§ 05 Handover

Handover: what to give AI, and what to keep.

A short video series for academics. Each episode hands one real part of the job to AI and shows what came back, how long it took, and where it fell short. The brief I used comes with every episode, so you can try it on your own work.

New episodes every Tuesday. Subscribe on YouTube to get them.

A Episodes

Newest first.

Episode 01 · 17 November 20265 min
01 / Handover

Last year's slides

I had AI check every study in my own lecture. It flagged 41 claims.

If you teach, you'll know how updating a lecture usually goes. You open last year's slides, change the date on the first one, and that's the lecture updated.

So I handed one of my own fourth-year lectures to an AI, along with the chapter that goes with it, and asked it to look up every study. Does it exist, have I described it properly, and has it held up since? About 40 minutes later it came back with 41 things worth a look, each with a link.

The brief I used (copy it, and change the parts in brackets)

I've attached the slides from one of my lectures: [a first-year social psychology lecture, last updated in 2011]. I want to bring it up to date before I teach it again.

Who it's for. [First-year undergraduates. It's an introductory survey lecture, so some claims are simplified on purpose.]

What I'd like you to do. Go through every slide. For each study, statistic or empirical claim, look it up and check three things:

  1. Does the study exist, and is it cited correctly (authors and year)?
  2. Is it described accurately (the design, the numbers, what was found)?
  3. Has it held up since? Look for later replications, meta-analyses, corrections, retractions, and newer figures that replace old ones.

Look things up rather than relying on memory.

What to give back.

  • My slides, with a note on each slide that needs attention. The note should say what the issue is and link to the source.
  • A one-page summary: how many claims you checked, how many are fine, and a table of the ones that need attention, most important first. Every row needs a link I can open.
  • A list of anything you couldn't verify.

What not to touch. Don't rewrite my slides, change my argument or delete anything. What stays in the lecture is my call. Some classic studies are worth teaching even if they haven't held up, so just tell me what has changed. Don't flag simplifications that are reasonable for [first years].

Before you tell me you're finished.

  • Open every source you cite and make sure it says what you say it says.
  • Sort each flag into one of four kinds: details wrong, hasn't held up, contested, or out of date.
  • Tell me how confident you are in each one.

A deck of more than 100 slides needs to be attached as two files. Claude reads the pictures in PDFs of up to 100 pages, and only the text in longer ones.

It can't open everything. In my run 30 pages wouldn't load or were behind a paywall, and about a quarter of the claims were checked against an abstract only. Open the link and read the paper before you change a slide.

On a personal account, Settings > Privacy > "Help improve our AI models" controls whether your chats are used for training.

This is a personal project, and the views in it are mine. The animation, music and sound were made with Claude, Remotion and ElevenLabs. The narration is me.