When the AI is wrong, and when it says no

If you read one post in this series, read this one. Everything else I have written is about what works. This is about the edges.

When it was wrong

A point of state law. While I was researching purchasing options, Claude told me with full confidence that a new state law did not cover services. It did. That language had been added during the legislative process. I caught it because I had read the bill. If I had not, the error would have gone into a presentation for City leadership.

When I pointed it out, it said plainly that it had been wrong and fixed the work. I still told it, later, that the mistake had cost it some of my trust. It acknowledged that too.

Assuming a decision. In a draft reply to a resident, it wrote as if we had chosen a provider and a purchasing method. We had done neither.

A wrong first guess. I asked about a scrape on my car. It guessed paint transfer. It was bare plastic, which is a completely different repair.

A rule it misstated, then corrected when I uploaded our actual policy.

The pattern: it is most often wrong where it is filling a gap with something that sounds likely.

When it corrected itself

Not every catch was mine. Claude once went back over a report it had written for me, unprompted, and corrected twelve of its own claims. It removed one comparison entirely because newer data no longer supported it. Its own estimate got smaller as a result.

When it said no

• I asked for a famous actor to appear in a joke movie trailer. It declined, because that would show a real person agreeing to something he had not. When I said he was fine with it, it held its position, politely.
• I asked for a jacket logo using a well-known rocket company's name and a job title that does not exist. It declined that too and offered an original design.
• I asked it to restart my home router. It would not enter the router's access code, and gave me the steps to do it myself.

What I take from this

Check anything that matters against the source. Tell it when it is wrong. And notice that an assistant that will tell you no is one whose yes means more.

---

Part of the series 27 ways I actually use AI. Claude drafted this post from the record of our work together.