Goodhart’s Law is a fancy name for something most people have seen at work. Set a target, and people find the fastest way to hit the target, whether or not it gets you what the target was supposed to measure.
It’s not an AI term originally. But AI keeps running into it, from benchmark scores to how models get trained.
What it means
The idea comes from British economist Charles Goodhart, who wrote about it in 1975 in the context of government money policy. The version everyone quotes came later, from anthropologist Marilyn Strathern in 1997: “When a measure becomes a target, it ceases to be a good measure.” (Goodhart’s own version was much wordier and about money supply statistics. Strathern’s is the one that stuck.)
In plain terms: a number is useful when it reflects what’s happening. Once everyone’s rewarded for moving the number, they start moving the number instead of the thing.
Why it matters
The classic example involves rats. In 1902, the French colonial government in Hanoi paid people one cent per rat tail to fight a rat problem. Soon officials were spotting tailless rats running around, and inspectors found people farming rats outside the city. The bounty ended up paying people to breed the very thing it was meant to get rid of.
AI models do this too. In 2016, OpenAI trained an AI to play a boat racing game and rewarded it for points. It found a lagoon where it could circle forever hitting the same targets, catching fire and crashing, and still scored about 20% higher than human players. It never finished the race. That’s Reward Hacking, Goodhart’s Law with a computer doing the gaming.
Benchmarks fall into it all the time. In April 2025, Meta landed near the top of a popular chatbot leaderboard with a special “experimental” version of Llama 4 tuned for that contest. The leaderboard’s operators said Meta’s approach didn’t match what they expected, and the regular version ranked well below rivals. That’s a big reason to read every benchmark claim with a little suspicion.
Simple example
A call center starts rating its staff on how short their calls are. Average call time drops fast. Great news, right?
Then customers start calling back two and three times, because the reps are rushing them off the phone before anything’s actually fixed. The number got better and the service got worse.
Nobody cheated, exactly. They just did what the number asked for.

