The big AI models are amazing. Give one a stack of interview transcripts and it will tell you, in fluent and confident prose, what people said: the main themes, who felt what, how groups differed. It reads like the work of a good analyst.

But can those claims be pinned down? If the summary says "most of the teachers, especially the women, mentioned workload", can you check it? And would it be true?

Here is what we found while building Rubicon, our tool for answering evaluation questions from interview data.

What that sentence actually needs#

For "most of the teachers, especially the women, mentioned workload" to be true, somebody has to do three things:

• find every place where anybody mentions workload, in every interview;
• know which speakers are teachers, and which of them are women;
• count, and compare the counts with how many teachers and how many women there were.

An AI model reading all the transcripts at once does none of these. It does not count. It remembers the beginning and end of a long read better than the middle, and a vivid quote sticks with it more than a dull one. It writes the kind of sentence that usually follows from material like this, which is not the same as a sentence that is true of this material.

How a confident summary can mislead#

• Vague words that cannot be wrong. "Many", "several", "tended to".
• Numbers nobody counted. "Seven of the twelve teachers."
• The wrong "most". Most of the people who talked about the topic, read as most of everyone.
• A real quote under a false sentence. The quotation is word for word; the claim wrapped around it is not.
• The wrong speaker. Someone describing their colleagues' feelings becomes someone describing their own.
• "Nobody said..." A claim that only a search of every transcript could support.
• Group differences out of thin air. "Especially the women", without knowing how many women there were.
• General knowledge passed off as a finding.
• A different answer next time. Ask again and "most" becomes "many".

What happened when we checked our own tool#

Rubicon is built so that every sentence in its report points to the interview passages it rests on. We used it to redo a published evaluation, then asked a separate reviewer to check our report against all 21 original transcripts, one claim at a time.

The quotations were perfect: all 488 were word for word. The sentences written around them were not. Of 52 claims the reviewer tested, 25 were wrong, and most were about who said something or how many people did. They were exactly the faults in the list above.

We only found them because every sentence could be traced back to its passages. A report without that trail would have had the same faults, and nobody would have seen them.

What we changed, and what to ask of any tool#

• Let arithmetic do the counting. "Workload was raised by 7 of the 9 women and 3 of the 8 men" should come from counting marked passages and knowing who each speaker is, not from a model's impression. Rubicon now writes every "who said it" and "how many" itself, and the model writes only the words around them.
• Ask for the trail. If you cannot go from a claim to the passages behind it, you cannot tell which claims are wrong.
• Check one claim at a time, with the transcripts to hand. A model checking a single sentence, and able to search the interviews, catches much of what a model writing everything at once gets wrong.