Why I Stopped Trusting AI Search Summaries and Started Checking the Link

The summary was confident, cited a source, and was wrong about the one number I cared about. That was the third time in a month, so I changed the habit. The AI search result is a lead now, not an answer, and I treat the citation as something to open, not something to trust.

The failure mode is subtle. These tools do not usually invent sources from nothing; they compress, and compression drops the condition. A study of five hundred people becomes everyone, last year's figure becomes current, a finding about one group becomes a rule for all. The summary reads clean because the qualifier got edited out in the name of brevity.

I now do one thing before quoting any AI summary: click the source and read the exact sentence. If the tool cannot show me the sentence, I do not use the figure. It costs thirty seconds and has saved me from three wrong statistics that I would have put in front of a client with a straight face.

This is not anti-AI posturing. The summaries are genuinely useful for finding where to look. The mistake is letting the middleman become the authority. The model is a fast librarian with a habit of paraphrasing; you are still the one who has to read the book.