Can an AI steer your attention without steering your answer — and still stop you from just copying its view?
Can interpretive guidance in AI output prevent anchoring without requiring active decisions?
This explores whether an AI that points out what is worth noticing, instead of handing over an answer, can stop people from latching onto the AI's view, and whether that works without adding friction such as making users commit to their own judgment first.
This explores whether an AI that points out what's worth noticing, instead of handing over an answer, can stop people from latching onto the AI's view, and whether it can do that without adding friction such as making users commit to a judgment first. The clearest material in the collection is the Learning to Guide work Can AI guidance reduce anchoring bias better than AI decisions?. It changes the usual setup, in which the AI either decides a case or hands it off to a human. Instead, the AI never gives a verdict. It highlights the parts of the input that matter, and the human makes the call. Anchoring happens when there is a ready-made answer to latch onto. If there is no answer on the screen, there is nothing to anchor to. That is the surprising part: the design avoids anchoring by leaving the answer out, not by asking the person to try harder.
There is a catch in the phrase "without requiring active decisions." Guidance doesn't remove the human's decision. It makes the decision unavoidable, and that is why it works. The friction is not gone. It has moved from an added step, like 'write down your guess before you see the AI's,' into the task itself. The person still has to think. They just aren't told what to conclude.
The collection also points to a weak spot. Highlighting is not neutral. The Rose-Frame analysis Why do people trust AI outputs they shouldn't? describes how people mistake AI output for reality, take fluent output as careful reasoning, and get their existing beliefs confirmed. These traps get worse when they occur together. An AI that chooses which features to emphasize is still steering attention, and a reader may anchor on what was highlighted instead of on a verdict. Research on reasoning traces Can we actually trust reasoning model outputs? adds a related warning. An AI's account of why something matters often doesn't reflect what actually drove its output. So guidance can look like a balanced pointer while carrying a hidden lean.
The collection is thin on the direct question. One paper tests guidance-instead-of-answers. It doesn't compare that design with other anti-anchoring approaches, and it doesn't measure whether people anchor on the highlights themselves. The answer it supports: yes, leaving the verdict out removes the most obvious anchor. But the anchoring risk moves to the highlights rather than disappearing, and nothing in the collection yet tests that next layer.
Sources 3 notes
Learning to Guide eliminates anchoring bias and unassisted hard cases by having machines supply interpretive guidance rather than autonomous decisions, keeping responsibility with humans while improving their judgment through enhanced perception.
Rose-Frame identifies map-territory confusion, intuition-reason conflation, and confirmation-bias reinforcement as traps that multiply their distorting effects when they co-occur. Evidence from cross-linguistic overreliance and architectural transformer biases confirms the compounding mechanism operates universally.
Research shows reflection rarely corrects errors, traces rarely explain decisions faithfully, and monitoring is vulnerable to two failure modes: omission (influence never reaches the trace) and laundering (problematic reasoning appears in clean language). These vulnerabilities persist even under evaluation pressure.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making
- Beyond Hallucinations: The Illusion of Understanding in Large Language Models
- Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation
- Learning To Guide Human Experts Via Personalized Large Language Models
- A Rational Analysis of the Effects of Sycophantic AI
- GenAI as a Power Persuader: How Professionals Get Persuasion Bombed When They Attempt to Validate LLMs
- On the Reasoning Capacity of AI Models and How to Quantify It
- Large Language Model Reasoning Failures