When a research tool tells you "users find onboarding confusing," ask one thing before you act on it. Did a user say that, or did the model?
Those two carry completely different weight, and most tools show them the same way. Same font, same confidence, same place on the page. So a model's paraphrase gets treated as if a person had said it, and a decision gets made on something nobody told you.
The tool's own words look like findings
The obvious case is the AI summary. It's the model's reading of the interview, and quoting it is quoting the model. Fine when it's shown as the model's view. A problem when it's shown as the user's.
The case people miss is the interviewer. When an AI runs the interview, roughly half the transcript is the AI. It asks questions, restates what it heard, offers a framing for the person to react to. "So it sounds like the export step is where you get stuck" will turn up in any search for export problems. But the participant didn't say it. The interviewer did, and the participant may have said "yeah, I guess" to keep things moving.
If those lines can be cited, the evidence fills up with the tool's own questions. It finds what it asked about, then reports it back as something users told it.
The test
Can you get from a claim back to a person saying it, in their words, in the interview it came from?
If yes, it's evidence. If the trail ends at a summary, or a line the interviewer said, or something a support rep typed into a ticket, it's an interpretation. Maybe a good one. Still not a user. This applies to a human researcher's deck too, where the quotes, if any, sit in an appendix nobody opens. The AI version is just faster and has more layers.
What it forces on the tool
You can't get this with a style guide, because the exceptions are exactly the ones that end up in the deck. It has to be how the data is stored. Every line carries who produced it, and the code that builds a citation refuses anything the participant didn't say. Summaries and recommendations are a different kind of object from quotes, shown differently, so you always know which you're reading.
What you give up
Some convenience. The model is often right about what someone meant, and the tidy paraphrase reads better than the real quote with its false starts. You don't get to use the tidy version as evidence.
Findings also get thinner. Some things you strongly suspect can't be shown, because nobody quite said them. The tool can say it thinks the pattern is there. It can't hand you a quote that proves it.
I think that's the right trade, because the alternative is a system where no layer is just the users. Every layer is the model reading the one below it. It looks like research. It's mostly the tool talking to itself.
In AskEngine, quotes are set in a different typeface from everything else, and each one opens the interview it came from. If you can't trace a claim back to a person, it isn't there.