Between 14 September and 19 October 2026, AIOS Scribe ran 380 transcription jobs. 256 completed and 124 failed. A third of all attempts failing is a lot, and the reasons are more instructive than the headline. Some of the failures were the AI's doing, some were ours, and one was a platform limit nobody could see because we had not turned on the logs. This post goes through all of them, because if you depend on AI transcripts you should know the ways they stop working.
A note on the sample: the jobs include many we ran ourselves to test the service, including deliberately difficult videos. The failure rate here is not what an ordinary visitor will see on an ordinary video.
The 124 failures, by recorded reason
| Reason | Jobs | What it means |
|---|---|---|
| Timed out | 51 | The job started and never reported back. Many were long videos (see below), but a failed job stores no length, so we cannot say how many. |
| Failed, no specific category | 30 | Anything we do not have a more precise label for. A visitor may retry these. |
| No reason recorded | 27 | We did not record a reason for these. |
| Too long | 15 | The video is longer than the service accepts. This one is permanent for that video. |
| Busy | 1 | The job waited too long for its turn and was dropped. |
Until 16 September a failed job told the visitor only “Transcription failed”, and the real reason sat in logs we could not search. Recording a reason on every failed job was the first fix, because you cannot improve what you cannot see.
Failure 1: long videos that died silently
From 6 to 18 October, 39 of the 51 jobs for videos longer than about 15 minutes failed, against 36 of 37 under 15 minutes that completed. Each long job was picked up within seconds, did about 45 cents of paid AI work, and then simply stopped, without saving a result or recording an error. Because they never finished, the stuck-job rule marked them timed out about 15 minutes later.
The cause was a limit we had not known to look for. A long video is transcribed in one-minute parts, and each part makes outside calls, so a long video makes many of them. The platform limits how many outside calls one job may make, and once a job passes it every later call fails, including the call that would have saved the failure. That is why it was silent. A controlled test with a 73-minute video, watched live, confirmed it. We raised the limit and turned logging on, and refused videos over 15 minutes up front while we fixed it properly: a long video is now transcribed as many small jobs instead of one, so no single job has to make all those calls. On 20 October, 15-minute, 58-minute and 104-minute test videos all completed, and the limit went back to two hours.
The lesson is not specific to us: a system that fails without writing anything down looks, from the outside, exactly like a slow one. If a long upload to any transcription service never returns, the failure may be on the service side, and the only honest fix is better visibility.
Failure 2: the AI refusing to transcribe famous talks
The AI sometimes declines a request and returns a refusal reason instead of text. On 16 September five well-known talks failed that way in production. We tested it directly. The original request asked for the transcript twice: once as a single block of text and once as timed segments. The refusals came from the single block. On the one talk we tried repeatedly, six of seven attempts with that prompt were refused. Asking for the timed segments only returned a full transcript every time in three attempts across two of those talks.
That is a small experiment, and the honest conclusion is limited: refusals are not deterministic. One talk was refused in production and then transcribed on a later attempt with the same prompt. So we retry a refusal automatically, and we never record a refusal as permanent for a video, because that would lock out a video the next attempt might transcribe. Only “too long” is permanent.
Failure 3: busy servers and a bad length estimate
On 15 and 16 September the main model answered “high demand” to several requests, including a fallback model. Moving to a paid account tier fixed the same video that had failed repeatedly on the free tier, which finished in 104 seconds right after the upgrade. Free-tier requests are the first to be dropped under load. Separately, the AI's own estimate of a video's length was often wrong, for example 97 seconds for a 16-minute lesson, so we now read the true length from YouTube instead.
What this means if you rely on transcripts
- If a transcript fails, look at the stated reason. “Too long” will not change; most others are worth one retry.
- Shorter videos fail less. Splitting a long recording yourself is a reasonable workaround.
- A transcript that completes can still be flagged. See what the automated reviews found.
- Do not assume that silence from any service means it is still working.
The limits we apply and why are listed in how AIOS Scribe works, and each change is dated on the what's new page.
Need a transcript to try this on?
Paste a public YouTube link and get a full transcription with timestamps. Free within fair-use limits, with a free account.
Try AIOS Scribe