How I’d Evaluate AI-Generated College Recommendation Letters

What one factual error taught me about designing GenAI systems that people can review and trust.

In my last article , I shared about Why Human-in-the-Loop Review Matters in GenAI

The follow up question is how to setup a evaluation to help system that we generated reports is accurate.

We refined the prompt after reviewing the drafts. From a solution architecture perspective, that experience raised a question I’d now address at the start of a project like this: How will we know whether the system preserves the meaning of the source records?

Test the claim against the evidence

I’d start with a small set of anonymized student records and ask teachers or counselors to verify the facts. For each achievement, I’d record the activity, the student’s role, the level or affiliation, and the source passage that supports it.

Then I’d break each generated letter into factual claims and check them individually. Here’s how that would work for the chess and cheerleading example.