An AI test can mark a good answer wrong
A preprint checked 16 AI test resources. None accounted for the needs of the people meant to use the results. In one paired test, a change in answer wording sharply cut exact-match scores.
AI AgencyRead story
Lead story
Feedback can change how people check AI advice
In a study of 244 people, feedback after an AI override changed how people thought through their next choice. It did not improve overall team results.
Source: arXiv · Oct 14, 2026Read original source
Covered by both lenses
Read story
More top stories
Lab safety needs more than written rules, essay argues
Zvi Mowshowitz argues that AI labs need safety expertise, backup checks and the will to slow a release. His claims about specific labs were not independently checked here.
AI AgencyRead story
Featured
Nurse scheduling AI showed uncertainty people could misread
In an expert test of a nurse-scheduling tool, people often misread a sign of uncertainty and felt less sure of their choice. Liked explanations did not clearly change confidence or advice use.
Beyond PromptingRead story
Culture & Human Agency
A kids’ streaming app tries to put a stop button back in
WIRED reports that Maka offers kids’ shows without ads or autoplay. Parents can set a session timer. Those features are not proof of better outcomes for children.
AI AgencyRead story
Research & Science
The real test may come after AI steps away
A preprint reviews four cases of human work with AI. It asks whether people keep their own skill and judgment after using a tool. Later effects differed across the cases.
Beyond PromptingRead story
