Does dictating an essay with speech-to-text and then cleaning it up look like AI to detectors?

I dictate first drafts for a living. Detectors do not know whether you spoke or typed. They see the final text and nothing else. So the useful question is what the text looks like at the end, not how it got there.

What I do now: I keep the raw transcript as its own file, untouched, with the date in the name. Then every edit pass is a new version. If a client ever questions a piece, I can show the ugly transcript, the tidier second pass, and the final one. Nobody fakes a messy transcript with filler words in it.

audio person here. dictated text gets flagged way more than typed text, basically always. speech is simple and repetitive and the auto punctuation puts commas where a language model would, so to a detector a transcript looks like machine output. honestly i would tell him to stop dictating for anything that gets scored.

Rob, I am sorry, he does not deserve that. Two practical things from the school side.

If the wrist injury is documented, ask the school for it to be recorded as an accommodation, in writing. Then the teacher has a reason on file for why this term's work sounds different from last term's, and a score has context.

Second, ask the teacher to look at the argument rather than the number. "It reads very smooth" is not evidence of anything. A smooth essay from a kid who revises is exactly what we are supposed to want. Offer to have him talk through the essay for five minutes. A boy who dictated it can do that without trying.

@Jules I would be careful with "basically always". I am not aware of anything showing that dictated text is flagged more often than typed text, and the mechanism you describe is a guess. Detectors score the final text, and they are known to misfire on plain, uniform, formulaic prose whoever wrote it, which can include a tidy human essay. What does seem plausible is that heavy automated rewriting afterwards changes how uniform the text is. That is a hypothesis for Rob to test with his own drafts, not a rule he should plan around. Our thread on detector accuracy covers why single scores should be held loosely.

Which cleanup step did he use? That matters more than the dictation. A grammar checker that fixes a comma is one thing. One that offers to rewrite whole sentences, with the student accepting most of the suggestions, can change the text more than people expect, and the result can be much more uniform than how a person writes.

I would compare the raw transcript with the final version line by line. If most sentences were rewritten by suggestion, that is a good guess at where the score came from. If only commas and names changed, it points elsewhere. The thread on false flags has a good list of what evidence to gather.

I dictate in English to practise my pronunciation and then edit, so I worry about this too. One of my essays was flagged last term. What helped me was that my document kept every change, and my advisor could scroll back through the history and see the messy first pass turn into the final one.

So Rob, I would also save each stage somewhere, not just the final copy. A flag with no process behind it is a number. A flag with the whole process behind it is much easier to talk about.

@Priya N. that is a very good question and I went and looked. He accepted a lot of the grammar checker's suggestions, probably most of them in the last three paragraphs because he was tired. The transcript has his own rambling phrasing and the final version is much smoother. So that may well be it. I have saved both versions and will keep every stage from now on.

Treat it like a chain of custody. Raw audio if the software keeps it, raw transcript, each edited version with timestamps, and a note of which grammar tool was on. Cheap to collect, annoying to reconstruct later.

And do not feed the essay to three more detectors to "check". Each extra score is another number the teacher can wave around, and they will disagree with each other, which proves nothing either way.

@Hana Sato fair, I was overstating it. I do not have a study, just a lot of transcripts that look weirdly tidy once the software has punctuated them. I will stop saying "always".