Last updated: April 2026
The anchoring effect is a cognitive bias where the first number you encounter distorts every judgment that follows, even when that number is obviously random. Discovered by Tversky and Kahneman in 1974, it inflates salary offers, real-estate prices, legal sentences — and, as 2025 research confirms, the outputs of large language models like GPT-4 and Claude. Debiasing requires more than “try to ignore it”: the brain adjusts away from the anchor, but it stops adjusting too soon.
What is the anchoring effect?
The anchoring effect is a cognitive bias in which people rely too heavily on the first piece of numerical information they receive — the anchor — when estimating an unknown quantity. Subsequent judgments are then adjusted from that anchor, but the adjustment is almost always insufficient. The final answer stays biased toward the original number, even when the number is irrelevant or explicitly random.
The effect was first formalized by Amos Tversky and Daniel Kahneman in their landmark 1974 Science paper “Judgment under Uncertainty: Heuristics and Biases”. It is now considered one of the most robust findings in behavioral economics — replicated across cultures, age groups, experts, and, as of 2025, large language models. If you’ve read our explainer on cognitive biases, anchoring is the textbook example of a System 1 heuristic in Kahneman’s dual-process framework.
How does the anchoring effect work?
Anchoring works through a mechanism called selective accessibility. When you encounter a number, your brain unconsciously retrieves information consistent with that number being the correct answer. This retrieved information then dominates the final estimate — even when you know the anchor is wrong.
In the classic wheel-of-fortune experiment (Tversky & Kahneman, 1974), researchers spun a wheel rigged to land on either 10 or 65, then asked participants to estimate the percentage of African countries in the United Nations. Those who saw the number 10 gave a median estimate of 25%. Those who saw 65 answered 45%. The wheel was openly random. It didn’t matter.
Anchoring is not a failure of effort. In follow-up experiments, offering cash rewards for accurate answers did not reduce the bias. The anchor operates below conscious control — you can know it’s irrelevant and still be pulled toward it.
7 real-world examples of the anchoring effect
1. “Original price” strikethroughs in retail
When a product shows $199 $89, the strikethrough isn’t information — it’s an anchor. Research by Biswas & Blair (1991) found consumers perceive the discounted price as a better deal when paired with a higher reference price, even when the “original” never applied. EU Directive 2019/2161 (Omnibus Directive), in force since May 2022, now requires retailers in the EU to display the lowest price from the last 30 days as the reference — a direct legal response to anchoring exploitation.
2. Salary negotiations
Whoever states the first number sets the anchor. A 2011 study by Galinsky, Ku & Mussweiler in the Journal of Personality and Social Psychology found that final negotiated salaries correlate 0.85+ with the opening offer. This is why HR departments often push candidates to disclose current salary first — and why 22 U.S. states and most EU member states have now banned this practice under equal-pay legislation.
3. Real estate listings
Northcraft and Neale (1987) gave real estate agents — professionals — identical property tours with different listing prices. Their appraisals moved in lockstep with the listing: an $11,000 higher listing produced appraisals ~$14,000 higher. The agents denied being influenced. They were wrong by 11-12%.
4. Judicial sentencing
In a widely cited 2006 study, German judges with 15+ years of experience rolled loaded dice before sentencing a hypothetical shoplifter. Judges who rolled a 9 recommended an average sentence of 8 months. Those who rolled a 3 recommended 5 months. The dice were explicitly random. The judges knew.
5. Restaurant menu design
Placing a $95 bottle of wine at the top of the list isn’t meant to sell $95 wine. It makes the $45 bottle feel reasonable. Menu engineering research (Yang, 2012, Cornell Hospitality Quarterly) shows this tactic consistently shifts orders toward the second-most-expensive item — exactly where restaurants place their highest-margin dishes.
6. CFD trading and technical levels
In my own trading on Plus500 using SMC methodology, anchoring is constant. The previous day’s high, a round number like gold at $2,000, or a clear swing point all become psychological anchors that traders cluster entries and stops around. This is why spread widens at these levels — market makers know the crowd is anchored. The SMC concept of “liquidity grabs” is essentially a trade setup built around exploiting anchoring in retail order placement.
7. LLM outputs (the 2025 finding)
A 2025 experimental study published in the Journal of Computational Social Science tested whether large language models exhibit anchoring bias. They do. When GPT-4, Claude, and Llama-class models were given biased numerical hints before a neutral question (“Is the Nile longer than 25,000 km?” vs. “longer than 1,000 km?”), their final estimates tracked the anchor. A parallel ACL 2025 paper on price negotiation simulations confirmed LLMs accept disadvantageous prices when a high anchor is introduced first. This matters because AI systems are increasingly embedded in hiring, pricing, and recommendation pipelines — and they inherit the same bias they’re supposed to neutralize.
How the anchoring effect is measured
The standard experimental design has two conditions:
| Step | Low anchor condition | High anchor condition |
|---|---|---|
| 1. Comparison question | “Is the Mississippi shorter or longer than 500 miles?” | “Is the Mississippi shorter or longer than 5,000 miles?” |
| 2. Estimation question | “How long is the Mississippi?” | “How long is the Mississippi?” |
| 3. Typical result | Median: ~1,400 miles | Median: ~2,700 miles |
The Anchoring Index (Jacowitz & Kahneman, 1995) quantifies the effect: (high-anchor median − low-anchor median) / (high anchor − low anchor) × 100%. Across hundreds of replications, this index consistently lands between 40% and 60% — meaning roughly half of any anchor’s magnitude leaks into the final judgment.
Why does the anchoring effect matter in 2026?
Three converging trends make anchoring more consequential now than at any point since 1974:
1. AI-mediated pricing. Dynamic pricing systems, AI-powered negotiation bots, and LLM-based recommendation engines all rely on first-mover numerical cues. If the LLM itself is anchored (as the 2025 research shows), the bias compounds rather than cancels.
2. EU AI Act and consumer protection. The EU AI Act (in force August 2024, high-risk obligations from August 2026) classifies systems that manipulate consumer decisions through known cognitive vulnerabilities as prohibited practices under Article 5. “Deceptive anchoring” in AI-driven retail is a live area of regulatory scrutiny.
3. Dark pattern regulation. The Digital Services Act (DSA) and Omnibus Directive jointly treat fake reference prices and deceptive anchors as unfair commercial practices. Fines reach 4% of annual EU turnover — Amazon, Wish, and Zalando have all been investigated on these grounds since 2023.
How to defend against the anchoring effect
The honest answer: you mostly can’t — but you can reduce it. Four evidence-based tactics:
Generate your own estimate first. Before looking at any listing price, salary offer, or menu, write down what you think the number should be. This pre-registers your judgment and creates a counter-anchor.
Consider the opposite. Mussweiler, Strack & Pfeiffer (2000) showed that explicitly asking “why might this anchor be wrong?” reduced anchoring by ~40%. Not eliminated — reduced.
Widen the reference set. Instead of comparing against one number, force yourself to think of three or four data points. This breaks the selective-accessibility mechanism.
For LLM use specifically: the 2025 ACL paper found that chain-of-thought prompting alone does not eliminate anchoring in models. What helped was structured multi-perspective prompting — asking the model to generate the answer from multiple independent angles before combining them. If you’re building an LLM pipeline for pricing or negotiation, bake this in.
In behavioral finance, anchoring manifests as reluctance to close losing positions below the entry price (the entry becomes the anchor) and reluctance to take profits far above a “round number” target. If your trading journal shows clustered exits at round numbers, you’re anchoring.
Anchoring effect vs. related biases
| Bias | What it is | Key difference from anchoring |
|---|---|---|
| Anchoring | First number biases estimate | — |
| Framing effect | Same info presented differently changes choice | No numerical anchor; changes decision via context |
| Availability heuristic | Recent/vivid examples bias probability estimates | Memory-based, not reference-point-based |
| Status quo bias | Current state becomes default preference | Anchor is existing state, not presented number |
| Loss aversion | Losses feel ~2× stronger than equivalent gains | Concerns outcome valuation, not numerical judgment |
For a full map of the biases listed above, see our cognitive biases explained overview. For Kahneman’s broader framework describing why the mind falls for them, prospect theory and nudge theory are the natural next reads.
Key takeaways
The anchoring effect is one of the most replicated findings in cognitive science: the first number you see pulls every subsequent judgment toward it, and that pull persists even when the number is random, irrelevant, or explicitly flagged as such. It shapes retail pricing, salary negotiations, real estate, legal sentencing, and — as 2025 research confirmed — the outputs of the AI systems we’re building to replace human judgment. You can’t fully escape it. You can reduce its grip by generating independent estimates, considering multiple reference points, and, when using LLMs, prompting for multi-angle reasoning rather than single chain-of-thought.
Frequently Asked Questions
Who discovered the anchoring effect?
Amos Tversky and Daniel Kahneman first documented the anchoring effect in their 1974 paper “Judgment under Uncertainty: Heuristics and Biases” published in Science. Kahneman later received the 2002 Nobel Prize in Economics partly on the strength of this line of research.
Does the anchoring effect work on experts?
Yes. Northcraft and Neale’s 1987 study on real estate agents and Englich, Mussweiler and Strack’s 2006 study on judges with 15+ years of experience both showed that professional expertise provides almost no protection against anchoring. Experts are often more confident they’re not anchored — but the measured bias is comparable to laypeople.
What’s the difference between anchoring and framing?
Anchoring involves a numerical reference point that biases a numerical estimate. Framing involves how information is presented (as gain vs. loss, for example) and biases the chosen option rather than the estimated quantity. Both are System 1 effects in Kahneman’s framework, but they operate through different cognitive mechanisms.
Do large language models (LLMs) show anchoring bias?
Yes. A 2025 study published in the Journal of Computational Social Science and a separate ACL 2025 paper on price-negotiation simulations both demonstrate that GPT-4, Claude, and Llama-class models exhibit measurable anchoring when given biased numerical hints before a neutral question. Simple chain-of-thought prompting does not eliminate the bias — multi-perspective prompting is more effective.
How does the anchoring effect apply to trading?
Traders anchor on entry prices (reluctant to exit below them), round numbers (e.g., gold at $2,000, EUR/USD at 1.1000), and previous session highs/lows. In SMC methodology, “liquidity grabs” exploit exactly this: price briefly sweeps an obvious anchor level to trigger clustered stop-losses, then reverses.
Is anchoring the same as confirmation bias?
No. Confirmation bias is the tendency to seek and interpret information that confirms existing beliefs. Anchoring is a narrower effect that operates specifically on numerical estimates, pulling them toward an arbitrary reference point. They can compound — an anchor becomes a belief, and confirmation bias then protects it — but they are distinct mechanisms.
Can you train yourself out of anchoring?
Only partially. Awareness alone doesn’t work — cash incentives for accuracy don’t eliminate it, and experts remain affected. The most effective techniques are procedural: generate your own estimate before seeing any external number, explicitly consider why the anchor might be wrong, and force yourself to reference multiple data points rather than one. These reduce but don’t erase the bias.
Bibliography
- Tversky, A. & Kahneman, D. (1974) — Judgment under Uncertainty: Heuristics and Biases. Science, 185(4157).
- Lou, J. et al. (2025) — Anchoring bias in large language models: an experimental study. Journal of Computational Social Science.
- ACL 2025 — How Does Cognitive Bias Affect LLMs? A Case Study on the Anchoring Effect in Price Negotiation Simulations.
- Nguyen, T. et al. (2024) — Human bias in AI models? Anchoring effects and mitigation strategies in large language models.
- Lin, Z. et al. (2025) — An Empirical Study of the Anchoring Effect in LLMs. arXiv:2505.15392.
- European Union — AI Act, Article 5 (Prohibited AI Practices).
- EU Directive 2019/2161 (Omnibus Directive) — consumer protection against fake reference prices.
- Englich, B., Mussweiler, T. & Strack, F. (2006) — Playing Dice with Criminal Sentences. Personality and Social Psychology Bulletin, 32(2).
- Northcraft, G. & Neale, M. (1987) — Experts, Amateurs, and Real Estate: An Anchoring-and-Adjustment Perspective. Organizational Behavior and Human Decision Processes, 39(1).
- Galinsky, A., Ku, G. & Mussweiler, T. (2011) — To start low or to start high? The case of auctions versus negotiations. Current Directions in Psychological Science.
- Jacowitz, K. & Kahneman, D. (1995) — Measures of Anchoring in Estimation Tasks. Personality and Social Psychology Bulletin, 21(11).
1 thought on “Anchoring Effect: 7 Examples & Why It Tricks AI Too”