When you read a candidate's AI transcript alongside their submission, the pattern you'd expect from strong engineers is a certain amount of mess. Wrong turns. Half-formed questions. A point where they ask about the wrong thing for ten minutes before realising. That kind of texture is what thinking in public usually looks like.

Which is why it's slightly unsettling when you open a transcript and everything reads like a tutorial. Five clean questions, five clean answers, no backtracking. It looks excellent at first glance. Look harder. It's often one of the weaker signals in a submission.

Why clean transcripts are suspicious

Real engineering thinking doesn't happen in five structured questions. It happens in twenty messy ones, most of which get abandoned halfway. When you're reasoning about a non-trivial problem with an AI, you typically go down the wrong path at least once, notice, double back, and ask a better question. That shape is visible in the transcript, and it's a good sign.

A perfectly structured transcript usually means one of two things. Either the candidate wasn't really reasoning, and asked the AI the most obvious version of the question and accepted whatever came back. Or they were crafting the transcript for the reader, treating it as a second polished artefact alongside their written answer. Both are concerning. The first suggests shallow engagement with the problem. The second suggests a candidate who's optimising for appearance, which is a different skill from the one you're hiring for.

There's a third possibility worth naming. Sometimes a clean transcript really is just a candidate who knew exactly what they needed and asked efficient questions. This does happen with senior engineers on problems well within their wheelhouse. If you see it, the rest of the submission should independently confirm that kind of depth.

What to look for when the transcript looks tidy

Length of each turn. A clean transcript with long candidate questions and long AI responses is usually the polished-for-appearance kind. The candidate spent time drafting each message. Compare this to a natural transcript, where some turns are one-liners and others are paragraphs.

Revision in the candidate's questions. Genuine thinking produces questions that reference previous answers, correct earlier misunderstandings, and build on each other. A transcript where each question stands alone, with no reference to the thread before it, often means the candidate was treating the AI like a search engine, not a collaborator.

Moments of disagreement. The strongest transcripts almost always include at least one exchange where the candidate pushes back on the AI. "That doesn't handle the case where…" or "I don't think that's quite right because…" A transcript with zero friction is one where either the candidate didn't notice the AI's mistakes or the AI didn't make any, and the second is statistically unlikely.

Abandoned threads. Did the candidate start asking about one approach, abandon it, and start over on a different one? This is a strong positive signal. They updated their mental model as they went. A transcript that proceeds in a single straight line from question one to the final answer is less convincing than one that changes direction mid-way.

How to bring it up in the follow-up

If you suspect a transcript was polished for appearance, the follow-up conversation is where you find out. Don't ask "did you really write these messages?" That's an accusation and it doesn't help you. Instead, ask a question about a specific turn: "In your third message you asked about the trade-offs between X and Y. What made you think about X first?"

A candidate who was really reasoning will be able to reconstruct their thinking in the moment. A candidate who was crafting an artefact will either hesitate or give a generic answer. The difference is usually obvious within thirty seconds.

This isn't about catching the candidate out. It's about confirming that the depth you saw in the transcript matches the depth they actually have. If it does, the interview gets richer. If it doesn't, you've learned something useful before you made an offer.

The healthier framing for candidates

A lot of what the "clean transcript" problem comes down to is that candidates don't know two things: that the reader actually wants the mess, and that they don't need to funnel every thought through the AI to produce a good submission. The instructions need to make both expectations explicit. In CriticCode, the candidate is told up front to use the AI however they would at work, that holding back to seem more certain doesn't help them, and that the interviewer cares about their reasoning more than the polish of the final answer.

A clean transcript isn't automatically a fake one. It's a prompt to read more carefully and ask a sharper follow-up question. That's usually the most useful response to anything suspicious in a submission.