Yes, often, but how well depends on the assistant and on the video. Gemini can take a public YouTube link and explain it, and accepts uploaded video files (a few minutes in total on a free account, longer on paid plans). ChatGPT accepts short uploaded clips but warns it may not analyse the whole video, and it does not reliably read a pasted YouTube link. Claude does not take video files at all. NotebookLM imports a YouTube video's transcript, not its pictures. The method that works with every assistant is to copy the video's transcript, paste it in, and ask for numbered steps. Whatever you use, anything shown on screen but never said aloud is the part most likely to be missed.

What each assistant actually does with a video

This is the state of things in late 2026. These features change often, so if something here no longer matches, trust the assistant's own help page.

Two things follow from that table. First, for most YouTube links the AI is reading the transcript, not watching. A video with no captions, or auto-captions that mangle the speaker's accent, gives a weak answer however good the model is. Second, when an assistant does look at frames, it samples them at intervals rather than reading every frame. A formula written on a board for two seconds can fall between samples.

The transcript method, which works everywhere

If you want a reliable step-by-step explanation, skip the guesswork about which tool can see what.

1. On the YouTube page, open the description and press Show transcript. Copy it. (On a lecture capture system or Coursera, there is usually a transcript or captions download.)
2. Paste it into whichever assistant you like, with the prompt below.
3. Separately, pause on any scene where something important is on screen but not spoken (a diagram, a worked equation, code) and screenshot it. Paste the screenshots too, or type what they show.

A prompt that gets steps rather than a vague summary:

Here is the transcript of a [length] video about [topic]. Explain it step by step for a [level] student. Number every step. For each step give: what the speaker says happens, why it happens, and the timestamp where it starts. If the transcript refers to something shown on screen that is not described in words, say "on screen, not in transcript" instead of guessing. List any term the speaker uses without defining it.

The line about "on screen, not in transcript" is the important one. Without it, assistants fill gaps with plausible content, and plausible content in a step-by-step explanation is exactly what you cannot tell apart from the real thing.

Worked example: a 14-minute lecture on the Krebs cycle

Say you have a recorded biology lecture on the Krebs cycle (also called the citric acid cycle) and the exam is next week. You paste the transcript with the prompt above. A good answer comes back looking like this:

1. Pyruvate becomes acetyl CoA (2:10). Before the cycle, each pyruvate from glycolysis loses a carbon as carbon dioxide and joins coenzyme A. This happens in the mitochondrial matrix.
2. Acetyl CoA joins oxaloacetate (4:05). The two-carbon acetyl group combines with four-carbon oxaloacetate to make six-carbon citrate. That is where "citric acid cycle" comes from.
3. Citrate loses two carbons (5:40). Over the next reactions, two carbon dioxide molecules are released and NADH is made.
4. Energy carriers are produced (8:20). Each turn makes 3 NADH, 1 FADH2 and 1 ATP (or GTP, depending on the cell).
5. Oxaloacetate is regenerated (10:50). The cycle ends where it started, ready for the next acetyl CoA.
6. Two turns per glucose (12:30). One glucose gives two pyruvates, so the cycle runs twice.
7. On screen, not in transcript (6:00 to 7:30). The lecturer points at a diagram and says "this one, then this one". The transcript gives no names here.

Step 7 is the honest answer, and it is the reason to screenshot the diagram. Now check the rest against your textbook. In explanations like this, the usual slips are counting NADH per glucose instead of per turn (it is 6 per glucose, 3 per turn), putting the cycle in the cytoplasm instead of the mitochondrial matrix, and quietly merging the link reaction (step 1) into the cycle itself. An AI explanation is a draft of your understanding, not a replacement for it.

Where AI explanations of videos go wrong

Knowing the failure modes saves more time than switching tools.

• Names and numbers from auto-captions. "Krebs" becomes "crabs", "Bazalgette" becomes nonsense. The assistant then explains the nonsense confidently.
• Visual-only steps. Maths lectures, coding tutorials and anything with "as you can see here" lose the most. Screenshot those moments.
• Very long videos. A two-hour recording is possible in some tools, but the step list gets coarser. Ask for one section at a time ("steps between 20:00 and 40:00").
• Invented steps. If a step has no timestamp, or a timestamp that lands on something unrelated, treat it as invented.
• The video was wrong. An assistant explains what the speaker said, including mistakes. A step-by-step breakdown of a bad video is still a bad video.

If you are using all this to revise, remember that reading an explanation of a video is still passive. How to study from video instead of rereading notes covers the parts that make it stick: pausing, predicting the next step, and testing yourself.

The flip side: turning the explanation into your own video

Once you have a clean, checked step list, you are closer to making a video than you might think. A numbered explanation with one idea per step is almost exactly a scene list. This is useful in two situations: you need to present the topic to a class, or you want a short revision video in your own words that you can rewatch.

Two rules keep it yours and keep it fair to the original creator:

• Rewrite in your own words. Do not paste the lecturer's transcript in as narration, and never re-upload or re-cut someone else's video. Use the step list as notes and write the explanation yourself.
• Credit the source. One line at the end: the lecture or channel name and the link.

Each step becomes a scene: two to four sentences of narration and one drawing that shows the step. For the Krebs cycle, that is a mitochondrion with the matrix shaded, a two-carbon piece joining a four-carbon ring, carbon dioxide leaving at two points, and the energy carriers counted at the end. Seven steps at about 25 seconds each is a three-minute video, roughly 450 words at 150 spoken words a minute.

Make it

1. 1

   Get the transcript and the screenshots

   Copy the transcript from Show transcript under the video. Screenshot every moment where something important is on screen but not said aloud.

2. 2

   Ask for numbered steps with timestamps

   Use the prompt above with any assistant. Insist on 'on screen, not in transcript' for gaps instead of guesses.

3. 3

   Check every step against a real source

   Textbook, class notes or the slides. Fix names that auto-captions mangled and delete any step without a believable timestamp.

4. 4

   Rewrite the steps as your own narration

   Two to four sentences per step, in your words, with one line saying what each scene should show. Add a credit line for the original video.

5. 5

   Paste it into openCanviz

   Choose Keep my wording so the narration is exactly what you checked, set a three-minute target and pick whiteboard, which suits processes. Fix any wrong label scene by scene, and replace the generated voice with your own recording if your teacher wants to hear you.

Common questions

Can AI explain a video that has no captions? Only an assistant that processes the audio or frames of an uploaded file, such as Gemini or ChatGPT with a short clip. For a YouTube link with no captions, most tools will fail or guess. If you have the file and the right to use it, upload it, or transcribe it first with a speech-to-text tool.

Can it explain a private or unlisted video? Mostly no. Gemini's YouTube features and NotebookLM work with public videos. For a private class recording, use the transcript your school's video platform provides.

Is it allowed to use AI to explain a lecture? For your own understanding, almost always. For graded work, follow your school's AI policy, and do not hand in the AI's explanation as yours. The step list is notes; the thinking you put on paper afterwards is the work.

Why did the assistant explain a different video? It probably could not read the link and answered from the title or from similar videos it knows about. Ask it to quote the first two sentences of the transcript. If it cannot, it never saw the video.

Can ChatGPT make a video from the explanation? Not by itself. It can turn the steps into a script, but not into a narrated, drawn video. Can ChatGPT make a video? explains what it can and cannot output.