Online research: only 1% of responses written with ChatGPT, but almost always paid even when it was forbidden

By linking ChatGPT histories given by participants to 712,930 responses, a study directly measures the use of AI on a scientific investigation platform.

Online participant platforms provide much of the data from the humanities and behavioral sciences, assuming that a human has written each answer well. To measure the actual use of AI, researchers surveyed 2,500 workers at a reputable platform, then cross-referenced ChatGPT histories given by weekly users with their submissions: 712,930 responses to more than 127,000 studies.

Rare, focused, but little punished

One in eight respondents say they have already used AI for a study, and 68% of the workers observed have done so at least once — yet assistance appears in only 1% of submissions, with no detectable increase in more than three years. It is highly concentrated: 5% of workers produce 64% of assisted responses. These are almost always paid, even when the instructions forbade AI, with approval rates comparable to others.

Half of the cases concern closed responses, which the platform's detector, designed for free texts, does not cover. The authors conclude that AI does not yet threaten online search, but that safeguards would be insufficient if the practice were to spread.