Fact-check passed; 20 overstatements remained, 18 found by AI

The author's diagnosis is that a sentence can pass a correctness check and still be wrong in strength. His seven types — comparison, causation, self-description, recorded actions, search scope, quoted wording and subject — all failed the same way: not by inventing facts but by writing them a notch stronger than the record warranted. One example, "the only difference was permissions," collapsed once the agent definitions were opened and the models and tools turned out not to match.

The most striking finding is who did the fixing. Across the 20 logged corrections from his two earlier articles, 18 came from the AI producing the article while opening primary records during pre-publication review, 1 from the fact-checking agent (a missing requirement in official documentation), and none from the author, whose role was approving fixes. He reasons the difference is what each step consults: fact-checking hits outside sources, while tracing hits the writer's own records.

A 2025 Royal Society Open Science study points the same way, the author notes. Ten LLMs summarizing 4,900 scientific and medical abstracts produced broader generalizations than the source even under accuracy-focused prompts, at an odds ratio of 4.85 versus human summaries, with three Claude models not showing a significant difference at the 5% level. The author flags the limits plainly: the study covers abstract summarization, not blog writing or working from records, and his own observations span just two articles, so the shared direction is not proof of a shared cause.