How to Summarize a Long YouTube Video with AI, Including When There Are No Captions

Not in the way you do. Almost every AI video summarizer reads the video's transcript, the caption text, and summarizes those words. It does not see the picture. That one fact explains most of the ways summaries go wrong, and it tells you what to do when a video has no captions: there may be nothing to summarize.

The quickest route is a sidebar tool that reads the transcript from the watch page for you. In Copilotly's YouTube summary, you open a watch page, press Alt+C, choose Summarize, and get a TL;DR, key points with timestamps and a few quotes, then ask follow-up questions. This guide covers that workflow, but also the parts other guides skip: what happens with no captions, how to handle a two-hour video, prompts that produce usable notes, and how to check a summary before you trust it.

Why "ChatGPT can't watch YouTube" is a common complaint

A chat model takes text in and writes text out. Pasting a YouTube link into a chat does not guarantee the model can open the page or fetch the captions. Some products can, some cannot, and I have not tested every one, so I will not claim which. The reliable method is to give the model the transcript text itself, or to use a tool that fetches it for you.

What a transcript-based summary can and cannot tell you

• Can: main topics, the order of arguments, steps a speaker lists, numbers they say aloud, quotes.
• Cannot: anything shown but not said (a diagram, a code editor, a product demo), whether the speaker is right, tone, sarcasm, or music and sound.

A tutorial that says "now click this, then that" over a screen recording will summarize badly. A podcast, lecture or talk will summarize well.

This is the verified flow from the YouTube summary feature page.

1. Open the YouTube watch page. You do not need to play the video.
2. Press Alt+C to open the Copilotly sidebar.
3. Choose the Summarize page action. The sidebar shows "Reading the video transcript", then a chip with the video title and its duration, then the summary streams in.
4. Type follow-up questions in the same chat. The transcript is saved with the chat, so later questions still have it.

What the summary contains

• TL;DR: two or three sentences on what the video is about and its main takeaway.
• Key points: five to eight bullets in the order they come up, each starting with a [mm:ss] timestamp.
• Notable quotes: two to four short verbatim quotes with timestamps, left out if nothing is worth quoting.

The model is told to use only what the transcript says and to quietly correct obvious caption slips, such as a misheard name. The side effect is that a slip it cannot spot stays in the summary.

Why follow-ups matter more than the summary

The summary is one message. The value is in asking next: "List the steps as a numbered list with timestamps", "What does the speaker say about pricing?", "What are the three main claims and what evidence is given for each?" Those questions pull detail the TL;DR dropped.

Picking a model

If one model glosses over detail, switch the picker to another in the same chat and ask again. The Claude and Gemini pages describe how each handles long inputs.

If a video has no caption track at all, Copilotly says so instead of guessing. The chip reads "No transcript - the link will be sent instead". The assistant is then told to say plainly that no transcript was available and to give only what the video's title and description support, not to invent what the video says.

That is the exact behavior, and it is deliberately conservative. A summary built from a title and a description is a summary of the marketing blurb, not the video. Do not treat it as one.

Why a video might have no captions

YouTube's automatic captions help page lists reasons automatic captions may not generate, as of November 2026: processing of complex audio is not finished yet, the video's language is not supported, the video is too long, the sound quality is poor or the speech is not recognized, there is a long silence at the start, or multiple speakers overlap or several languages are spoken at once. Live streams have extra limits: automatic captions for live streams are English only and are being rolled out to channels. Uploaders can also add their own captions in YouTube Studio, as described on YouTube's subtitles and captions page, and videos with captions turned off by the uploader fall in the same group as having none.

What to do instead

Checking whether a transcript exists

YouTube's help page on viewing video transcripts says the path is the "Show transcript" control in the video description. Per the search result I saw for that page, the transcript scrolls with the video and selecting a line jumps to that moment. If that control is missing, the video probably has no captions, and no summarizer can read what is not there.

What I have not verified

The link is sent to the model "so it can try to read the video", according to the feature's privacy notes, but the model is also told not to invent content. I have not confirmed that any model can reliably read an uncaptioned video from a link, so do not rely on it. If the sidebar says no transcript, assume you are getting a title-and-description answer.

A two-hour video is longer than Copilotly's transcript limit, so the summary will cover only the first part unless you work in sections. The limit is 45,000 characters of transcript. If the video is longer, the transcript is trimmed, and the chip says "trimmed" so you know the summary covers only the part that was read.

How much video is 45,000 characters?

This is my arithmetic, not a vendor figure. Assume an average of 6 characters per spoken word, counting the space after it. At 120 words per minute that is 720 characters a minute, so 45,000 characters is about 62 minutes. At 150 words a minute it is 900 characters a minute, about 50 minutes. At 180 words a minute it is 1,080 characters a minute, about 42 minutes. So expect the cap to cut in somewhere between 40 and 65 minutes of speech.

A 2-hour video at 150 words a minute is about 108,000 characters, so the tool reads roughly the first 42 percent: 45,000 / 108,000.

Strategies that work

1. Read the trimmed summary honestly. If the chip says trimmed, the later part of the video is not in the chat. The feature guidance is to ask about the early part and watch the end yourself.

2. Ask for specific topics, not "everything". "What does the speaker say about pricing?" beats "summarize everything", because the model searches a smaller target inside what it has.

3. Feed the rest in as text. Open the transcript from the video description ("Show transcript"), copy the second half, and paste it into the chat, or into the text summarizer. Summarize each half separately, then ask for one combined outline. This is a manual workaround, and it works because the model then reads the text you gave it.

4. Use chapters as seams. If the creator added chapters, split at chapter boundaries so no point is cut in half.

5. Summarize in two passes. First pass: section headings only. Second pass: details for the two sections you care about.

A prompt for the second half

The text below is the second half of a talk. The first half covered [topic from the first summary]. Summarize this half as: (1) a three-sentence overview, (2) the main claims in order, (3) any numbers or names mentioned, (4) anything that contradicts the first half. Do not add information that is not in the text.

When the summary is not enough

For a long interview or lecture you plan to quote, treat the summary as a map. Jump to each timestamp and listen to the passage before you repeat it.

Copy these into the sidebar chat after the first summary. Replace the [placeholders].

Lecture notes and study guide

Turn this lecture into study notes. Format: a heading per topic in the order it appears, 2-4 bullets each, then "Key terms" with one-line definitions as the speaker defined them, then 5 practice questions with answers. Put the [mm:ss] timestamp next to each heading. Use only what the transcript says, and mark anything unclear with [check].

Meeting or webinar recording

List the decisions made, the action items with the owner if a name was said aloud, the open questions, and the deadlines mentioned. If an owner or date was not stated, write "not stated". Include timestamps.

Tutorial or how-to

Extract the steps as a numbered list with timestamps. List the tools, settings, versions and prices the speaker names. Then list the steps where the speaker refers to something on screen without describing it, so I know where to watch.

That last line is the fix for the transcript-only limit: ask the model to flag where the speaker says "as you can see here" or "this setting", because that is where you must look at the video yourself.

Interview or opinion video

What are the speaker's three main claims? For each, give the evidence they offer, quote the sentence, and add the timestamp. Then list anything they assert without evidence.

Quick triage

Give me a verdict in three lines: who this video is for, what I would miss if I skipped it, and the one timestamp I should watch if I only have two minutes.

Turn the summary into something else

• "Rewrite these notes as an email to [colleague] with three takeaways and one question."
• "Turn the key points into flashcards: question on one line, answer on the next."
• "Make a checklist I can tick off while following the tutorial."

For text you want polished after the summary, select it and use rewrite anywhere. To go deeper on a claim from the video, run a web search and compare. The guide on AI tools by profession has more workflows.

The feature page does not specify how non-English videos are handled, so I will say what is known and what to test. Copilotly reads the caption track from the watch page, and that includes auto-generated captions. YouTube's help page says automatic captions are available in many languages, for example English, Spanish, French, German, Japanese and Hindi, and that for videos with multiple audio tracks captions are generated in the default language only (as of November 2026).

What to try

1. Run Summarize and read the output language. The model may answer in the video's language or in yours.
2. Ask for the language you want. "Answer in English, but keep names and quotes in the original language." Models generally follow this, but check.
3. Quote check. Ask for the key quote in the original language plus a translation, then compare the original to what you hear.
4. Translate the surrounding page. The description and comments can be read with page translation, which shows the translation under each original block. The feature does not translate subtitles on the video itself.

Where it breaks

• Automatic captions are weakest on accents, dialects, background noise, mixed-language speech and technical terms, per YouTube's help page.
• Names and places get transliterated inconsistently.
• If a language is not supported for automatic captions, there may be no captions, which puts you back in the no-captions case above.

For anything important, such as a legal, medical or immigration video, use the summary to find the passage, then get the original checked by a qualified human.

A summary is as accurate as the captions beneath it, and it can still add errors of its own. YouTube says automatic captions come from machine learning, that quality may vary, and that they can misrepresent speech because of mispronunciations, accents, dialects or background noise (as of November 2026). The summary then inherits those mistakes.

I have not measured an accuracy percentage for any tool and will not invent one. Here is a five-minute check you can run on any summary instead.

The 5-point spot check

1. Timestamp test. Pick three key points. Open each timestamp in the video and listen for 30 seconds. Does the speaker actually say it there? If two of three are off, distrust the whole summary.
2. Number test. Find every figure, date and name in the summary. Check each against the video. Numbers are where captions and summaries fail most.
3. Quote test. Play each notable quote. Does the wording match? Captions may differ from what was said.
4. Omission test. Watch the last five minutes yourself, especially if the chip says trimmed. Conclusions and caveats often sit at the end.
5. Negation test. Look for "not", "unless", "except" and "only". Summaries sometimes flatten a qualified claim into a clean one.

Red flags

• A summary with perfect, evenly spaced timestamps and no detail, which suggests a thin transcript.
• Details about what is "shown" or "demonstrated", which a transcript-only model cannot know.
• Confident statements about a video that said "no transcript" in the chip.

The ask-for-evidence prompt

For each key point, quote the exact sentence from the transcript that supports it, with the timestamp. If you cannot find supporting text, say "not in transcript".

This forces the model to point at its source, and a missing quote tells you where to look. A confident summary of a flawed argument is still a summary of a flawed argument, so for claims you plan to repeat, verify them elsewhere, for example with web search.

There is no single free answer, because it depends on what you need.

The no-account fallback

Open the transcript on YouTube, copy it, and paste it into the text summarizer. You lose automatic timestamps and the chat, but it needs no account and no extension.

Where a dedicated tool may suit you better

Dedicated video summarizers can have extras Copilotly does not: panels built into the YouTube page, export to notes apps, and summaries of whole playlists in one go. Copilotly is an assistant that happens to summarize videos, so the summary lives in a chat. That is better for follow-ups and worse for bulk work. I have not compared output quality with specific tools and will not claim one is better.

Paid plans

Pro is $6.99 a week, $19.99 a month or $149.99 a year, with a 7-day trial for first-time subscribers (see pricing). If you only summarize a few videos a week, start with Free and see whether the daily limit ever bites.

What is read and what is sent

On a watch page, Copilotly reads the caption track and basic details: the video title, channel, length and the first part of the description, using your own browser session, without playing the video. The transcript, title and description are included in the message to the model that writes the summary, together with your question. If there is no transcript, the link is sent instead.

Which videos to avoid

Do not use any summarizer on private or unlisted videos you are not allowed to share with an AI provider, such as a company meeting recording or a client call. Treat the transcript as text you are sending to a third party.

Copyright, in plain terms

This is practical guidance, not legal advice, and the rules differ by country. A summary for your own study or triage is a different thing from republishing it. If you publish notes, write them in your own words, credit and link the creator, and keep quotes short. Copying large amounts of a creator's wording into a post or a course is risky whatever tool produced it. If it matters commercially, ask the creator or a lawyer.

If you are the creator

Add accurate captions in YouTube Studio. It helps viewers who need them, and it also means summarizers have something accurate to read.

1. Before you watch: run Summarize, read the TL;DR and key points, and decide whether the video deserves your time. Note the timestamp of the one part you most need.
2. While you watch: jump to that timestamp. Ask the chat about terms you do not know.
3. After you watch: ask for notes in your format, then compare them with what you saw. Fix anything the transcript misheard.
4. Before you share: run the 5-point spot check above and cite the video, not the summary.

For a pasted transcript, the text summarizer does the same job without the extension, and the model comparison helps you choose which model to ask when detail matters.