AI transcripts are a strong first draft, not a verified record. This page lists where AIOS Scribe is known to go wrong, how to spot it, and what to do. Where we have seen a problem in our own testing, it says so and gives the date; where we have not tested a category, it says that too rather than guess.
Problems we have actually seen
| Problem | What happens | What to do |
|---|---|---|
| Songs and music | The model sometimes declines copyrighted songs. When it does transcribe one, lyrics can be missing: on a 3 minute 25 second pop song the first version left 40 to 50 second stretches with about ten words (fixed in October 2026 by sending short sections). | Retry once. For lyrics, compare against the video; treat the result as incomplete. |
| Well-known talks | Some famous talks were declined as “recitation” in September 2026, then transcribed on a later attempt of the same video. | Try again in a few minutes. A refusal is not final. |
| Timestamps | The model's clock can be off in either direction: about 1.6 times fast on two talks, but short on another (an 844-second talk ended at 715 seconds). In two songs it also switched notation partway, so later lines showed as 0:01 to 0:03. A clock that runs fast and the notation switch are corrected; a clock that runs slow is not, and every corrected time is an estimate. | Click the timestamp and listen before you quote a time. |
| Lines at the one-minute cuts | Long videos are processed in one-minute sections. A short repeated line heard on both sides of a cut can appear twice. | Delete the duplicate if you are editing the transcript. |
| Video length | Videos that YouTube reports as over two hours are refused. The service can also pause new videos when its monthly capacity is reached. | Transcribe long recordings in parts, or try again later. |
| Private and unlisted videos | They cannot be read at all. | Use a public video. |
Categories we have not measured
Names and spelling, technical terms, accents, overlapping speakers, background noise, code read aloud, abbreviations, numbers and punctuation are the usual weak points of speech-to-text in general. We have not yet tested AIOS Scribe on each of them in a controlled way, so we give no figures and no examples for them here. The benchmark plan describes how we intend to.
How to check a transcript quickly
- Search the transcript for names, numbers and technical terms that matter, and confirm each against the video.
- Click the timestamp of any line you will quote, and listen to that moment.
- Look for stretches with far fewer words than the speech in the video would produce. That is the sign of a gap.
- Do not publish, cite or submit a transcript as exact without this check.
If you find a video that fails or an error pattern we should know about, tell us on the contact page with the link.
Frequently asked questions
- Is a transcript from AIOS Scribe exact?
- No. It is an AI draft. Check names, numbers and any line you will quote against the video.
- Why was my video declined?
- The AI model sometimes declines copyrighted music and very well-known talks. The same video can succeed on a later attempt, so try again.