Does the social penalty for using AI fade once the tool becomes familiar, and how would you test that over time?

This explores what a study over time would need to show whether the competence penalty for using AI fades as the tool becomes ordinary. The short answer is that the corpus names this exact gap but contains no such study. People expect to be rated as less competent for using AI. Nobody has tracked whether that judgment softens with familiarity. There is also a competing explanation: the penalty may come from AI's *agency* (the sense that the tool did the work), not its novelty, and agency doesn't wear off with habit Does the social penalty for AI use fade as the tool becomes ordinary?. A good design therefore has to separate those two explanations, not just watch a number over time.

The closest model in the collection is a partner-selection experiment with 975 participants. Bots whose identity was disclosed were avoided at first. Over repeated rounds they won people over, because participants learned that bots behaved reliably and generously Do humans learn to prefer AI partners over time?. That gives you a template: disclosed AI use, repeated rounds, and a behavioral choice instead of a survey answer. Adapted to this question, the same evaluators would judge the same colleagues over weeks or months, with each colleague's AI use disclosed. You would then vary two things independently. One is how much hands-on exposure the evaluator has to the tool. The other is how much of the work the AI did (light editing versus drafting the whole thing). If the penalty fades with exposure regardless of how much the AI did, novelty was the cause. If it persists whenever the AI did a lot of the work, the cause is how credit for the work gets assigned.

Two cautions from neighboring research shape how you'd measure it. First, early impressions are unreliable. Studies of the chatbot Mitsuku found that the social dynamics present in first sessions faded predictably with repeated use, so single-session results can't predict long-term behavior Do chatbot relationships lose their appeal as novelty wears off?. You need several measurement points and a group with no added exposure, so that general cultural change isn't mistaken for familiarity. Second, not every rating reflects a real attitude. Work on annotation data separates genuine preferences from non-attitudes and from preferences people construct on the spot, and the test is whether a response holds steady across measurement conditions Do all annotation responses measure the same underlying thing?. If people's competence judgments change with question wording at the start of the study, they may never have held a firm view. In that case, an apparent 'fading' could just be noise settling down rather than a real change in attitude.

One tempting shortcut is to simulate the study with LLM personas before running it on people. The corpus suggests this would fail in this particular case. Personas reproduce published experimental effects about 76% of the time, but success depends on the effect being strong, and small effects produce both false positives and false negatives Can AI personas reliably replicate human experiment results?. Personas built from behavioral data predict A/B test direction well for large effects and poorly for near-zero ones Can behavior-based personas predict A/B test outcomes?. A slowly fading social penalty is the kind of small, gradual effect that simulation handles worst. So this question needs real people tracked over real time.