The AI summary of my own notes keeps lying, and I finally saw why
A summary is not a fact. It is a compression, and compressions drop the part that would have saved me.
I trusted the one-line summary my notes app generated for a meeting, then acted on it and was wrong. The summary was not false so much as hollow: it kept the conclusion and dropped the caveat that changed the conclusion.
That is what summaries do. They trade detail for length, and the detail they drop is usually the exception, the condition, the 'unless.' Those are exactly the parts that matter when you decide something.
So I stopped reading the summary as the note. I read it as a signpost: 'this file is about X,' not 'this file says X is true.' Then I open the source before I act. The summary earns the click; it does not earn the decision.
The fix is not a better model. It is a habit. Any claim I will act on gets traced to the line it came from. If the summary and the source disagree, the source wins, because the summary was built by throwing things away.
An AI that summarizes your own notes is a great librarian and a poor witness. Treat it like the first, not the second.
Frequently asked questions:
Q: Are summaries useless then?
A: For recall, no. For decisions, yes, until you check the source.
Q: Why does it drop the important part?
A: Compression keeps the common thread and sheds the exception, which is often the load-bearing detail.
Q: How do I use summaries safely?
A: As a pointer to the file, not a substitute for reading it before you act.