Psychology and biases
Anchoring
Any number you see before estimating pulls your estimate towards it, even when you know the number is random.
Amos Tversky and Daniel Kahneman, Judgment under Uncertainty: Heuristics and Biases, Science (1974).
You are asked to guess a price, a date or a percentage, and shortly before, a number crossed your path: a list price, an opening offer, a figure in a headline. Your guess lands nearer that number than it would have, and it feels like your own judgement. Tversky and Kahneman described anchoring in 1974 as starting from an initial value and adjusting towards the answer, with the adjustment stopping too early. The anchor can be supplied by someone else, as in a negotiation, or by your own partial calculation. High school students asked to estimate 8 x 7 x 6 x 5 x 4 x 3 x 2 x 1 within five seconds gave a median of 2,250; those shown the same product in ascending order gave 512. The correct answer is 40,320.
It works because adjustment takes effort and stops as soon as the answer feels plausible, and because the anchor makes anchor-consistent facts easier to retrieve. Paying subjects for accuracy did not remove the effect in the original study. In practice: the first number in a salary negotiation shapes the final one; the 'was $200' tag makes $120 feel cheap; the original estimate on a ticket shapes every re-estimate. A practical defence is to produce your own estimate, from base rates or a quick calculation, before looking at anyone else's number, and to treat an extreme opening offer as information about the other side's strategy rather than about the value.
Unlike several textbook effects, anchoring has survived large replication. Klein and colleagues' Many Labs project in 2014 re-ran 13 classic findings across 36 samples and 6,344 participants. The four anchoring questions replicated in every one of the 36 samples, with effect sizes larger than the original estimate: Cohen's d of roughly 1.2 to 2.4 against 0.93 in the original study. What is still argued is the mechanism, insufficient adjustment against selective accessibility, and how much of the effect survives when the estimator has real expertise. The practical point stands either way: an anchor you cannot avoid seeing is one you must consciously counter.
Example
In the 1974 Science paper, Tversky and Kahneman had subjects estimate quantities such as the percentage of African countries in the United Nations. Before each estimate a wheel of fortune was spun in front of them to produce a number between 0 and 100, and they were asked whether the true value was higher or lower than that number, then to estimate it. Groups whose wheel landed on 10 gave a median estimate of 25 percent; groups whose wheel landed on 65 gave a median of 45 percent. Paying subjects for accuracy did not reduce the gap.
Where it applies
Write your own number down before you open the vendor's quote, the previous sprint's estimate or the other party's offer. When a language model is asked to rate or score something, a number present in the prompt or in a few-shot example pulls the score the same way, so keep stray numbers out of evaluation prompts.
Think about it
Which number did you see this week before you made an estimate, and what would you have guessed without it?
Applied it to something real? Log it.
Test yourselfA wheel of fortune lands on 65 before someone guesses a percentage. What happens to the guess, and why?
It rises towards 65. People start from the visible number and adjust too little, even when they know it is random and are paid for accuracy.
Related models
- First principlesBreak a problem down to what you know is true and reason up from there, instead of copying what others have done.
- The availability heuristicWe judge how likely something is by how easily examples come to mind, so vivid and recent things look more common than they are.
- The planning fallacyPeople predict their own projects will finish sooner and cost less than almost every comparable past project did.
- The map is not the territoryEvery model, metric or plan is a simplification, and the parts it leaves out can matter more than the parts it shows.