A study of 22 documents from three public AI projects argues that official records can leave out objections, including calls to stop a project. Counting comments is not the same as showing how they shaped a decision.
In his Import AI newsletter, Jack Clark covers a poll on AI rules and an analysis of parallel AI workers. Splitting work may cut waiting time while raising the total amount of computing used. Neither finding proves a general speed gain.
In a 16-person writing study, people could preview style changes and return to their original text. They reported less mental effort with this tool. The study did not establish that the new words were more accurate.
Researchers asked reviewers how they check sources. Their study included 12 interviews and a survey of 203 people. Reviewers used different methods. A linked paper alone does not show that the sentence beside it is right.
In two experiments, people felt in control of a computer's first action, yet judged their role in its later outcome differently. The study makes a useful distinction. It did not test whether review steps make real AI systems safer.
Kevin Roose told WIRED he used AI to sort notes, find leads and critique a draft of his book. He says he wrote the book himself and kept a human fact-checker. This is his account, not a quality study.