Survey questions about trusting 'AI' can accidentally measure something else entirely — how do you word them to hit the real target?
How should machine heuristic items be worded to match their theoretical purpose?
This explores how survey questions that measure the 'machine heuristic' (the mental shortcut that machines are objective, unbiased and precise) should be phrased so they capture that general belief rather than something nearby, such as trust in one specific system.
This explores how survey questions that measure the 'machine heuristic' should be worded so they capture the belief they are meant to capture. The machine heuristic is the snap assumption that a machine is more objective, neutral or accurate than a person. The direct answer is that this collection has nothing on it. None of the retrieved notes cover survey design, scale validation or the communication research where the machine heuristic comes from. Questions like whether items should be phrased about 'machines in general' or 'this AI', or whether they should be split into objectivity, accuracy and lack of bias, can't be answered from what's here.
The corpus does offer one adjacent point. It suggests why the wording problem matters: there is often a large gap between how a machine's output looks and what produced it. Chain-of-thought text looks like step-by-step logic but works by copying familiar reasoning patterns Does chain-of-thought reasoning reveal genuine inference or pattern matching?. Models that seem to reason carefully about constraints often just default to the harder-looking option Are models actually reasoning about constraints or just defaulting conservatively?. Models asked to run numerical methods produce confident, precise-looking numbers that are wrong Do large language models actually perform iterative optimization?.
For anyone writing scale items, that gap is the case for keeping two things apart: the heuristic itself (a general, prior belief about what machines are like) and a judgment about how well a particular system performed. If an item asks whether 'the AI's answer was accurate', it measures performance. If it asks whether 'machines are less biased than people', it measures the heuristic. The research above suggests these can come apart sharply, so mixing them in one item would hide the very cases where the shortcut misleads people. To go further on item wording itself, you would need sources outside this collection, such as Sundar's work on the MAIN model and later validation studies of the machine heuristic.
Sources 3 notes
CoT works by constraining models to reproduce familiar reasoning patterns from training, not by enabling novel symbolic reasoning. Performance degrades predictably under distribution shifts—the signature of imitation rather than capability emergence.
Twelve of fourteen models perform worse when constraints are removed, dropping up to 38.5 percentage points. Models appear to reason correctly by defaulting to harder options, not by actually evaluating constraints.
Research shows LLMs cannot perform iterative procedures in latent space. They recognize optimization problems as template-similar and emit plausible-looking but incorrect values, a failure mode that persists across model scale and training approaches.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Can Large Language Models Reason and Optimize Under Constraints?
- Beyond Accuracy: Evaluating the Reasoning Behavior of Large Language Models -- A Survey
- CoT is Not True Reasoning, It Is Just a Tight Constraint to Imitate: A Theory Perspective
- Measuring Faithfulness in Chain-of-Thought Reasoning
- Hierarchical Reasoning Model
- When More is Less: Understanding Chain-of-Thought Length in LLMs
- Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens
- Advancing Language Model Reasoning through Reinforcement Learning and Inference Scaling