Cognitive-forcing nudges that fight AI rubber-stamping help most the people who barely needed the help in the first place.

This explores which personal traits decide whether someone actually gains from an AI decision aid rather than just using one. The corpus has one direct answer, and it's uncomfortable. In an experiment with 199 participants, cognitive forcing functions helped people high in Need for Cognition much more than people low in it. Need for Cognition is a measured trait for how much someone enjoys effortful thinking. Cognitive forcing functions are design tricks that make users think before they see or accept the AI's suggestion. So the fix designed to stop people from rubber-stamping AI output works best for the people who were least likely to rubber-stamp it in the first place. The aggregate numbers looked like a success, but they hid a gap that the intervention itself created Do cognitive forcing functions help all decision makers equally?.

That result changes the question. The trait that matters may not be 'who is good with AI' but 'who is willing to do the extra thinking the tool asks for.' It also points to a design tension: a guardrail that relies on effort will widen gaps between thinkers and non-thinkers unless it is built for the people who would rather skip the work.

A second clue is how people mentally size up an AI partner. When users model a dialogue agent, perceived competence accounts for about half of their impression (49% of the variance). Human-likeness and communicative flexibility explain the rest How do users mentally model dialogue agent partners?. This doesn't measure benefit directly. It does suggest that how much someone gets from a tool may depend on how accurately they judge its competence, and that people differ in how much weight they give its human-like surface.

Trust also changes with experience, which complicates any idea of fixed traits. In repeated partner-selection games with 975 people, participants at first avoided partners they knew were AI. Over many rounds they came to prefer them, because the bots behaved more reliably and with less variance than humans Do humans learn to prefer AI partners over time?. An anti-AI bias at the start doesn't predict long-term avoidance. Repeated exposure with feedback can override it.

The corpus is thin here. Beyond Need for Cognition, it has no direct evidence on expertise, numeracy, age, or prior AI experience as predictors of benefit. Most of the nearby material is about the 'traits' of AI agents rather than of human users. The takeaway is that the best-documented predictor is a motivational trait rather than a skill, and that a well-meant design can quietly reward people who already think hard.