservable in the system, and how did the lack of observability facilitate gaming?
e) What happened (or would happen) if the incentive were removed? Would the system return to its pre-incentive state, or would the removal create additional problems?
C3. Design an "incentive audit" protocol for your professional domain. The protocol should:
a) Identify all major incentives operating in the domain (formal and informal, monetary and non-monetary).
b) For each incentive, assess its vulnerability to each of the five laws of perverse incentives.
c) Establish a monitoring system that would detect emerging cobra effects early.
d) Define trigger conditions that would indicate when an incentive should be modified or removed.
e) Specify a decision-making process for modifying incentives that accounts for the fifth law (removing an incentive can be worse than never having created it).
Part D: Synthesis
These exercises require integrating ideas across multiple chapters.
D1. Cobra Effects and Feedback Loops. Chapter 2 introduced positive (reinforcing) and negative (balancing) feedback loops.
a) For the cobra bounty example, diagram both the intended feedback loop (negative/balancing) and the actual feedback loop (positive/reinforcing) that emerged. Label each link in both loops.
b) Identify the point in the system where the loop switched from balancing to reinforcing. What conditions allowed this switch to occur?
c) Design a modified bounty system that includes a secondary feedback loop specifically to detect and counteract the cobra effect. What information would need to flow to which decision-makers?
d) The chapter argues that the welfare cliff creates a "poverty attractor." Using the feedback loop language from Chapter 2, explain what makes this attractor stable and what perturbation would be required to escape it.
D2. Cobra Effects and Goodhart's Law. Chapter 15 showed that metrics used as targets are corrupted by optimization pressure.
a) For three examples from Chapter 21 (cobra bounty, emissions trading, welfare cliff), identify the Goodhart metric -- the measure that became a target -- and explain how the metric decoupled from the underlying goal.
b) The chapter claims that "every cobra effect involves a Goodhart failure, but not every Goodhart failure escalates to a full cobra effect." Construct examples of each: a cobra effect (Goodhart failure that makes the problem worse) and a Goodhart failure that merely fails to solve the problem without making it worse.
c) Design a metric for one of the examples that is resistant to both Goodhart corruption and cobra effects. Is this possible? What trade-offs must you accept?
D3. Cobra Effects and Iatrogenesis. Chapter 19 analyzed how well-intentioned interventions produce harm.
a) The chapter argues that every cobra effect is also an instance of iatrogenesis. Construct a precise argument for this claim, identifying the "healer," the "patient," and the "harm" in three cobra effect examples.
b) Is there a meaningful distinction between a cobra effect and iatrogenesis, or are they the same pattern with different labels? Make the case for each side.
c) The five laws of perverse incentives from this chapter and the Intervention Calculus from Chapter 19 both provide frameworks for predicting when an intervention will go wrong. Compare the two frameworks. Are they complementary, overlapping, or contradictory?
D4. Cobra Effects and Redundancy. Chapter 17 argued that redundancy is not waste.
a) Explain how the cobra bounty system lacked redundancy. What would a "redundant" cobra control system look like?
b) For the emissions trading example, explain how the offset system's lack of redundancy (relying solely on offset certificates as proof of emission reduction) made it vulnerable to cobra effects.
c) Propose a general principle: "Incentive systems should include redundancy in their measurement mechanisms." Explain what this means in practice and illustrate with an example.
Part E: Advanced Challenges
These exercises push beyond the chapter's material into deeper or more speculative territory.
E1. Research the concept of "mechanism design" in economics (the work of Leonid Hurwicz, Eric Maskin, and Roger Myerson, who won the Nobel Prize in Economics in 2007). Explain the revelation principle and its implications for designing cobra-resistant incentive systems. Does the revelation principle suggest that cobra effects are avoidable in principle, or does it reveal fundamental limits on what mechanism design can achieve?
E2. The chapter discusses the Gibbard-Satterthwaite theorem, which shows that in many settings, no mechanism can be simultaneously incentive-compatible, efficient, and budget-balanced. Research this theorem and its implications. Choose a real-world incentive system (voting, auctions, matching markets, or carbon trading) and explain which property the system sacrifices and why.
E3. The concept of "incentive ecology" in this chapter is informal. Attempt to formalize it. Define an incentive ecology as a set of agents, a set of possible strategies for each agent, a payoff function, and an incentive structure. Then, using this framework, define a "cobra effect" formally: under what conditions does the equilibrium behavior of the agents produce an outcome that is opposite to the designer's intention? Can you specify necessary and sufficient conditions?
E4. The chapter argues that "the harder the system is to observe, the more vulnerable it is to cobra effects." Research the concept of "information asymmetry" in economics (Akerlof, Spence, Stiglitz). How does information asymmetry create vulnerability to cobra effects? Is the solution more information (transparency), or can information itself create new cobra effects? (Hint: consider how transparent grading rubrics can be gamed more effectively than opaque ones.)
E5. Design a cobra-resistant incentive system for one of the following deeply challenging contexts. For each, explain your design, identify its remaining vulnerabilities, and assess whether a cobra-free incentive system is achievable even in principle:
a) Compensating executives in a way that aligns their interests with long-term shareholder value and broader stakeholder welfare.
b) Paying physicians in a way that incentivizes good patient outcomes without creating incentives to cherry-pick healthy patients or undertreat costly conditions.
c) Scoring academic research in a way that incentivizes quality, originality, and social impact without creating incentives to game citations, fragment studies, or pursue fashionable topics.
Part M: Mixed Practice (Interleaved Review)
These exercises mix concepts from Chapters 14-21 to build integrated understanding across all of Part III.
M1. A government agency measures its success by the number of small businesses it helps create each year (Goodhart target, Ch. 15). To meet its targets, the agency relaxes standards for what counts as a "business," counting sole proprietors with no revenue, hobby operations, and paper entities (cobra effect, Ch. 21). The agency's training programs, designed to help entrepreneurs, are standardized curricula that ignore the local knowledge and specific skills that each entrepreneur needs (legibility, Ch. 16). The programs strip away mentorship and hands-on support (redundancy reduction, Ch. 17) in favor of scalable online courses. When a high-profile failure occurs -- a funded business collapses, taking several others with it -- the failure cascades through the local economy (cascading failure, Ch. 18). The agency responds with more oversight and more standardized requirements (iatrogenesis, Ch. 19), which drive the most capable entrepreneurs to avoid the program entirely. Diagnose this system using all relevant failure modes and propose a comprehensive reform.
M2. A police department implements a crime-reduction program that measures success by reported crime statistics (Goodhart, Ch. 15). Officers, incentivized by the statistics, discourage victims from filing reports and reclassify serious crimes as minor ones (cobra effect, Ch. 21). The department's crime-mapping software, which relies on reported data, overfits to the manipulated statistics, concentrating patrols in areas that appear dangerous on paper rather than areas where crime is actually occurring (overfitting, Ch. 14). Meanwhile, community policing programs -- the informal, relationship-based approaches that actually reduce crime -- are defunded as "inefficient" compared to data-driven patrols (redundancy vs. efficiency, Ch. 17, and legibility, Ch. 16). When a major crime wave hits, the department's data-driven model fails catastrophically because it was trained on manipulated data (cascading failure, Ch. 18). Trace the complete chain of failure modes, identify the root cause, and design a system that avoids all identified failure modes simultaneously.
M3. A university ranks departments by research funding secured (Goodhart target, Ch. 15). Departments respond by hiring researchers who can win large grants rather than those who can teach well or do transformative research on small budgets (cobra effect, Ch. 21). The emphasis on grant funding drives departments to adopt legible, quantitative research methods over qualitative, interpretive, or field-based methods that are harder to fund (legibility, Ch. 16). Departments become intellectually homogeneous, losing the diversity of methods that once produced breakthrough insights (redundancy reduction, Ch. 17). Junior faculty overfit their research to the patterns that have historically won grants, ignoring emerging fields and unconventional approaches (overfitting, Ch. 14). When a paradigm shift occurs in the field, the department is intellectually unprepared, and its grant funding collapses (cascading failure, Ch. 18). The administration responds by imposing more metrics, more reporting requirements, and more oversight (iatrogenesis, Ch. 19). Analyze this system and propose an alternative governance model.
M4. A social media platform introduces an algorithm that promotes content based on engagement metrics (Goodhart, Ch. 15). Content creators discover that outrage, controversy, and misinformation generate more engagement than thoughtful, accurate content (cobra effect, Ch. 21). The algorithm overfits to engagement patterns, creating filter bubbles that reinforce users' existing beliefs (overfitting, Ch. 14). The platform's content moderation team, overwhelmed by volume, relies on automated systems that can only evaluate legible features of content (keywords, flagging patterns) and miss context-dependent harm (legibility, Ch. 16). The platform strips away human moderators as "inefficient" compared to automated systems (redundancy reduction, Ch. 17). When a misinformation campaign goes viral, the automated systems cannot contain it, and the cascade propagates through the platform's interconnected network (cascading failure, Ch. 18). The platform responds by banning more keywords and tightening algorithmic filters, which drives harmful content to less detectable forms and pushes legitimate speech off the platform (iatrogenesis, Ch. 19, feeding back into the cobra effect, Ch. 21). Diagnose and redesign.
M5. Design a "failure mode audit" template that can be applied to any incentive system, organizational policy, or regulatory framework. The template should:
a) Systematically check for each of the eight failure modes from Part III. b) Assess the interactions between failure modes -- which pairs tend to reinforce each other in this particular system? c) Identify the most vulnerable point in the system -- the place where a failure in one mode is most likely to trigger failures in others. d) Propose targeted interventions for the most critical failure modes, and assess whether each proposed intervention is itself vulnerable to any of the eight failure modes. e) Include a "meta-check": is the audit itself vulnerable to Goodhart's Law (will people game the audit), overfitting (is it trained on historical failures that may not predict future ones), or legibility traps (does it only detect legible failure modes)?