Anchoring

011 min

Where anchoring shows up

An illustrative case, using a role rather than a person. A team is estimating a migration in sprint planning. Before anyone speaks, the tech lead says "I think this is about three weeks, but tell me what you think." The estimates that follow land between two and four weeks. The work takes nine.

Change one input and run it again. This time each person writes their estimate privately before anyone speaks, and the numbers are revealed together. They range from two weeks to two months. The spread itself is the finding: somebody in the room already knew this was large, and in the first run that person never said so.

The team's knowledge was identical in both runs. What differed was whether a number existed before the estimates were formed.

The moment the effect operates is earlier than people expect. It is not during the argument about the estimate. It is at the instant the first number enters the room, before anyone has committed to anything. By the time the discussion starts, the range has already been set, and the discussion happens inside it.

The same structure appears wherever a number arrives before a judgment: a budget mentioned before scoping, a competitor's price seen before setting yours, a salary band named before an offer, a severity rating suggested before a bug is triaged.

022 min

Where anchoring comes from

Amos Tversky and Daniel Kahneman named the effect in Judgment under Uncertainty: Heuristics and Biases, published in Science in 1974. Anchoring appears there as the third of three heuristics, alongside representativeness and availability.

The demonstration is worth knowing exactly, because its design is what makes the result hard to explain away. Participants were asked to estimate quantities stated as percentages — for example, the percentage of African countries in the United Nations. Before each estimate, a number between 0 and 100 was generated by spinning a wheel of fortune in the participants' presence, so nobody could believe the number carried information. Participants first said whether the true value was higher or lower than that number, then gave their estimate.

The paper adds one sentence that closes off the obvious objection: "Payoffs for accuracy did not reduce the anchoring effect." Paying people to be right did not help.

A second demonstration in the same paper shows the anchor does not have to be handed over at all. Two groups of high school students had five seconds to estimate a product written on a blackboard. One group saw 8×7×6×5×4×3×2×1, the other saw 1×2×3×4×5×6×7×8. The median estimate for the ascending sequence was 512. For the descending sequence it was 2,250. The correct answer is 40,320. Both groups anchored on the result of the first few multiplications they managed, and both adjusted upward far too little.

032 min

Why anchoring happens

The 1974 paper offered a mechanism in its name: people start from the anchor and adjust, and the adjustment stops too early. That account is intuitive and turns out to be incomplete, because it cannot explain anchors that nobody is adjusting from.

The better-supported account is selective accessibility, proposed by Fritz Strack and Thomas Mussweiler in 1997. The claim is that considering whether the answer might be near the anchor makes anchor-consistent information easier to bring to mind. Deciding whether a migration takes more or less than three weeks causes you to retrieve the reasons it might take about three weeks. Those reasons are then sitting in front of you when you produce your estimate, and the estimate is built from them.

The difference between the two accounts matters practically. If anchoring were insufficient adjustment, trying harder would help. If it is selective accessibility, the anchor has already changed which evidence is available, and effort applied afterwards works on a biased sample.

The evidence that separates them

Englich, Mussweiler and Strack tested the accessibility account directly rather than assuming it. In the fourth study of their 2006 paper, they measured which arguments came to mind after exposure to an anchor. The accessibility of incriminating arguments was higher for participants given a high anchor than a low one. The anchor had not merely shifted a final number. It had changed what the participants could readily think of.

This also explains one of the strangest features of the effect. An anchor works when you know it is random, because knowing a number is meaningless does not stop it from making related thoughts available.

041 min

Individual effects

For one person, the distorted output is any number they produce after seeing another number.

The direction of the error is toward the anchor, and its size scales with how uncertain the person is. Someone who knows the answer precisely is unaffected. Someone estimating in an unfamiliar domain is moved a great deal. This is why anchoring does the most damage exactly where estimates matter most: new work, unfamiliar systems, first-time decisions.

The specific cost is a false sense of independent judgement. The estimate feels reasoned, because reasons were genuinely produced. They were just the reasons the anchor made available.

There is a second cost that shows up across many decisions rather than in any one. Because anchored estimates cluster, a group of anchored individuals produces a narrow range and reads that agreement as confirmation. Consensus arrived at from a common anchor looks exactly like consensus arrived at from independent analysis.

051 min

Systemic effects

Across an organisation, anchoring compounds through sequence: whoever produces a number first sets the range that everyone after them works inside.

The compounding mechanism is that anchored outputs become the next round's anchors. Last quarter's estimate anchors this quarter's. A first-pass budget, produced quickly and with the least information anyone will ever have about the project, anchors every refinement that follows. The organisation experiences this as gradual convergence on a number, when it is really repeated adjustment from an initial guess.

The pattern is stable over time in a way that looks like institutional knowledge. Pricing that was set once against a competitor's page, headcount plans that move by increments from a figure nobody can source, sprint velocities that hover near a number chosen in the team's first month — each of these can be an anchor propagating rather than a measurement.

The reason it is hard to detect is the same reason it is hard to correct in one person. There is no record of estimates that were never made. An organisation can see the numbers it produced. It cannot see the range it would have produced had the first number been different.

062 min

Examples

Expert judges, and dice. Birte Englich, Thomas Mussweiler and Fritz Strack published Playing Dice With Criminal Sentences in Personality and Social Psychology Bulletin in 2006. It is the strongest demonstration available that expertise does not protect against the effect.

In their first study, 42 experienced German legal professionals — 23 judges and 19 prosecutors, averaging 129.9 months of courtroom experience, more than ten years — read realistic case material about an alleged rape, designed with experienced trial judges so that it matched real case files. A sentencing figure was introduced through a journalist's question. Participants given the high anchor gave a mean sentence of 33.38 months. Those given the low anchor gave 25.43 months.

Then the researchers removed every remaining excuse. In their third study, 52 junior lawyers determined the sentencing demand themselves, by throwing a pair of dice in front of them. The dice were loaded so that everyone saw one of two demands, 3 or 9 months on probation. Participants who rolled the high demand gave final sentences averaging 7.81 months. Those who rolled the low demand averaged 5.28 months.

These are trained legal professionals, deciding a sentence, having personally generated the number by throwing dice. The paper also compared the size of this effect against a version where the same figures came from a prosecutor, where they would be legitimate input. The sentencing decisions moved by a similar amount either way.

A contrast worth holding next to it. The 1974 wheel-of-fortune result and this one differ in one variable: whether the participants were experts working in their own domain on a decision with real consequences. The effect survived that change. That is the finding, and it is the reason anchoring is treated as a structural feature of judgement rather than a laboratory curiosity.

072 min

How anchoring shows up in product, design, and AI

Deliberately used: pricing pages. A higher-priced plan displayed beside the one a company expects to sell changes what the middle plan looks like. Original-price-struck-through is the same device. Neither is deceptive when the higher price is one somebody genuinely pays. Both become deception when the anchor is fictional, which is why advertised reference prices are regulated in many places.

Deliberately used: negotiation. The first number named sets the range. This is the most practically valuable application of the concept, and it applies to salary, contracts and scoping conversations equally.

Accidentally suffered: estimation. Covered in the opening scenario, and worth restating as a rule: any number spoken before estimates are collected has already changed them.

Accidentally suffered: metrics and dashboards. A default date range, a y-axis that does not start at zero, a target line drawn on a chart — each supplies an anchor before the viewer forms a judgement. A graph with a goal line at 95 percent produces different reactions to the same data than one without it.

In model evaluation. Two places to watch. When humans rate model outputs, showing a reference answer or a previous score anchors the rating, which is why blind, independently-ordered evaluation matters. And when a model is asked to score or estimate, numbers earlier in its context influence the output in a way that is at least superficially similar; the mechanism is not established to be the same, so treat that as a reason to randomise order in evaluation harnesses rather than as a claim about how models think.

082 min

How to guard against anchoring

Awareness is close to useless here, which the evidence states plainly: the judges in the 2006 study knew the sentencing demand was produced by dice they had thrown themselves. What works is changing the procedure so that the anchor arrives later or not at all.

Collect estimates before discussion. Written, private, simultaneous, then revealed together. This is the single highest-value change available, it costs about ninety seconds, and it converts a sequence into a set of independent judgements. The spread you see is information that the anchored version destroys.

Consider the opposite, concretely. Because the mechanism is selective accessibility, the counter is to deliberately retrieve evidence the anchor suppressed. Before settling, write two reasons the answer could be much higher and two it could be much lower. Generating the evidence is the operative step, not merely intending to be open-minded.

Use reference classes rather than intuitions. Estimate from what comparable past work actually took. A distribution of real outcomes is an anchor too, and it is one with information in it.

Control who speaks first and in what order. In a review, ask the most junior person before the most senior. In a planning meeting, ask for the range before the point estimate.

Avoid re-anchoring on your own last answer. When revising an estimate, rebuild it rather than adjusting the previous number. Adjustment from your own prior figure is the same mechanism applied to yourself.

092 min

Common misunderstandings

"You can avoid it by ignoring the number." The 2006 dice result exists specifically to rule this out. Participants generated the anchor themselves at random and were still moved by it. The number does not need to be believed to have its effect.

"It only works on people who do not know the subject." Uncertainty increases the size of the effect, but expertise does not remove it. The German legal professionals had more than ten years of courtroom experience each.

"Anchoring is just priming." They are related and not identical, and the difference is worth keeping. Priming is a broad claim about one stimulus making related concepts more available. Anchoring is a specific, numerically measurable shift in quantitative judgement toward a presented value. This matters for how much confidence to place in each: much of the social priming literature has had serious replication trouble, while anchoring is among the effects that has come through replication work in good shape.

"A high anchor always helps the person who sets it." An implausible anchor can be rejected outright and can cost credibility. The effect is strongest for anchors inside the range of things a person might believe.

Where the anchor is doing useful work. An anchor drawn from real data is exactly what a good estimate should start from. Reference-class forecasting is deliberate anchoring on an informative number. The problem is never that a starting point exists; it is that arbitrary starting points work just as well as informative ones.

101 min

Anchoring vs. nearby concepts

The fast test in a live discussion: was a number stated before the judgement was made? If yes, anchoring is in play whether or not anyone thinks the number was meaningful.

112 min

Where the evidence is contested

Anchoring occupies an unusual position among the judgement biases: the effect itself has held up while the original explanation for it has not.

On the effect, the Many Labs replication project led by Richard Klein and colleagues, published in Social Psychology in 2014, tested thirteen classic and contemporary effects across 36 samples and 6,344 participants. Ten replicated consistently, two failed outright, and the anchoring effects were among those that replicated. This is a meaningful result given how much of the surrounding literature from the same era did not survive.

On the mechanism, the disagreement is real. The 1974 anchoring-and-adjustment account holds that people start at the anchor and stop adjusting too soon. Strack and Mussweiler's 1997 selective accessibility account holds that the anchor changes which information is retrieved, and that no adjustment process needs to be running at all. There is also a live question about whether anchors that are merely incidental, rather than explicitly compared against the target, produce the same effect and by the same route.

Stated fairly, the strongest version of the case against the original account is that it predicts effortful correction should help, and it generally does not. The strongest case for keeping some role for adjustment is that people do appear to adjust when the anchor is self-generated, as in the multiplication example from the original paper.

Where this leaves a practitioner is straightforward and mostly unaffected by the outcome. The effect is real enough to design against. The competing mechanisms recommend the same countermeasures, because both are defeated by the anchor never arriving first. That is the reason the practical advice in this entry is about sequence and procedure rather than about thinking harder.

?8 questions

Questions people ask

▶3 videos

Watch

• 

  Understanding Anchoring

• 

  CRITICAL THINKING - Cognitive Biases: Anchoring [HD]

  Wireless Philosophy

• 

  What is Anchoring? Learn from Haseeb Taj | Construction Explained | DAO PropTech

  DAO PropTech

§5 sources

Sources