Anchoring Bias
People and AI systems both let a final judgment drift toward whatever number, date, or price entered the conversation first, even when that number is irrelevant, random, or came from someone with an interest in moving you. The fix is not to distrust every number. It is to build your own independent estimate before you let someone else's number into the room.
Anchoring bias is the pull toward a first number, even an irrelevant or random one, when forming a later estimate or judgment.
What is the first number I heard, and would I land somewhere different without it?
Salary and price negotiations, real estate offers, project estimates, performance ratings, and reading AI-generated figures that state a number before showing the reasoning behind it.
Fast overview
Psychologists Amos Tversky and Daniel Kahneman demonstrated the effect in a 1974 experiment. They spun a wheel rigged to stop on either 10 or 65, then asked participants whether the percentage of African countries in the United Nations was higher or lower than that number, and to give their own estimate. People who saw 65 guessed a higher percentage on average than people who saw 10, even though everyone knew the wheel was random and had nothing to do with the real answer.
Anchoring shows up in at least three places: the number someone states before you form your own view, a random or irrelevant figure that happens to be nearby, and a number you generated yourself earlier that keeps pulling later estimates toward it. All three work the same way: the mind treats a starting point as informative and adjusts too little away from it.
Mental Model Compass
You are about to state or hear a number, price, deadline, or rating, and the stakes are real.
A number mentioned before you have formed your own view, especially one chosen by someone with something to gain, or one you generated yourself earlier without realizing it would anchor you.
What would I estimate if I had never heard this number at all?
Assuming experience makes you immune. Experienced real estate agents anchored on a property's listing price in their own valuations, even though most denied using it (Northcraft and Neale, 1987).
Learn it progressively
Identify the first number, date, or price that entered the conversation, and who introduced it.
Before hearing further numbers, form an independent estimate using your own reasoning or comparable data, on paper if possible.
People typically adjust too little away from an anchor. If your independent estimate differs from the anchor, push your figure further than intuition suggests, then recheck it.
Pull in reference points unconnected to the anchor: market data, historical base rates, or a second opinion from someone who has not heard the anchor.
Make the call, and write down the anchor you were exposed to and how far your final number sits from it, so a later review can check whether the anchor still leaked in.
See it in different settings
Salary negotiation
A recruiter asks for a candidate's desired salary before naming a range. Whatever number the candidate states becomes the reference point the employer negotiates down from, even when it is below market.
Real estate
A buyer sees a comparable listing price before viewing a property, and even professional appraisers' final valuations drift toward it despite training meant to prevent exactly that (Northcraft and Neale, 1987).
Retail pricing
A tag reading "was $120, now $60" anchors shoppers to $120, making $60 feel like a bargain regardless of what the item is actually worth.
Project estimation
A stakeholder mentions offhand that a project "should take about two weeks" before the team scopes the work. The team's own estimate clusters near two weeks even after they uncover more work than expected.
Working with an AI assistant
An AI states a rough figure early in a chain of reasoning, and both the assistant's later steps and the person reading the answer keep treating that early guess as fixed rather than provisional.
Worked application
Negotiating a starting salary offer
Notice the anchor: The recruiter opens with a number in the offer email, or asks for your expected range before discussing the role in detail.
Build your own estimate first: Before responding, check published salary bands and industry compensation surveys for the role, independent of any number the recruiter has mentioned.
Adjust further than feels natural: If your independent research lands meaningfully above the offer, counter further than feels comfortable. Studies of negotiation find that first offers anchor the final settlement more than most negotiators expect, so a timid counter still concedes ground to the anchor (Galinsky and Mussweiler, 2001).
Bring in outside comparables: Ask a peer in a similar role, or check a public compensation database, for a second data point untouched by the original offer.
Decide and record: State your counter-offer, note the anchor you were given, and record the gap between them so your next negotiation starts from your own research rather than someone else's opening number.
Common mistakes
Naming your own number second, which lets the other side's opening figure set the anchor instead of yours.
Adjusting only slightly from a stated anchor because a small move feels like enough correction.
Assuming professional experience or expertise makes you immune to anchoring.
Treating a random or irrelevant number, an old contract figure, a prior year's budget, a rigged spinning wheel, as if it carries real information.
Using the label "anchoring" to explain away every high or low number you disagree with, instead of checking whether an actual anchor was present.
When not to use it
Not every early number is a meaningless anchor. Sometimes the first figure genuinely is the best available estimate, based on real data rather than chance. Do not use this model to dismiss every asking price or quote as manipulation, to justify refusing to ever name a number first out of pure caution, or to claim you were not influenced simply because you do not feel influenced. The model is a check on whether a number shaped your judgment through mere order of exposure, not a reason to treat every figure you hear as suspect.
Related models
Confirmation Bias
Once anchored to a number, people tend to search for reasons that number was right, rather than testing it directly.
Bayesian Updating
The disciplined alternative: move your estimate in proportion to genuine evidence, rather than in proportion to whichever number arrived first.
Loss Aversion
An anchor often works by framing a decision as a loss or a gain relative to that reference point, which is what gives it emotional pull.
Expected Value
Calculating an expected value independently, before hearing any other figure, is one of the most reliable defenses against an anchor.
Try It Yourself
Save this as a Skill
This turns the model above into a short, reusable instruction set you can keep and paste into any AI assistant before you react to a number.