Why the Anchor Won't Let Go

You already know anchoring happens — throw a number at someone and their estimate drifts toward it. The first lesson showed you the dent. This lesson opens the hood and asks the harder question: why? Why does the dent survive when the number is obviously plastic? Why can you know an anchor was pulled from a hat and still lean toward it?

The answer is genuinely surprising: there isn’t one mechanism. There are (at least) two, and they run in different situations. Understanding which engine is turning is the difference between a defense that works and a pep talk that doesn’t.

Two engines, not one

For years anchoring was explained with a single tidy story — you adjust from a starting value and stop short. It’s a good story. It’s just incomplete. In the late 1990s two research teams pulled the phenomenon apart and found that “anchoring” is really a family of effects with different causes depending on where the anchor comes from.

• When you generate the starting number yourself, one engine runs: slow, effortful anchoring-and-adjustment.
• When someone hands you a number to consider, a different engine runs: automatic selective accessibility.

The reason this matters is practical. The two engines respond to completely different interventions. One yields — a little — to effort, warnings, and motivation. The other shrugs those off. Confuse them and you’ll aim your willpower at the wrong target.

Engine 1: Anchor-and-adjust (and quit too early)

Anchoring-and-adjustment is the mechanism Nick Epley and Thomas Gilovich pinned down (in work around 2001 and 2006) for self-generated anchors — cases where you know a relevant value and consciously move away from it.

The analogy. Picture inching along a rope in thick fog. You can’t see the answer; you can only feel whether you’ve reached “plausible” yet. So you shuffle outward from your known starting point and let go the instant the answer feels possible — not correct, just no longer obviously wrong. You stop at the near edge of the plausible range instead of pressing on to its center.

A worked example. What’s the boiling point of water at the summit of Everest? Most people start from a value they do know — water boils at 100°C at sea level — and adjust downward because they recall that thinner air lowers the boiling point. Adjustment is effortful, so they stop as soon as they hit a number that seems believable, maybe 90°C. The real answer is closer to 70°C. They adjusted in the right direction and quit too soon.

Another classic: when was George Washington elected president? Few people have the date memorized, but almost everyone knows the Declaration was signed in 1776. So they start at 1776 and inch forward — “a bit after, but not much” — landing around 1780. The actual year is 1789. Same fingerprint: correct direction, premature stop.

The tell-tale sign: this engine is partly fixable

Because anchoring-and-adjustment runs on effort, anything that buys more effort buys less bias. Epley and Gilovich showed the self-generated flavor shrinks when people are warned, paid for accuracy, highly motivated, or simply alert rather than tired or tipsy. Give people a reason to keep inching down the rope and they travel further before letting go.

The pitfall. This is exactly the finding that fools people into thinking all anchoring is willpower-solvable. It isn’t. This partial fix only applies when the anchor was self-generated. Aim the same effort at an externally supplied anchor and it barely moves — because a different engine is running.

Engine 2: Selective accessibility (the anchor rigs the search)

When someone hands you a number — a list price, an opening offer, a “higher or lower than X?” question — Thomas Mussweiler and Fritz Strack (1999) showed you’re not sliding a dial at all. You’re running a biased memory search.

The mechanism. To answer “is the true value higher or lower than the anchor?”, your mind runs a positive test (also called confirmatory testing or hypothesis-consistent testing): it assumes the anchor might be the answer and retrieves knowledge that fits that assumption. Evaluating “could a used car be worth 40,000?” pulls up its leather seats, low mileage, clean history. That anchor-consistent evidence becomes highly accessible — mentally “warm” and easy to recall — and when you then produce your own estimate, the warm evidence dominates. The anchor doesn’t push your number; it rigs which facts are lying around when you reach for an answer.

The analogy. Someone asks “is this restaurant more or less than 5 stars amazing?” To check, you dredge up every good thing about it. Then they ask what you’d rate it — and your head is now full of the good stuff you just retrieved. You were never told to inflate the score; the question quietly stocked your memory with reasons to.

Why this reframes everything. If the bias lives in what came to mind, then telling someone to “just adjust more” is useless — there’s no conscious sliding step to lengthen. The distortion already happened upstream, during retrieval, before you felt like you were estimating at all. This is why the accessibility engine is the stubborn one.

Absurd anchors still pull — belief is not required

Here’s the finding that makes the accessibility story impossible to ignore. If anchoring were just “you partly believed the number,” then a ridiculous number should do nothing. It doesn’t play out that way.

Strack and Mussweiler (1997) asked people whether Gandhi died before or after age 9, or before or after age 140 — figures everyone knows are nonsense. Both anchors still shifted the subsequent age estimates, in opposite directions. The absurd-low anchor dragged guesses down; the absurd-high anchor dragged them up. Nobody believed Gandhi died at 9 or 140. The pull happened anyway.

Chapman and Johnson pushed further: even numbers generated by an obviously random process — a spin of a wheel, the last digits of a Social Security number — moved people’s valuations and estimates. You can watch the number get invented at random and still be nudged by it.

The lesson. An anchor doesn’t need to be believable to work, because the mechanism never required belief. Testing the comparison (“older or younger than 140?”) still surfaces age-related knowledge and still leaves it accessible. The absurdity gets rejected by your conscious judgment while the retrieval it triggered quietly biases the next step.

Why anchoring is so brutally robust

Put the two engines together and you can predict exactly how stubborn anchoring will be — and why the obvious fixes underperform.

The right-hand column is the grim one. Timothy Wilson and colleagues (1996) showed that forewarning — literally telling people “there’s an anchoring effect, don’t fall for it” — did little against externally supplied anchors. Financial incentives for accuracy don’t reliably rescue people either. Expertise doesn’t immunize you: real-estate agents anchored on listing prices, judges on random sentencing numbers, all while insisting the anchor hadn’t affected them. Even mood barely dents it.

The reason is now obvious given the mechanisms. Warnings, cash, and expertise all try to buy more effort. That works — a little — on the engine that runs on effort (self-generated adjustment). It fails on the engine where the damage is already done during automatic retrieval, before conscious effort has anything to grab.

You now know why the anchor won’t let go: partly because you quit adjusting too early, and mostly because the number quietly rigged which facts came to mind. Willpower reaches the first problem and misses the second. In the next lesson, we stop trying to try harder and start changing the process itself.