A Canadian politician has delivered one of the best accidental demonstrations yet of what happens when AI is used without the final step — the one where a human actually reads what the machine has written. On June 9, Bill Oliver, a Conservative member of New Brunswick’s legislature, was roughly halfway through a 30-minute speech when, after an entirely ordinary sentence, he calmly continued: “Here’s a more natural, flowing version of that section that reads like a legislative speech rather than a series of short points.” He did not stop. The chamber did not appear to notice either. A few minutes later, Oliver read another similar instruction referring to a “more developed, flowing version.” The footage only went viral in July. Oliver later acknowledged using artificial intelligence while preparing the speech and said the AI instructions had remained in the final version when it was printed. The AI had therefore done exactly what it was asked to do: make the politician’s speech sound more natural. The only person who apparently did not read it before it was delivered was the politician himself.

The incident is also a neat example of what is sometimes called “AI residue” — traces accidentally left behind when chatbot output is copied without a final human review. They can include phrases such as “here is an improved version,” comments addressed directly to the user, introductory notes, or explanations of how the text was rewritten. The Oliver episode illustrates a useful distinction: the problem does not begin when a person asks AI for help writing something. It begins when the person stops being the final editor of his own words. A chatbot can draft, rewrite and organize. The Print button still doesn’t know how to proofread.