The videos that fall through
The gap is not random. Transcript databases are built by crawling what is popular, recent and English. The videos that get missed are the ones people most often need a transcript for:
- University and conference lectures. A 2013 physics lecture with 40,000 views has captions on YouTube and no reason for a transcript service to have crawled it.
- Small channels. Below some threshold of views, nobody indexed it.
- Non-English uploads. Coverage drops sharply outside English, even where YouTube has full captions.
- Auto-caption-only videos. Some services index only creator-uploaded captions and skip machine-generated ones entirely — which is most of YouTube.
- Anything recent. Uploaded yesterday means crawled by nobody.
None of these have anything to do with whether the transcript exists. They are facts about someone else’s index.
Why the server gets refused even when the video is indexed
There is a second failure underneath the first. YouTube treats a request from a data centre differently from a request from a person browsing. Tools that fetch captions server-side run into rate limits, consent walls and outright blocks — the open-source libraries most of these products are built on ship named error codes for exactly this.
The symptom is confusing, because it does not look like a block. It looks like an empty result. The tool finds a caption track, gets handed something that is not captions, and reports that the transcript could not be read.
Check the video once, by hand, just to settle it. If the transcript is there and the tool says it is not, the problem is on the tool’s side — either it never indexed the video, or it was refused. If the panel is empty, the video genuinely has no captions, and no product on the market can read it. Anything that returns a summary anyway is describing the title, the description, or its own guess.
What to do about it
The fix is structural, not a workaround: use something that goes to the video instead of querying an index. YouTube Conversation is a free ChatGPT plugin that does exactly that. You paste the link and ask your question — the video opens in your own Chrome, on your own connection, and the transcript is read from the video itself. You do not open panels, click through menus, or copy anything. There is no database to have missed your video, and no server address for YouTube to turn away.
That also changes what a failure means. When it cannot read the real transcript, it tells you. It is not allowed to fall back to a transcript service, a search result, a summary site, or its own re-transcription of the audio. Every other tool in this category will answer anyway. This one is not permitted to — so an answer you get is an answer from the video.
The honest limit
If the video has no captions at all, this does not help you either. A private video, a removed video, a region-blocked video, or an upload the creator never captioned and YouTube never auto-captioned: there is nothing to read, and the plugin will say so rather than invent something. Tools that promise summaries “with or without subtitles” are re-transcribing the audio on their servers, which is a different product with a different failure mode — and it is worth knowing which one you are using.
For what ChatGPT itself can and cannot do with a video, see can ChatGPT watch or listen to YouTube videos. For the failure where it answers without reading anything at all, see why ChatGPT can’t watch YouTube videos.