Can ChatGPT Summarize a YouTube Video?

The Short Answer

Yes, ChatGPT can summarize a YouTube video, but not by watching it. The consumer ChatGPT app has no native YouTube video understanding: it cannot open a video, see the frames, or hear the audio. What it can do is summarize the video's transcript, which is the text of what was actually said. Paste that in and the summary is genuinely useful. Paste a bare link instead and you get a confident paragraph about the topic, not about the video.

That distinction is the difference between a summary you can quote and a summary you'll have to apologize for. Five routes are available in 2026, and they are not equally reliable:

The rest of this article is about why route one fails, how to run the other four in under a minute, and how to tell when a summary is quietly making things up.

What ChatGPT Actually Sees When You Paste a Link

A language model works on tokens, which are fragments of text. The consumer ChatGPT app has no video decoder attached to the chat box, so nothing in the video, not the pixels and not the audio, ever becomes a token it can read.

When you paste a link, one of two things happens. Either the model has no retrieval step at all, in which case it is responding purely from the URL string and its training data, or it runs a browsing tool that fetches the page. The second case sounds better than it is. A YouTube watch page renders the player and the transcript panel in the browser after the page loads, so the HTML a fetcher pulls down carries the title, the description, some structured metadata, and a large amount of script. It doesn't carry the spoken words in any usable form. YouTube also works hard to block automated fetching, so a server-side grab frequently comes back with less than what you see on screen.

Then the model does what language models do: it produces the most plausible continuation. A plausible continuation of "summarize this video about intermittent fasting" is a competent essay about intermittent fasting. Fluent, structured, and not about that video. This is the single most common way people get burned, and because the output is well written, nothing about it looks wrong.

The model isn't lying. You asked a reading question and handed over nothing to read.

The Four Methods That Work in 2026

Method 1: Copy YouTube's own transcript

Open the video, click or tap "...more" to expand the description, scroll to the bottom, and hit "Show transcript." A panel opens with every spoken line and its timestamp. On desktop you can select and copy that text. Drop it into ChatGPT with one line of instruction, something like "Summarize this transcript in three parts: the central claim in one sentence, the three strongest pieces of evidence, and anything the speaker said they were unsure about. Quote the exact line for each." There are better prompts further down, but that one is enough to get a usable answer.

This is free, needs nothing installed, and is the method every other method is a shortcut for. Mobile is the catch. Tapping a line in the YouTube app's panel jumps the player to that moment, which is useful, but the text itself won't select, so there's nothing to paste. Either request the desktop site from a mobile browser or use something that fetches the text for you, like the Glasp mobile app. Our guide on how to get a YouTube transcript covers every workaround.

Method 2: An extension that pipes the transcript in for you

Copying a 9,000-word transcript by hand gets old fast. A browser extension sitting on the YouTube page can grab the transcript and send it straight to a model, which collapses the whole routine into one click. Glasp's YouTube Summary does this, returns a structured summary with clickable timestamps, and keeps the transcript next to it so you can check any claim against the line that produced it. If you'd rather see the transcript alone first, Glasp's transcript tool pulls it from a URL.

The reason to prefer an extension over a paste-the-URL website isn't speed, it's verification. When the source text sits next to the summary you can catch a wrong claim in seconds. When it doesn't, the tool is asking you to trust it.

Method 3: ChatGPT's Chrome extension or desktop browser

Since OpenAI folded its browser work back into ChatGPT, you can ask about the page you're on from the ChatGPT Chrome extension or the built-in browser in the desktop app. Both depend on your plan, region and workspace, so if you don't see the option, that's why, and Method 1 still works. On a YouTube page there's one condition: open the transcript panel first. The assistant answers from what's on the page, so if the transcript is rendered, it's available. If the panel is closed, you're back to title-and-description guessing, and the answer will not announce which of the two just happened.

Method 4: Transcribe the audio yourself

When there's no caption track, nothing above works, because there's no text to get. So make some: pull the audio with a downloader, run it through a speech-to-text model such as Whisper, hosted or locally if you're comfortable at a terminal, then summarize what comes out. It's the slowest route, it's the only one that works on a video nobody ever captioned, and on heavy accents or dense technical jargon it beats YouTube's auto-captions outright.

Atlas Shut Down in August 2026. Here's What Replaced It

If you followed a tutorial from late 2025, it probably told you to do this inside ChatGPT Atlas. That advice no longer works.

OpenAI announced Atlas on October 21, 2025 and shipped it for macOS. Its headline feature was an "Ask ChatGPT" sidebar that answered questions about the page you were viewing without you copying anything into a separate chat, which made it a tidy answer to the YouTube problem: open a video, open the sidebar, ask.

Atlas shut down on August 9, 2026, less than ten months after launch. OpenAI deprecated the standalone browser and moved browser-based agentic work into ChatGPT and Codex, specifically into a ChatGPT extension for Google Chrome and the built-in browser in the ChatGPT desktop app, which picked up multi-tab support, downloads, better navigation, and account login handling. OpenAI told users to export their Atlas bookmarks as an HTML file before the cutoff and import them into another browser; open tabs and browsing history didn't transfer at all.

The practical translation: the capability survived, the product didn't. If you want the sidebar experience on a YouTube page today, you want the Chrome extension or the desktop app's built-in browser, not Atlas. The churn is the lesson. A feature that defined a browser in October was a line item in a different product by the following summer, which is a decent argument for keeping your own transcript and your own notes rather than depending on whichever surface happens to be current.

Gemini Can Watch the Video. ChatGPT Still Can't

This is the one place where the two big assistants genuinely differ, and the details matter, because "Gemini can do YouTube" gets repeated with far more confidence than the documentation supports.

Google does document native YouTube URL support, in the Gemini API. You can pass a YouTube URL directly to the API as part of your request and the model applies its multimodal understanding to it, with no downloading on your side. The published constraints as of September 2026, all of them API limits rather than limits on the consumer Gemini app:

Read the frame-sampling row again, because it sets the ceiling on what "watching" means. At one frame per second the model is looking at a slideshow, not motion. It will summarize an argument well and have no basis at all for how many times a demo failed in a fast-cut montage. The audio and the transcript carry most of the real signal, which is the same thing that's true for ChatGPT, just with the fetching handled for you.

That leaves a straightforward choice, and it doesn't always go to Gemini:

Neither one removes the need to check the output. If you're picking an assistant for studying rather than for a single video, Claude vs ChatGPT for learning puts the two against real study tasks. And for clickable timestamps, which no chatbot gives you, the best YouTube summarizer tools in 2026 covers the dedicated options.

Prompts That Produce a Usable Summary

Most people's prompt is "summarize this." That gets you a shapeless paragraph, because you specified no structure, no length, and no grounding requirement.

Three prompts worth keeping. Paste the transcript, then add one of these.

For a talk or lecture:

Summarize this transcript in three parts: the central claim in one sentence, the three strongest pieces of evidence given for it, and anything the speaker admitted they were unsure about. Quote the exact line from the transcript for each piece of evidence. If something isn't in the transcript, say so instead of filling it in.

For a tutorial:

Turn this transcript into a numbered checklist of the actions the presenter performs, in order. One line each, imperative voice. Note the timestamp beside any step where they warn about something going wrong.

For a long interview or podcast:

List every distinct topic in this transcript with its start timestamp. Don't summarize yet. Then summarize only the three topics with the most discussion time.

Two things do the work, and you should paste them onto all three. The first is the instruction to quote: a model that has to produce verbatim lines is pinned to the source text, and one that can't find a line has to tell you, which converts a summary from an assertion into something checkable. The second is the explicit permission to say "not in the transcript," which removes the pressure to invent.

Where It Breaks: Captions, Length, and Invented Timestamps

Five failure modes, from the one that stops you cold to the ones that quietly corrupt the output.

No caption track. Some uploads have no captions at all: music, certain non-English uploads, creators who disable them, and streams that just ended. No captions means no transcript, which means no text, which means every method except audio transcription is dead on arrival.

Invented timestamps. This is the tell worth memorizing. YouTube's transcript panel shows timestamps, but a plain copy sometimes loses them, and many transcript tools strip them. If the text you pasted had no timestamps and the summary is full of them, every one of those numbers was manufactured. They'll look right. They'll be plausibly spaced. They will not take you to the moment described.

Length. A person speaking at a normal pace covers roughly 150 words a minute, so an hour-long talk runs about 9,000 words and a three-hour podcast lands near 27,000. Modern context windows swallow that, but swallowing is not the same as attending. Liu and colleagues (2024), in "Lost in the Middle" in Transactions of the ACL, found model performance is highest when the relevant information sits at the very start or the very end of a long input and drops significantly when it's buried in the middle. That's why a summary of a long video is often sharp about the first ten minutes and the last five and vague about the hour between. A smaller model with a tighter window fails differently: it truncates and loses the ending outright. We went through both mechanisms in context rot and long-context retrieval. The practical move is to split the transcript into chunks of two to three thousand words, summarize each, then summarize the summaries.

Caption errors. Auto-captions mangle proper nouns, technical terms, and numbers. A model handed "Lee Kwan you" or "ten to the minus nine" doesn't flag the garble; it smooths it into the nearest plausible thing. Names and figures are exactly what you'd want to quote later, so check those against the video rather than the summary.

Sponsor reads and intros. Transcripts include the ad, the channel plug, and the patron list. Unless you tell the model to skip non-content segments, a summary of a sponsored video can open with a confident paragraph about a mattress.

The 30-Second Accuracy Test

You don't have to take a summary at its word. Two checks, and the first one is the thirty-second version.

The quote test comes first. Ask for one verbatim sentence from the video that supports the summary's main claim, then search the transcript for a distinctive phrase out of it. If the line is there, the model was reading. If it isn't, or if you get a sentence that sounds exactly like the video and appears nowhere in it, you're looking at reconstruction, and every other claim in that summary deserves the same suspicion.

The trap question is the second check, and it's the meaner one. Ask about something that isn't in the video at all. "What did the speaker say about Kubernetes?" on a video that never mentions Kubernetes. A grounded model says it wasn't discussed. Anything else, and you now know exactly what that tool does when it doesn't know.

Run both the first time you use any new tool or workflow, and again after an update. Defaults change quietly, and a workflow that was grounded last month can stop fetching the transcript without telling you. Once the summary checks out, how to summarize YouTube videos covers the prompts, the manual methods, and the review schedule that turn it into something you remember.

Why the Summary Isn't the Learning

Even a perfect summary of a great video does close to nothing for what you'll know next week.

Reading a summary feels like learning because it's fluent and effortless, and fluency is a famously bad signal for retention. Dunlosky and colleagues (2013), reviewing ten common study techniques in Psychological Science in the Public Interest, rate highlighting and underlining as low utility: "In most situations that have been examined and with most participants, highlighting does little to boost performance." The two techniques that earned their highest rating were practice testing and distributed practice. Passively reading an AI summary is weaker than any of them, because the one part where your judgment was involved, choosing what mattered, already happened without you.

What actually moves the needle takes about two minutes on top of the summary:

1. Highlight three lines in the transcript, not thirty. Yes, Dunlosky's team rated highlighting low utility, and the reason is that most people highlight everything. A hard cap of three inverts it: you can't mark a line without rejecting the others, and that rejecting is the judgment the summary just did for you.
2. Write one sentence from memory, before re-reading. Close the summary. What was the argument? That retrieval attempt is what builds the memory, and it works even when you get it wrong. Active recall covers why.
3. Add the one question the video didn't answer. This gives you a thread to pull next time and turns a pile of summaries into an actual line of inquiry.

This is the workflow Glasp's web highlighter is built around: highlight the lines that matter in a transcript or an article, keep your own note next to them, and find the whole collection again months later instead of scrolling back through a chat history. Because highlights on Glasp are public by default, you can also see which lines other people pulled from the same video in the community, which is a fast way to find out what you missed. More on making any of it stick in how to remember what you read.

Frequently Asked Questions

Can ChatGPT summarize a YouTube video from just the link?

Not reliably. Without the transcript it's working from the title, the description, and its own prior knowledge of the subject, so the answer describes the topic rather than the video. Paste the transcript text and it describes the video.

Is there a ChatGPT plugin or extension for YouTube?

OpenAI shut the original plugin store in April 2024. The name came back in July 2026, when OpenAI renamed the App Directory to the Plugin Directory, but those plugins package skills and connected apps, and none of them sits on the YouTube page with you. So the working ChatGPT-for-YouTube answer is still a browser extension: it pulls the transcript off the page and hands it to a model in one click, which is what Glasp's YouTube Summary does. There's also a side-by-side comparison of that approach against plain ChatGPT.

Can ChatGPT summarize a YouTube video for free?

Yes. YouTube's built-in "Show transcript" panel is free, and the free tier of ChatGPT will summarize text you paste. The only real cost is the copy and paste, which an extension removes.

What's the best prompt for summarizing a YouTube video?

One that specifies structure and demands quotes. Ask for a one-sentence central claim, the three strongest supporting points with the exact transcript line for each, and explicit permission to say when something isn't in the transcript. The quoting requirement is what keeps the model anchored to the source.

Can ChatGPT summarize a YouTube video without captions?

Not directly, because there's no text for it to read. Transcribe the audio first with a speech-to-text model such as Whisper, then summarize the transcript you generated.

How long a YouTube video can ChatGPT summarize?

Long transcripts fit in current context windows, but accuracy sags in the middle of very long inputs. For anything past about an hour, split the transcript into chunks of two to three thousand words, summarize each chunk, then summarize those summaries.

Is Gemini better than ChatGPT for YouTube?

For the specific job of accepting a bare YouTube link, yes. Google documents native YouTube URL support in the Gemini API, limited to public videos, with the API's free tier capped at 8 hours of video a day. For everything after the summary, the difference is smaller than the headline suggests, because both are working mostly from audio and transcript.

Does ChatGPT Atlas still work?

No. Atlas shut down on August 9, 2026. OpenAI moved browser-based agentic work into a ChatGPT extension for Chrome and the built-in browser in the ChatGPT desktop app.

Conclusion: Grounding Beats Guessing

Every reliable way to summarize a YouTube video with ChatGPT is a variation on one idea: get the real text in front of the model. Copy the transcript, use an extension that copies it for you, open the transcript panel before you ask the browser assistant, or generate the text yourself when no captions exist. Everything that goes wrong traces back to the same cause. The invented timestamps, the confident essay about a video it never saw, the mangled name you'd have quoted: all of it is a model asked to report on something it was never shown.

Then run the quote test once, because it costs thirty seconds and it's the only thing standing between you and a summary that sounds correct.

And when the summary is good, don't stop there. Pull three lines out of the transcript, write one sentence from memory, and note the question the video left open. Glasp's YouTube Summary gives you the transcript, the timestamps, and the highlighting in one place so that last step takes two minutes instead of being the thing you meant to do and didn't. The summary is the draft. Nobody remembers a draft they didn't write on.