Anchoring
Anchoring is the finding that an arbitrary value, once it enters your mind, drags your subsequent numerical judgments toward itself — even when you know it's arbitrary, even when you're warned, even when it has nothing to do with the question. It is one of the most robust effects in the whole troubled field of judgment-and-decision-making, and that robustness is exactly why it deserves a careful, honest reading. Most "cognitive biases" you've heard about have wobbled badly under the replication crisis. Anchoring has wobbled less than almost any of them. That makes it more important to understand precisely, not less — because a real effect, misapplied, does more damage than a fake one.
Let's pull it all the way down.
What it is (and where it came from)
The canonical source is Amos Tversky and Daniel Kahneman's 1974 paper in Science, "Judgment under Uncertainty: Heuristics and Biases." This is the paper that put "heuristics and biases" into the language. In it, anchoring is presented as one of three core heuristics — alongside representativeness and availability — that people use to make estimates under uncertainty. The claim: when forming a numerical judgment, people start from an initial value (the anchor) and adjust from it, but the adjustment is typically insufficient — they stop too soon, landing closer to the anchor than they should.
The demonstration that everyone remembers is the wheel of fortune. Tversky and Kahneman had participants spin a wheel marked from 0 to 100 — and here is the detail that matters — the wheel was rigged to stop on either 10 or 65. After the wheel stopped, participants were asked whether the percentage of African nations in the United Nations was higher or lower than the number the wheel had just produced, and then asked to give their own estimate.
The result is the whole point of the field in miniature. The group that saw 10 gave a median estimate of about 25%. The group that saw 65 gave a median estimate of about 45%. A number generated in front of them by a wheel they watched spin — a number with no possible informational relationship to African UN membership — moved their answers by roughly twenty percentage points. They knew it was a wheel. They watched it spin. It didn't matter.
That is the signature of anchoring, and why it unsettled people: it isn't a case of someone being fooled by a misleading-but-relevant cue. The cue was transparently irrelevant and it worked anyway.
A word on intellectual honesty about origins. Tversky and Kahneman named and framed the effect, but the broader observation that an initial reference point distorts later judgment has older roots — in psychophysics (how a prior stimulus shifts your perception of the next one) and in earlier work on judgment and assimilation effects. What 1974 did was isolate it cleanly, give it a mechanism (anchoring-and-adjustment), and fold it into a research program that would eventually win Kahneman the 2002 Nobel Memorial Prize in Economics. (Tversky had died in 1996; the prize is not awarded posthumously, and Kahneman, who died in 2024, was consistent and public that the work was a partnership.)
It's worth being precise about why this landed as hard as it did, because it tells you what anchoring is really a demonstration of. The dominant model of human reasoning in mid-century economics and much of psychology assumed something like rational calculation: a well-informed person, given a question, retrieves relevant facts and computes an answer, and irrelevant inputs wash out. Anchoring is a direct refutation of that picture at its most uncomfortable point. The irrelevant input doesn't wash out. It contaminates the answer, the contamination is large, and — this is the part that makes it philosophically heavy rather than just a parlor trick — the person can't feel the contamination happening. Introspection reports a clean, reasoned judgment. The judgment was poisoned upstream of awareness. That gap between what reasoning feels like from the inside and what it is mechanically is the whole engine of the heuristics-and-biases program, and anchoring is its cleanest single illustration.
The real science / evidence
Here is where anchoring earns the unusual statement I made at the top — that it's one of the sturdiest effects in a field that has otherwise taken a beating. Let me be specific about what is well-established and what is softer.
The basic effect replicates, broadly and across domains. Anchoring has been shown in estimates of geography, prices, probabilities, dates, the length of the Mississippi River, the height of redwoods, the year of an event, and on and on. It survived the large-scale "Many Labs" replication project (Klein et al., 2014), which re-ran a batch of classic psychology findings across many sites with large samples specifically to see which ones held up. A number of celebrated effects did not hold up in that project. Anchoring did — it was among the findings that replicated reliably. This is the single most important fact to carry: in an era where you should be skeptical of any pop-psych "bias," anchoring is one of the few that has paid its dues.
It works with random anchors. The wheel of fortune is not a fluke. Later studies extended the random-anchor finding in ways that are almost comic. In a frequently cited line of work, researchers had people consider the last two digits of their own Social Security number, then bid on consumer goods (Ariely, Loewenstein & Prelec, 2003, the "coherent arbitrariness" paper). People with higher SSN digits bid more. Again: a number drawn from the person's own ID, with zero relationship to the value of a bottle of wine or a cordless keyboard, shifted willingness to pay. (Note the honest qualifier — this specific "anchoring on your own SSN" paradigm has had a mixed replication record; some attempts reproduce it, while larger replications (e.g. Chapman et al., 2023) found much weaker effects. The general phenomenon of arbitrary anchors moving valuations is solid; this one charismatic demonstration of it is less bulletproof than the headline version. Hold the principle loosely where the single study is shaky.)
It survives warnings, and it survives expertise. This is the finding that should genuinely bother you. Telling people about anchoring before they make a judgment does not reliably immunize them. Offering money for accuracy doesn't reliably eliminate it. And — the demonstration with the most real-world teeth — experts are not immune. A landmark study by Northcraft and Neale (1987) gave real estate agents a genuine house to appraise, complete with a tour and a full information packet. The only thing manipulated was the listing price printed in the packet. The agents' appraisals moved with the listing price — and, tellingly, when asked, most agents denied that the listing price had influenced them. They believed they were valuing the house on the merits. They weren't. The anchor moved them and hid its own fingerprints.
So the evidence base, stated honestly:
- Well-established: arbitrary and irrelevant anchors shift numerical judgments; the effect replicates at scale; warnings and incentives provide weak protection; experts are affected in their own domains.
- Softer / contested: the exact size of the effect in any given setting, certain individual flagship demonstrations (the SSN study), and — as we'll get to — why it happens.
The mechanism
This is the genuinely interesting part, and it's where you need to resist the urge to have a tidy answer, because the field doesn't have one cleanly settled. There are two major accounts, and the honest position is that both appear to be real, operating in different conditions.
Account 1 — Anchoring-and-adjustment (the original). Tversky and Kahneman's own explanation: you take the anchor as a starting point and adjust away from it toward your true estimate, but you stop adjusting as soon as you reach a value that seems plausible rather than continuing to the value that's actually best. Because the plausible region is wide, you stop early, and "early" means "still near the anchor." On this view anchoring is a story about effortful, deliberate adjustment that quits prematurely.
The strongest evidence for genuine adjustment comes from work on self-generated anchors — cases where you produce the anchor yourself from something you know. Epley and Gilovich (in a series of papers in the 2000s) showed that for these self-generated anchors, you really do seem to adjust step-by-step from the anchor, and you adjust too little — and you can be pushed to adjust further by giving incentives or by warning people. (Example: "When did George Washington become president?" Many Americans don't know, but they know it was after 1776, so they anchor on 1776 and adjust up — and stop short. The true answer is 1789.)
Account 2 — Selective accessibility (the challenger). Developed largely by Strack and Mussweiler (1997) and elaborated by Mussweiler through the late 1990s and 2000s. Their argument: anchoring isn't mainly about adjustment at all. When you're asked "is the answer higher or lower than [anchor]?", you test that hypothesis by selectively retrieving information consistent with the anchor being correct. You essentially ask your memory, "could it be around the anchor?" — and your memory dutifully surfaces anchor-consistent evidence. That evidence is now sitting at the front of your mind, accessible, so when you make your own estimate it's built out of anchor-consistent material. No deliberate adjustment required — the anchor biased your retrieval, and your retrieval built your answer.
The decisive evidence for selective accessibility is that it predicts anchoring should work even when there's nothing to "adjust from" in the classic sense — and especially when the anchor is externally provided (the experimenter hands you the number, as in the wheel). For externally provided anchors, the adjustment story struggles and the accessibility story fits well.
The synthesis that most researchers now accept: two routes, sorted roughly by where the anchor comes from. Self-generated anchors → genuine, insufficient adjustment (Epley & Gilovich). Externally provided / experimenter-given anchors → selective accessibility, a priming-like memory effect (Strack & Mussweiler). The original 1974 label, "anchoring-and-adjustment," turned out to name the less common of the two mechanisms — accurate for the cases where you build your own starting point, but not the best description of the wheel-of-fortune demonstration that made it famous. That's an unusually clean example of a founding intuition being partly right and partly superseded, and it's worth sitting with: the most famous demo of "anchoring-and-adjustment" is probably not mostly adjustment.
There's also a deeper, unresolved question lurking underneath both accounts: how much of anchoring is a genuine belief distortion versus a response-scale or numeric-priming effect — i.e., the mere presence of a big number making big numbers feel more "available" as answers. This is live, technical, and not settled. The honest summary: the what of anchoring is rock-solid; the why is two-thirds mapped.
A note on what "insufficient adjustment" really implies, because it's easy to read past. The original framing isn't that people fail to adjust — it's that they adjust and stop too early, at the near edge of a range of answers that feel acceptable rather than at the single best answer. The reason this matters for the mechanism debate is that "stopping at the near edge of plausibility" is an effortful-process story: it predicts that anything which makes you work harder or care more — incentives, accountability, more time, being told to keep going — should push the stopping point further from the anchor. For self-generated anchors, that prediction roughly holds (Epley & Gilovich found you can extend the adjustment). For experimenter-given anchors, it largely doesn't, which is precisely the evidence that those cases aren't running on adjustment at all but on biased retrieval. The two mechanisms aren't just two theories competing for the same data — they make different predictions about what fixes the bias, and the data sorts them by anchor source. That's about as clean as mechanism disputes get in this field, which is to say: clean enough to teach, not clean enough to close.
Sharp examples
Negotiation and the first offer. This is the most practically important arena, and the evidence is among the strongest applied evidence anchoring has. In negotiation studies, the side that makes the first offer tends to pull the final agreement toward that offer — the first number anchors the bargaining range. This is why "let them name a price first" is often bad advice when you have a reasonable sense of value: by speaking first with an aggressive-but-defensible number, you set the gravity well the whole negotiation falls into. (The qualifier: this advantage assumes you have decent information. If you're badly informed, speaking first can anchor you to a number that gives away how little you know, or that's far off market. The first-offer advantage is real but conditional.)
Pricing and "anchor pricing." The entire architecture of retail pricing runs on this. A "$1,200, now $499" tag works because $1,200 is the anchor — your sense of whether $499 is "a lot" is computed relative to it, not relative to the item's actual worth. The decoy on a menu — a wildly expensive bottle of wine nobody buys — exists to anchor the second-most-expensive bottle into looking reasonable. Manufacturer's Suggested Retail Price is, functionally, a legally-sanctioned anchor. "Most people choose the $99/month plan" anchors the comparison. None of this requires the buyer to be fooled about anything factual; it just sets the reference point from which "expensive" and "cheap" are felt.
Judicial sentencing. Here is the most alarming applied result, and it should be cited carefully. Englich, Mussweiler, and Strack (2006) ran experiments with German legal professionals — including experienced judges — in which an irrelevant number was introduced before a sentencing decision. In one striking version, judges read a case, then rolled dice (rigged) before recommending a sentence; those who rolled higher numbers recommended longer sentences. In related versions, a sentencing demand coming from a non-expert (a journalist's phone question, or a computer science student) still moved the professionals' recommendations. Read that again: the length of a prison sentence, in a study with real judges, moved with a number from a pair of dice. The honest framing: these are experimental studies, not audits of live courtrooms, and the dice scenario is deliberately extreme to isolate the mechanism. But the direction of the finding — that even trained, high-stakes, expert judgment is anchorable — converges with the real-estate-agent result and is hard to wave away. It is the strongest argument for building anchoring out of consequential decision processes structurally, rather than trusting individuals to resist it.
Forecasting and estimation at work. Last quarter's number anchors this quarter's target. The initial project estimate anchors the final budget even after scope balloons. The first salary figure mentioned in a hiring conversation anchors the offer. In each case the anchor often entered the room arbitrarily — someone's gut, a round number, a prior unrelated to current reality — and then governed everything downstream.
Medicine and probability estimates. A subtler, higher-stakes example: a clinician told that a test result is "higher than the cutoff of X" can have their probability estimate of disease pulled toward the framing of X, and an initial diagnostic hypothesis anchors the interpretation of every later piece of evidence (closely related to confirmation bias — the anchor sets the hypothesis, and selective accessibility surfaces confirming data). The first plausible diagnosis is sticky in exactly the way the first number is sticky, and for an overlapping reason: once a candidate value or hypothesis is in mind, the mind goes looking for reasons it's right rather than reasons it's wrong. Treat this example as illustrative of the mechanism rather than as a single decisive study — the clinical-anchoring literature is real but more heterogeneous than the negotiation or pricing work.
Round numbers and the power of precision. A small, practical wrinkle worth knowing: there's evidence that precise anchors (an asking price of \$486,500) can pull harder than round ones (\$485,000), apparently because precision signals that the number was carefully computed and therefore deserves to be taken seriously as a hypothesis. This sits well with the selective-accessibility account — the more credible the anchor as a candidate answer, the more anchor-consistent evidence your mind retrieves. The qualifier, as always: this is a real and replicated-enough finding to be worth using, not an iron law; the size of the precision premium varies.
Where it gets distorted / oversold (the honest cut)
Because anchoring is real, it gets stretched to explain things it doesn't, and that's where you get hurt or fooled.
- "Anchoring explains everything about pricing/persuasion." It doesn't. Anchoring is specifically about numerical judgments pulled toward a salient number. A lot of what gets called "anchoring" in marketing-guru content is really framing, priming, or social proof — related cousins, different mechanisms. Calling all of it "anchoring" makes you sloppy about which lever you're actually pulling, and sloppy levers don't reliably move.
- Inflated effect sizes and just-so stories. The lab effect is robust as a direction; the magnitude in any real situation is variable and often smaller than the dramatic demos imply. Someone who tells you "anchor high and you'll get 30% more, guaranteed" is overselling. The effect is reliable enough to use and too variable to promise. Treat it as loading the dice slightly in your favor, not as a control knob.
- The SSN/charismatic-single-study trap. As flagged above, some of the most fun demonstrations (last-two-digits-of-your-SSN drives your wine bid) have shakier replication records than the general principle. This matters because pop accounts cite the coolest version, which is often the least replicated version. The discipline: trust the phenomenon (arbitrary anchors move valuations — replicates broadly), be skeptical of any single jaw-dropping demo of it.
- "Just be aware of it and you'll be fine." This is the most dangerous oversell, because it's the opposite of true. The whole disturbing core of the evidence — warnings don't reliably work, incentives don't reliably work, experts deny influence they demonstrably received — says awareness is weak armor. Defending against anchoring is mostly a structural problem (change the process), not a willpower problem (try harder to be objective). Anyone selling "now that you know about it, you've beaten it" hasn't read the awkward studies.
The debates & frontier
- Mechanism, still open. As covered: anchoring-and-adjustment vs selective accessibility, now mostly reconciled as a two-route model split by anchor source (self-generated vs provided), with a residual unsettled question about how much is belief-distortion vs numeric/response priming. This is genuinely active, not a textbook closed case.
- How much survives debiasing? Most simple debiasing attempts fail, but not all. Generating reasons against the anchor, "consider the opposite," and forcing explicit consideration of a wide range have shown some reduction in some studies. The frontier question is whether any of these are robust enough to deploy in real decisions or whether the only reliable fix is structural (blind the decider to the anchor; have independent estimators commit before seeing each other's numbers; use formulas instead of holistic judgment for high-stakes calls).
- Moderators. Mood, expertise, knowledge of the domain, cognitive load, and individual differences all modulate anchoring, but the literature on moderators is messier and less replicated than the core effect. Be appropriately skeptical of confident claims like "smart people anchor less" — the evidence there is mixed and sometimes contradictory.
- The replication crisis context. This is the meta-point and the reason for the careful tone throughout. Psychology's last fifteen years have been a reckoning: many famous effects (some in this very heuristics-and-biases tradition) shrank or vanished under rigorous replication. The right stance toward any bias claim is now "show me the replication." Anchoring is one of the findings that, having been put to that test, largely passed. That doesn't make it magic or unlimited — it makes it one of the few you're entitled to actually rely on, within its real boundaries.
How to apply it
- Set the anchor on purpose, or someone else will. In any numerical negotiation where you're reasonably informed — price, salary, scope, budget, timeline — get a defensible number into the room first. The first credible number sets the gravity. Silence cedes the anchor to the other side.
- Make your anchor aggressive but justifiable. An anchor that's transparently absurd gets discarded (and damages credibility); one that's ambitious but defensible pulls hard. The pull comes from the number being taken seriously, even briefly, as a hypothesis to test.
- Defend structurally, not by willpower. When you are the one being estimated at, the move isn't "try to ignore their number" — that's been shown not to work. The moves that have a chance: commit your own independent estimate before you hear theirs; consciously generate the strongest case the anchor is wrong; and for anything consequential, decide against an external reference (comparable sales, market data, a formula) rather than against the number they handed you.
- Blind the high-stakes decisions. This is the lesson of the judges and the real-estate agents. If you run a process where one person's number can poison everyone's judgment, fix the process: collect independent estimates before anyone shares; keep the irrelevant number (the wishful target, the prior unrelated figure) out of the room until after the judgment is committed.
- Audit your own anchors for arbitrariness. Half the anchors governing your decisions entered arbitrarily — last quarter's figure, a round number, a competitor's price, the first guess someone blurted. Ask of any number running your decision: where did this actually come from, and does it deserve the authority it has? Usually it doesn't.
- Use it, don't trust it. It loads the dice; it doesn't control the roll. Anchor deliberately, expect a real-but-modest pull, and never build a plan that requires the anchor to land a specific amount.
How It Connects
the Law of Relativity most directly — Relativity says nothing has fixed value, everything is judged relative to what it's set against; anchoring is the cognitive machinery of exactly that, the proof that "expensive," "long," "high," and "fair" are computed against a reference point rather than read off reality. Also: framing and priming (the salient number is a prime that biases retrieval — selective accessibility is anchoring's bridge to priming) · adjustment / insufficient effort (System 2 quitting early — Kahneman's Thinking, Fast and Slow) · availability (its sibling heuristic from the same 1974 paper) · the reference-point logic of prospect theory (Kahneman & Tversky 1979 — gains and losses are felt relative to a reference, the same relativity principle running through value itself). Thinkers: Tversky & Kahneman (origin), Strack & Mussweiler (selective accessibility), Epley & Gilovich (adjustment for self-generated anchors), Northcraft & Neale and Englich (experts are not immune).
The first number wins not because it's right, but because everything after it is measured from where it stands — so either you plant it, or you fight uphill from someone else's.