What is anchoring bias?

Anchoring bias is the tendency to let the first number that appears in a discussion drive every estimate made after it, even when that number is arbitrary and the person deciding knows perfectly well that it is. Amos Tversky and Daniel Kahneman described it in Science in 1974, in the paper "Judgment under Uncertainty: Heuristics and Biases" (185, 4157, pp. 1124-1131). Participants watched a wheel of fortune stop on either 10 or 65, said whether the percentage of African countries in the United Nations was higher or lower than that number, and only then gave their own estimate. The median estimates were 25 for the group that started from 10 and 45 for the group that started from 65. The mechanism has a technical name, anchoring and insufficient adjustment, which in plain words means one thing: you start from the first figure available and then you move too little.

See also: Planning fallacy · The systematic tendency to estimate a project's time and cost on the best-case scenario, ignoring how similar projects actually went.

Why the anchor holds even when it is absurd

The unpleasant part is that you do not have to believe the anchor for it to work. In the 1974 experiment the number came out of a wheel, in plain sight, and it still moved the median answer by twenty points. What came after needs splitting, though: versions using openly random anchors have struggled to replicate, while the form in which the anchor is simply put to you is among the few results from that era to have survived the replication crisis. In the Many Labs project (Klein and colleagues, Social Psychology 45, 3, pp. 142-152, 2014) it is that form, in Jacowitz and Kahneman's 1995 version, that replicates across 36 independent samples, while other classics on the same list did not hold. Two things are worth keeping apart. The first is pure homonymy: in an AI context one speaks of anchoring a model's answer to verifiable sources, but that is grounding and has nothing to do with this bias. The second is a family resemblance: anchoring is the generic mechanism at work on most numeric estimates, while the planning fallacy is the specific optimism about project time and cost.

An enterprise example

A vendor sends the quote to rebuild the order portal and writes 45,000 euros on it. From that moment the negotiation no longer runs between 25,000 and 70,000: it runs around 45,000, and even the competitor's later bid is read as expensive or cheap relative to that figure rather than to the work involved. The same happens in a planning meeting: the first person to say "three weeks for me" has already fixed the field, and everyone else's estimates arrange themselves around three weeks. And it happens on a price list, where the most expensive plan on the page makes the one next to it look reasonable, which is part of the reason psychological pricing works.

Why it matters for decision makers

The countermeasures that hold up are procedural and cost very little. Estimate blind: everyone writes their own figure down before anyone says one out loud, and only then are they compared, so the meeting starts from a distribution instead of an anchor. In supply negotiations, ask for the quote before stating the budget, because if the budget comes out first the quote will settle on top of it. And replace the arbitrary anchor with a real one, meaning start from the data of comparable projects already finished, which is reference class forecasting. None of the three asks people to be more rational, a request that in testing shrinks the anchor slightly and never removes it: they change the order in which numbers enter the room, which is the one lever you can actually pull. One uncomfortable fact is left to look at squarely: anchoring works against you when you receive the first number, and for you when you are the one giving it.