Mission: stop AI outputs and my own estimates from being pulled toward a number I supplied.
Anchoring is the pull an initial number or suggestion exerts on later estimates, even when that number is irrelevant. Kahneman's book covers the anchoring experiments in its Part II heuristics discussion (see Sources).
It works on AI in two directions. When your prompt contains a number, a draft, or a conclusion, the model's answer tends to drift toward it. And when the AI gives you an anchor first, your own judgment drifts toward it.
A 2025 benchmark tested anchoring across tiny, standard, advanced, and reasoning models, including GPT-4o, DeepSeek-R1, and Qwen3. Anchoring appeared in all of them, and the mitigations it tested reduced the effect only modestly, by up to about 10 percent. Reasoning was the exception and helped somewhat (Huang et al., arXiv 2505.15392). A separate mechanistic study on smaller open models found anchors reweight the output distribution, and its effects varied across prompt designs (Valencia-Clavijo, arXiv 2511.05766, a preprint).
Choose the bias the scenario shows.
Anchors in your prompt steer the answer, and anchors in the answer steer you. Form your own rough estimate before you ask, and ask for a range rather than a single number.
Try it this week: ask the AI for an estimate with no number in the prompt, then compare it to one you gave it. Note the gap.