The case file described a woman, called Lena M. in the materials, caught stealing from a supermarket for the 12th time. The young lawyers reading it at a postgraduate program in Speyer, Germany, were recently out of law school and in their first jobs as judges. Their task was to sentence her.
They were told the study was about the best order of questions in a courtroom. Then each one picked up a pair of dice, rolled, added the two faces, and wrote the total in as the prosecutor’s demand, in months on probation.
The dice were loaded. One set always landed on 1 and 2. The other always landed on 3 and 6.
Those who rolled a 9 gave Lena M., on average, about two and a half months more than those who rolled a 3. For the identical theft, sentences in the room ran from 1 month to 12.
Where the adjustment stops
The first number you hear can pull your estimate toward it, even one you know is random.
The psychologists Birte Englich, Thomas Mussweiler and Fritz Strack, who published the dice study in 2006, built it to settle a doubt. In an earlier study in the same paper, experienced judges had been pulled by a demand printed with a warning that it was random. Maybe they hadn’t believed the warning. So the researchers put the randomness in the judges’ own hands. Another study looked for the mechanism: after a high roll, participants sorted incriminating arguments about the case faster than after a low one, as if the number had brought the case against the defendant to mind.
The name goes back to 1974. In a paper in Science, Amos Tversky and Daniel Kahneman described people estimating quantities such as the percentage of African countries in the United Nations. First, a number between 0 and 100 was set by spinning a wheel of fortune in front of them, and they said whether the answer was higher or lower. Then they gave their estimate. The median estimates were 25 for the group that got 10, and 45 for the group that got 65. Paying for accuracy didn’t reduce the effect.
Tversky and Kahneman’s explanation was that people start from the number they’re given and adjust, and “adjustments are typically insufficient.” They called it anchoring. A boat on an anchor can drift, but only as far as the chain.
How well it holds up
In 2014 the Many Labs project reran 13 classic and newer psychology effects across 36 samples. Two priming effects did not replicate at all. Anchoring on factual questions, like the height of Mount Everest after a high or low figure, did, and the original studies had underestimated its size. That was anchoring on estimates of quantities, not a rerun of the judges’ dice experiment.
Other forms are shakier. Studies of anchoring on what people will pay for products have given inconsistent results. A 2019 analysis that pooled them found the effect in every study but one, though smaller than in the best-known original. A 2021 meta-analysis of 29 legal anchoring studies found the effect again, with some signs that published studies overstate it. Professionals in those studies were anchored a little less than laypeople, by a margin too small to be statistically significant.
The first number in the meeting
Imagine your team is sizing a new feature. Before anyone has opened the code, the director says, “This should be about two weeks, right?” Then the estimates go around the table. Each one is a little higher than two weeks, and none is double.
The director’s number came from a hunch. It still became the place where everyone else started counting. Nothing in the room stops it. Prior experience may not either: in the judges’ studies, specialists in criminal law were pulled by the number about as much as lawyers from other fields.
A number you write down before you’ve heard anyone else’s at least protects your first estimate. If each person writes privately and the slips are read out together, the director’s two weeks starts as one number among several. The discussion afterwards can still drift toward it, so ask what each estimate rests on, and let evidence, not the first figure spoken, move the numbers.
When you hand the model a number
The same pull has been measured in language models. In a 2024 study in the Journal of Behavioral and Experimental Finance, Jeremy Nguyen asked GPT-4, Claude 2, Gemini Pro and GPT-3.5 for forecasts after mentioning a high or a low value. According to the paper’s abstract, which is what this page relies on, all four models’ forecasts moved toward the number. Two fixes, asking for step-by-step reasoning and an “ignore previous” instruction, had what the study called “limited and varying degrees of effectiveness.” Those are older models; newer ones may behave differently, and you can check.
Ask for the same forecast in fresh chats, a few times for each version: no figure, a high one mentioned, a low one mentioned, with everything else in the prompt the same. If each version’s answers cluster toward its figure, the model may be swayed by it, and that’s worth checking with more runs; a single run of three is only a hint. Keep irrelevant figures apart from relevant facts, too. A model that updates its forecast on last quarter’s real sales isn’t anchored just for doing so, though it can still lean too hard on one number.
It runs the other way too. Write your own estimate down before you read the model’s, the way you’d want the judges to decide before anyone rolls.