On this page
You missed Tuesday's physics class, or you watched it half-asleep, so you paste the recording into an AI tool and get a neat one-page summary. It looks complete. But what did it leave out?
An AI summary is a useful map of a lecture, not a safe replacement for it. Here is why, and a 20-minute way to check one.
Can you trust an AI summary of a lecture video?
Most tools build a lecture summary in two steps. First, speech-to-text turns the audio into a transcript. Then a language model shortens that transcript. Errors can enter at both steps.
At the first step, a 2024 study of OpenAI's Whisper speech-to-text service, as it worked in 2023, found that roughly 1% of audio transcriptions contained whole phrases or sentences that were not in the audio at all. A long class produces many sentences.
At the second step, researchers Uwe Peters and Benjamin Chin-Yee compared 4,900 summaries from 10 language models with the original scientific texts. Most models produced broader generalisations than the originals, even when prompted for accuracy, and three of them (DeepSeek, ChatGPT-4o and LLaMA 3.3 70B) did so in 26 to 73% of cases. In a direct comparison, the AI summaries were nearly five times more likely than human-written ones to contain broad generalisations. The authors suggest a likely cause: summaries can omit the details that limit a conclusion.
Length is a third issue. When language models had to find one relevant fact in a long input, they often did best when it sat at the beginning or end, and performance dropped when it was in the middle. That is a good reason to spot-check the middle of a 90-minute class, not just the start.
And a transcript only holds what was said aloud. If your teacher wrote a derivation on the board while saying "so this becomes this", the summary has nothing to work from.
What do AI lecture summaries usually miss?
In science and maths classes, watch for five kinds of gap:
Dropped steps. A five-line derivation becomes one line with the final formula.
Lost conditions. "Only for small angles" or "assuming no friction" disappears, which is the kind of overgeneralising the study above describes.
Board-only content. Diagrams, sign conventions and units written but not spoken.
Squashed worked examples. The teacher's solved problem shrinks to its answer.
Warnings. "Students always forget to convert to radians here" rarely survives.
How do you check an AI summary against the video?
Ask for a summary you can check. A prompt like this helps:
"Summarise this lecture as numbered points. Give the timestamp for each point. List every worked example separately with all its steps. List every condition or assumption the teacher states. Mark anything you are unsure about."
Then run this check, which takes about 20 minutes for an hour-long class:
Open three timestamps at random, including one from the middle of the class, and confirm the point is really said there. Timestamps can be wrong too.
For every formula, go to its timestamp and count the steps on the board. Write in any the summary skipped.
Next to each result, write its condition in the margin: when does it apply, and when does it not?
For each worked example, pause before the teacher solves it, try it yourself, then compare.
Cross out anything you cannot find in the video.
What does a checked summary look like?
Take the simple pendulum. Say the summary reads: "Time period of a pendulum: T = 2π√(L/g)." You jump to that timestamp and find the teacher actually said more. The restoring force is −mg sin θ, and only for small angles can sin θ be replaced by θ (in radians), which gives simple harmonic motion and the formula. The teacher then worked an example: for L = 1 m and g = 9.8 m/s², T = 2π√(1/9.8) ≈ 2.0 s.
Your checked notes now hold the formula, its condition, the step behind it and a solved example. The summary held only the formula.
Why keep the teacher's worked examples?
The worked example is where the method lives: which equation to pick, what to substitute and where students slip. Research on learning from worked examples compared with solving from scratch explains why studying full solutions helps beginners. A summary that keeps only answers throws that away.
It also changes what you do. Reading a summary feels like progress, but it can turn into the kind of offloading covered in what studies found about leaning on AI. Pausing, predicting and testing yourself, as in how to learn from video lectures, keeps the thinking with you.
If a step still will not make sense after you rewatch it, the Fermi Problem Library has worked problems with a step-by-step solver that explains the reasoning rather than printing an answer. And when a whole topic feels shaky, Fermi's AI tutor works out what you already know and builds a plan for the session, so you start from your own gaps.
Tonight, take one summary you already have, open three of its timestamps and see what it dropped.
