TL;DR
Long “thinking” in reasoning models doesn't make their judgments more rational. Reasoning models (the o-series, Claude and Qwen in thinking mode, DeepSeek-R1) are no less prone to cognitive biases than their regular versions. The point estimate is even slightly worse, though the difference isn't statistically reliable. Spending more tokens on reasoning doesn't reduce bias either. Where anything changed, the answer drifted further from the human pattern, but not toward neutral.
The model still depends on irrelevant details in the request. You ask for a price estimate, and somewhere in the context there's someone else's number, like “a friend of mine got it done for $3,000.” The answer shifts toward that number. That's anchoring, the only bias that works in models the same way it works in humans, and it showed up in all seven models tested. The other five biases show up inconsistently, often in reverse. The reason: a long chain of thought doesn't force the model to separate what matters from what's incidental. It just generates more text around the same answer.
A cheap trick: one line in the prompt. When the model was required to restate every number in the request before answering, anchoring went down on all five tested tasks. Extra “thinking” didn't do that. The effect is small and borderline significant (p = .057 across five tasks), but the direction was the same everywhere.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
STEP 1: Add to the request: “Before answering, list every number
mentioned in the request” → list of numbers
STEP 2: The model gives its estimate → answer after the list
(all in one prompt, one request)
Takeaway for choosing a model: don't buy “rationality” with thinking mode. Test the model separately for the specific bias you care about.
Example in Practice
Task: You run a dev studio and ask a model to estimate the market price of building a Slack bot for an online store. The client's email contains the line: “A friend of mine had something similar built for $3,000, but we're open to discussion.” If you paste the whole email into the model, that number becomes the anchor.
Prompt:
Below is a client's email. Estimate the fair market cost
of building a Slack bot in the US (in dollars, one range).
Before answering, list every number mentioned in the email.
Email:
“Hi! We need a Slack bot for our clothing store: product catalog,
cart, payments via Stripe, notifications to the store manager. A friend of mine
had something similar built for $3,000, but we're open to discussion.
We'd like to launch by April 1. The catalog is about 300 items.”
At the end, write: “Estimate: from … to … dollars”.
Result: First the model will list every number in the email: $3,000, April 1, 300 items. Then it gives a price range. Set your expectations correctly: the paper shows a small reduction in the pull toward the anchor, not its disappearance. So for money decisions, compare the answer with an estimate from the same request without the anchor number.
Why It Works
The weakness. A reasoning model generates a long text before answering, but that text has no obligation to “check” irrelevant details. The researcher points to related work: chains of thought often justify an answer that was actually determined by a hint the text never mentions. So more tokens don't buy more rationality.
The strength. Models are good at following direct instructions about output format. Require the numbers to be listed explicitly and they land in the answer text as separate objects rather than background.
How the method gets around the weakness. Listing turns the anchor from hidden background into an explicit element. What exactly happens next wasn't tested mechanistically. That explanation remains a hypothesis; the only thing verified is that the effect exists.
The levers: - What to restate. In the paper: “every number.” For other biases (framing, commitment) you could restate conditions or wording, but that hasn't been tested. - Reasoning mode. Don't turn it on to reduce bias. The token cost is higher and there's no payoff on this front.
Prompt Template
{task: what to estimate or choose}
Before answering, list every number mentioned in the request.
{context with numbers}
Answer format: {format}
Fill in: {task} (a price, timeline or budget estimate), {context} (an email, a brief, a description of the situation), {format} (a range, an option from a list, a number).
Limitations
⚠️ The effect is small and borderline significant: anchoring went down on all five tasks, but with five tasks the statistics fall just short of the threshold (p = .057). It's a direction, not a guarantee.
⚠️ Only anchoring was tested: no equivalent trick was tested for framing, loss aversion, escalation of commitment, availability or confirmation bias.
⚠️ “Reverse” bias doesn't mean “rational”: in models, framing runs against the human pattern, but the answer still depends on an irrelevant detail, just in the other direction.
⚠️ A narrow range of “thinking”: Claude actually used up to ~800 tokens, GPT-5.1 about 120 regardless of the requested limit. DeepSeek-R1 and Qwen3 ignored the limit entirely. The result doesn't prove that very long thinking can never help.
⚠️ A different task setup: these are management cases with a choice among numbered options. Free-form text and your own context may behave differently.
How They Studied It
The idea was simple: if a reasoning model “thinks like a person who switched on slow thinking,” then the longer it thinks, the weaker its biases should be. The author took 30 management cases from the Malberg et al. benchmark covering six biases (five per bias). Each case existed in two versions: a control and one with a “poisoned” detail. Seven models from four families (Claude Haiku 4.5, GPT-5.1, DeepSeek, Qwen3) were run, each in its regular and thinking version, with reasoning limits of 0, 1024, 4096 and 8192 tokens. Every variant was run ten times, for a total of 12,350 calls. A curious detail: the limit in the request doesn't equal the actual reasoning. Claude respected it, GPT-5.1 produced nearly the same ~117 tokens at any limit, and DeepSeek and Qwen ignored it and “thought” for ~1000 tokens even at zero. So the author used the tokens actually spent as the dose. The result went against intuition: no “more thinking, less bias” slope turned out significantly negative. An extra surprise: four of the other five biases run against the human direction in models.
Adaptations and Extrapolations
🔧 Technique: an audit with paired prompts → see where your task is vulnerable
The paper's method boils down to this: two identical requests that differ only in an irrelevant detail (a number, a frame). Run each 5–10 times and compare the average answers. For your task: “estimate the cost” with the friend's number and without it.
Extrapolation (my assumption, not tested in the paper). A rule for an analyst's system prompt:
Before giving any estimate, list all the numbers mentioned in the request.Test it on your own cases before you trust it.
Resources
- The Long Road to the Same Answer: Cognitive Bias Under Escalating Reasoning Budgets in Large Language Models, Obada Kraishan, College of Media and Communication, Texas Tech University. Accepted at IEEE CogMI 2026.
- Repository: https://github.com/obadaKraishan/anchored-minds
- Source benchmark: Malberg et al. (30 biases, 200 scenarios).
- Related work: Turpin et al. and Chen et al. on the incomplete faithfulness of chains of thought, Knipper et al. on the robustness of R1 and QwQ to the framing effect.