Incentives, Feedback & the Cobra Effect

So far you’ve learned the move: name the obvious result, then ask “and then what?” and trace the chain. But a move tells you what to do, not where to point it. If you have to ask “and then what?” about every decision with equal suspicion, you’ll exhaust yourself before lunch. What you actually want is a map of where the surprises come from — the underlying machinery that keeps turning sensible first-order plans into second-order disasters.

Good news: there basically are only three engines. Almost every backfire you’ll ever meet is one of them, or a couple working together. Learn the three, and “and then what?” stops being a vague anxiety and becomes a targeted search: you’ll know exactly which rocks to look under.

Engine 1 — Incentives & gaming (the Cobra Effect)

Imagine you run a vending machine that pays out a dollar every time someone drops in a dead cockroach. You wanted a cleaner building. What you’ll actually get is a roach farm in the basement, because you didn’t pay people for fewer roaches — you paid them for roach corpses, and corpses are something an enterprising person can mass-produce. The incentive points at the proxy, not the goal, and people follow incentives the way water follows gravity.

This is the Cobra Effect, and it’s named after a story that actually happened. Under British colonial rule in Delhi, officials were alarmed by the number of venomous cobras in the city. Their solution was elegantly simple: put a bounty on dead cobras. Bring in a dead snake, collect a reward. First-order, it worked beautifully — people went out, killed cobras, and the snake count started dropping. Mission accomplished.

Then the second-order machinery woke up. Locals noticed that a dead cobra was now worth money, and that the cheapest way to produce dead cobras on demand wasn’t to go hunting dangerous wild snakes — it was to breed them at home. Cobra farms sprang up. People raised snakes for the sole purpose of killing them and collecting the bounty. When the government eventually figured out what was going on, they did the sensible thing and cancelled the scheme. Now those breeders were holding warehouses full of suddenly worthless cobras — so they released them. The end result: more cobras in Delhi than before the program started. The policy didn’t just fail; it ran in reverse.

If you think that’s a one-off colonial blunder, history obligingly ran the experiment again. In French-colonial Hanoi, a rat infestation in the sewers led officials to pay a bounty per rat — and, to save people the trouble of hauling in whole carcasses, they accepted a rat tail as proof. You can guess what happened. Rat-catchers started catching rats, snipping off the tails, and releasing the now-tailless rats alive — because a living rat breeds more rats, and more rats means more tails to collect. Inspectors eventually found rat farms on the edge of the city, run by people growing rats for their tails. Same engine, different rodent.

The technical name for what these bounties created is a perverse incentive: a rule whose reward structure makes people produce the very thing the rule was meant to prevent. The home discipline here is economics and policy design, but the principle is universal. The diagnostic is always the same — the reward was attached to a proxy (dead snakes, rat tails) instead of the actual goal (fewer live snakes, fewer live rats), and a proxy can be manufactured.

When to use it

Reach for the Cobra Effect lens any time a plan offers people a reward, target, or punishment tied to a measurable proxy. Ask: “If I wanted to collect this reward in the laziest, most cynical way possible — without actually achieving the goal — how would I do it?” Whatever you come up with, somebody out there will too.

Goodhart’s Law

There’s a cleaner, more general way to state the Cobra Effect, and it comes from a British economist named Charles Goodhart, who was writing about monetary policy in 1975. Boiled down to its bumper-sticker form by later economists, Goodhart’s Law reads:

“When a measure becomes a target, it ceases to be a good measure.”

The intuition: a metric is useful precisely because, in its natural state, it correlates with the thing you care about. Test scores correlate with learning. New-account counts correlate with happy customers. But the moment you announce “hit this number or else,” you give people a reason to optimise the number directly — and the cheapest way to move a number is almost never the thing that made the number meaningful in the first place. The correlation snaps, and you’re left optimising a proxy that no longer points at the goal.

The crispest real-world example is Wells Fargo. The bank set aggressive cross-selling targets — employees had to open a certain number of new accounts and products per customer, on the theory that a customer with more products is a more engaged, more loyal customer. The metric (accounts per customer) was a measure of a good thing. Once it became a hard target with jobs on the line, employees did the rational, desperate thing: they opened accounts customers never asked for. By the time regulators were done counting, employees had created roughly 3.5 million unauthorised accounts, the bank ate a $185 million fine plus billions more in settlements, and the CEO was gone. The metric had perfectly correlated with customer engagement right up until it became a target, at which point it measured nothing but fear.

The pattern repeats everywhere a number becomes a quota:

Notice the through-line: Goodhart’s Law is the Cobra Effect’s grown-up sibling. The cobra bounty failed because the reward (money for dead snakes) was attached to a proxy. Goodhart generalises it to any measurement: the act of turning a measure into a target is what corrupts it, because it hands everyone an incentive to optimise the proxy rather than the goal.

When to use it

Pull Goodhart’s Law out the instant anyone proposes managing by a single number — KPIs, OKRs, test scores, leaderboard rankings. It doesn’t mean “never measure.” It means: assume that any metric, once it carries consequences, will drift away from the goal it was chosen to track, so build in slack, multiple measures, and human judgement rather than worshipping one figure.

Engine 2 — Feedback loops

The second engine isn’t about people cheating — it’s about amplification. Picture a microphone held too close to its own speaker. A tiny sound goes into the mic, comes out the speaker louder, goes back into the mic louder still, and within half a second you’ve got an ear-splitting screech from a whisper. Nothing changed the rules; the output fed back into the input, and a small disturbance exploded. Now picture a thermostat instead: room gets too warm, heater clicks off, room cools, heater clicks back on — the output also feeds back into the input, but here it cancels the disturbance and holds things steady.

Those are the two flavours of feedback loop, and the distinction is the whole engine. A reinforcing (amplifying) loop feeds an effect back in a way that makes it bigger — the screech, the snowball, the viral post. A balancing (stabilizing) loop feeds an effect back in a way that shrinks it toward an equilibrium — the thermostat, the body sweating to cool down, supply rising to meet a price spike. This is the core idea of systems thinking, the discipline that studies how parts of a system loop back on each other; feedback loops are deep enough that they get their own dedicated course later in this track. For now, the second-order point is just this: loops compound, so a change you’d shrug off as tiny in a static world can run away with the whole system.

The textbook example is a bank run. Suppose a rumour spreads that a bank is shaky. First-order, a few nervous customers withdraw their savings. But banks lend most deposits out — they don’t keep all the cash on hand — so visible withdrawals make the bank look shakier, which scares more depositors, who withdraw, which makes it look shakier still. Each withdrawal increases the fear that causes the next withdrawal: a textbook reinforcing loop. A bank that was perfectly solvent on Monday can be genuinely insolvent by Friday, killed not by any real problem but by the loop the rumour started. The same machinery drives panic-buying (empty shelves → “I’d better stock up” → emptier shelves) and viral growth (each new user invites two more, who each invite two more).

Engine 3 — Substitution & adaptation

The third engine is the sneakiest, because nobody’s cheating and nothing’s amplifying — people are just quietly rerouting around your rule. Picture a river and a dam. You build the dam to stop the water; the water doesn’t argue, doesn’t game the system, doesn’t loop — it simply finds the lowest path around the edge and keeps flowing. People do the same with rules. Block one path and behaviour flows to the next-easiest one, often undoing much of what you intended.

Formally, substitution is what happens when a rule makes one option more costly, so people switch to an untaxed, unbanned, or unmeasured substitute that serves the same need. The home discipline is economics, and the everyday case is taxes. Slap a steep tax on cigarettes in one state, and first-order, cigarette sales there drop — but a chunk of that “drop” is just smokers driving across the state line, buying online, or switching to a cheaper untaxed product. The behaviour didn’t vanish; it moved. Tax sugary soda, and some people switch to equally sugary juice. The first-order intent (less of X) collides with the second-order workaround (more of substitute-for-X).

A subtler cousin is risk compensation, captured by the Peltzman effect. Economist Sam Peltzman argued that when cars are made safer — seatbelts, airbags, anti-lock brakes — some drivers respond by driving a little less carefully, because the added safety makes risky driving feel cheaper. The safety improvement is real, but part of it gets “spent” on faster, more aggressive driving, partly offsetting the gain. Same mechanism shows up with bike helmets, ski safety gear, even financial bailouts: make an activity feel safer and people will rationally take on more of the risk you just protected them from. Nobody’s being malicious — they’re adapting to the new conditions, which is exactly the second-order reaction first-order plans forget to model.

When to use it

Use the substitution lens whenever you restrict, tax, ban, or make-safer something. Ask: “If this path gets more expensive, what’s the next-cheapest path that serves the same need — and does it route around my goal?” Water finds the gap; assume behaviour will too.

Putting the engines together

Here’s the payoff. You no longer have to vaguely worry that something might go wrong. When you predict an effect, you run it past three specific questions, and most backfires light up at least one:

Plenty of real disasters run two engines at once. A poorly designed sales target (gamed incentive) can spread through a team as a culture of cheating (reinforcing loop). A tax (substitution) can trigger a smuggling industry that grows the more it’s enforced (loop). You don’t need to label them perfectly — you need the checklist to make you look in the three places where the trouble actually hides.

The limit: not every rule is doomed

Time for the honest caveat, before you cancel every incentive program you’ve ever designed. These three engines are a flashlight, not a prophecy. They tell you where to look for backfires — they do not tell you that every plan is secretly cursed. Plenty of rules work exactly as intended: most people do pay their taxes, vaccines do reduce disease, well-designed bonuses do motivate good work. If you treat “second-order thinking” as “assume everything backfires,” you’ll become the paralysed cynic who never ships anything, which is its own kind of failure.

The skill isn’t pessimism — it’s targeted suspicion. The engines hand you three high-yield questions to ask of any plan; most plans pass at least the first time you ask. And when one does light up, you usually don’t have to scrap the plan — you fix the design: pay for the goal instead of the proxy, cap the loop, close the substitute. The cobra bounty wasn’t doomed by physics; it was doomed by paying for dead snakes instead of for verified-reduced live snake populations, which is a fixable design flaw.

When to use it

Run the three-engine checklist whenever you’re designing a rule, metric, incentive, tax, ban, or safety feature — anything that tries to change what people do. Skip the doom-spiral; just check the three spots, fix what lights up, and ship the rest.

Recap

Where this goes next

You now have the engine map — the three sources of nearly every second-order surprise. In Lesson 04, “Five Cases,” we put the map to work on five real-world backfires from policy, business, and technology, naming the engine (or engines) driving each one and showing how a second-order thinker would have caught it in advance. Theory, meet the wild.