Turning a transcript into notes with an AI assistant

Why the raw transcript misleads the model

An assistant has no way to tell a transcript from a document someone wrote. It reads fluent, correctly punctuated prose and reasons over it as fact. The errors a speech model makes are exactly the ones that survive this: a plausible wrong name, a confident wrong number, a looped phrase over a silence.

The result is not a model making things up. It is a model faithfully summarising text that was already wrong, which is harder to catch, because the summary is internally consistent and reads well.

Given the same text and told what it is, the behaviour changes. Names get hedged, figures get quoted as stated rather than asserted, and the model will say a passage is unclear instead of smoothing it into something definite.

The preamble, and why it is not padding

The transcript export here puts a short note in front of the text saying the words were produced by a speech model, and naming the places such models go wrong: proper nouns, acronyms, numbers, overlapping speech, and stretches with no speech. That is the whole mechanism.

It is tempting to shorten this to something like may contain errors. That works much less well, because it tells the model there is risk without telling it where. The specific list is what lets it apply caution to the right words rather than hedging everything equally, which is just a less useful summary.

If you are copying a transcript out of something that does not add this, write it yourself. Two sentences at the top changes the output more than any amount of prompt wording after it.

What to ask for, and what not to

The requests that work treat the transcript as evidence to be quoted. The ones that go wrong ask the model to be certain about details the audio may not support.

• Ask for quotes with timestamps rather than claims. A quoted line can be checked against the recording in seconds; a paraphrase cannot
• Ask what was decided and, separately, what was raised but not resolved. The second list is usually the valuable one and summaries drop it
• Ask it to flag anything that looks like a transcription error, especially names and figures. It is good at spotting a word that does not fit its sentence
• Do not ask who said what unless the transcript has speaker labels. Without them the model infers speakers from context and presents the guess as fact
• Do not ask for figures you intend to rely on. Have it quote the line containing the number so you can check the one thing that matters

The part that is actually an upload

Worth being blunt about, because this page would otherwise quietly undermine the rest of the site. Transcribing here involves no upload: the model runs in your browser and the file never leaves your machine. Pasting the resulting transcript into a hosted assistant uploads it.

The audio stayed put, which is not nothing. But depending on the provider and the plan, the text may be retained, logged, or used for training, and the text is the part with the names and the numbers in it. If the recording was one you could not upload, the transcript is very often one you cannot paste either.

What follows: for a confidential recording, use a model running on your own machine, or an assistant with a contractual no-retention guarantee, or do the summarising yourself. Transcribing locally and then pasting into a free chatbot is the same disclosure with extra steps.