When an AI writes code or a document for you, do you really check it less — or just check it later, outside the chat?

This explores whether the drop in scrutiny after an AI produces an artifact is real, or whether people just move their checking out of the chat window, by running the code, editing the document or showing it to a colleague. The short answer is that the corpus can't settle it, but it explains why the question is so hard to answer from conversation logs. It also gives a strong competing explanation.

The starting observation comes from an analysis of 9,830 Claude conversations Do artifact outputs reduce how critically users evaluate them?. When users are building an artifact, they put more effort in up front: they clarify goals and iterate before it's produced. Afterwards they are less likely to fact-check it or question its reasoning. The catch is that this was measured from transcripts. Anything a user does after closing the tab is invisible to that kind of study.

This is where a formal result from a different area becomes useful. Work on AI self-reflection proves that a judge who reads only the generated text cannot reliably tell whether a reflection helped when the truth depends on something outside that text. Judges that can check the actual environment succeed Can transcript alone tell whether a reflection helps?. The same logic applies to researchers studying users. If verification happens in a compiler, a spreadsheet or a meeting, two very different users can leave identical transcripts: one who checked carefully offline and one who never checked. The scrutiny drop could be partly an artifact of where the researchers were looking. Research on AI judges points the same way: evaluators that go out and collect evidence are far more stable than ones that only read outputs Can agents evaluate AI outputs more reliably than language models?.

The corpus also gives good reasons to doubt that offline checking explains everything. Polish itself seems to switch off judgment. Evaluators have rated AI-written documents as both human-written and better than real human submissions Does polished writing actually signal better quality work?. Models trained to imitate ChatGPT's confident, fluent style fooled human raters without getting any more accurate Can imitating ChatGPT fool evaluators into thinking models improved?. The model side may make this worse. Preference training rewards confident answers over clarifying questions and checks on understanding, cutting those moves to 77.5% below human levels Does preference optimization harm conversational understanding?. So a finished-looking artifact arrives with fewer of the signals that would prompt anyone to question it.

The takeaway is that both explanations are plausible, and transcript data alone can't separate them. Doing so would take studies that follow what happens to an artifact after the conversation, such as whether code gets tested or documents get revised. The corpus doesn't contain that kind of follow-up research yet, so treat the 'scrutiny drop' as an upper bound on how much users stop checking, not a confirmed measurement.