We all know AI hallucinates. They fill in the gaps. Make up citations (the bane of librarians these days are students looking for help finding sources that don’t exist).
But do they outright lie?
Call me naive. But I was surprised.
I was using Claude to analyze open text survey responses and accidentally used a prompt from the week before that was quite specific about the names of columns in the dataset. Imagine there was a column of data that called “Students>88”, representing a flag to identify students with higher than an 88% grade.
I was looking for patterns in responses to see if there were any differences between B+(ish) students and students earning a lower grade.
But the previous week I’d completed a similar analysis using a column called, “B+Stu” and had a crisp prompt that referred to that column title explicitly.
I uploaded date with the “Student>88” column but asked Claude to compare responses between students marked “B+Stu” and the others.
Claude happily chewed on the data and came back with a compelling analysis. Claude used the column title in its analysis, which is how I discovered I’d used an outdated prompt (all of this was in a new chat, unrelated to the previous one).
I wrote to Claude, “but there is no column called “B+Stu” in the data I provided.
Claude immediately apologized, and somewhat disarmingly, explained its error in a very forthright way. But it hadn’t noticed that my very particular reference didn’t match the data I provided…and continued anyway?
Bonus learning: at least with Claude, it is best to upload data in one prompt, then ask your questions in a second prompt. Maybe Claude wouldn’t lie using this approach. Perhaps I’ll create some fake datasets to reverse engineer Claude’s immorality.
