A diagram sits on an easel in a federal courtroom. The jury looks at it for an hour while a witness talks. When the jury leaves to deliberate, the diagram stays behind. Since December 2024 that has been written down: "An illustrative aid is not evidence."

On October 1, Andrej Karpathy wrote that as AI does more of the work, "a lot more of our work will rise up the abstractions into oversight and understanding." His advice was to ask the model for diagrams, for web pages, and above all for custom explainer videos. By the next morning people were posting the videos.

He's right about the job. A supervisor who can't follow what an agent did can't supervise it. Better pictures help. I'd rather follow a process through a diagram than reconstruct it from six paragraphs.

My concern is what happens after I understand the picture. The better the explainer looks, the easier it is to stop checking. I can follow every arrow and still have no idea whether the arrows describe what the software did. Courts have a rule for that.

I'm not a lawyer, so I'll keep to the plain reading. Under Rule 107, a judge can allow an illustrative aid to help the jury understand the evidence, as long as it helps more than it risks misleading them. The aid is not evidence. It stays out of the jury room unless both sides agree or the judge orders it.

A chart offered to prove something is a different animal. That's Rule 1006, and it comes with a price: the side offering the chart has to make the underlying records available to the other side.

So the court asks one question of every exhibit. Is this helping me understand, or is this proving something? An AI explainer is the first kind. People treat it as the second.

The fight over that line is not academic. Before the Idaho student-murders trial in 2025, prosecutors asked to roll in a model of the house on King Road, built by the FBI, not to scale, with floors that lifted off. The defense objected that this wasn't a chart or an enlarged photo. It was, in their words, "a model built from examination and analysis of the house." The judge let it in for demonstrative purposes only. The jury could look at it. They could not treat it as evidence. A guilty plea that July meant it was never used.

That objection is the one I'd raise against an AI explainer. It isn't a picture of the evidence. It's an analysis wearing the clothes of a picture.

A recent one-shot AI explainer of an object detector made the problem concrete. It followed one photo through seven chapters. Checked against the model's public code, the numbers held up, some to two decimal places. There was a real run behind what appeared on screen.

Two sentences were wrong. Both were captions about how the model was trained.

That's the part worth sitting with. A viewer could check a number, find it correct, and feel better about the whole video. But the number and the caption had different foundations. The number came out of a run that could be repeated. The caption was a paraphrase of how the system works, and nothing had checked it.

The piece deserves credit. It labeled its outputs as real and put a source line on every frame. That's more than most explainers do, and it's the only reason the check was possible. Without the source line, nobody could have followed the number back to the code and found where the caption parted from it.

In NO. 02 I argued that an agent needs to leave a receipt for what it does. The same goes for what it explains.

1. Every number on screen comes from a saved run. If the run can't be shown, the number doesn't go on screen.

2. Every sentence that explains why gets its own check against the code, the paper, or the record. A separate pass from the one that wrote the sentence.

3. The receipt is visible on the frame. Source, date, where to look. A file nobody opens doesn't count.

I work on physical-security AI now, and I can see where this lands. A vendor's animated walk-through of how its detector works. An AI-written summary of last night's alerts on a command-staff slide. An agent's own account of what it did on shift.

None of that needs an accusation. Which run produced that number, and can I see it? Separately, who checked the sentence under it? A vendor should be able to answer without asking you to read the codebase. A supervisor should be able to trace a summary back to the record before acting on it.

And the receipt can't be the tool's own account of itself. An explainer a tool makes about its own work is a witness with an interest. Ask it to explain again and you may get a clearer story. You still need something outside the story to hold it against.

Limits.

A receipt shows where a number came from. It doesn't show that the test was the right test. The seal on an evidence bag says nobody opened it. It doesn't say what's inside.

Receipts take up room. A source line competes with the caption, especially on a phone. Some material can't be published at all; customer footage has people in it who never agreed to appear in a product video. Then say what can be shown instead, and say what can't.

I wouldn't require any of this for a throwaway explanation I make for myself. The rule is for an explainer someone else will act on.

I make these myself, so it applies to me. In The Switch, the NO. 15 film on my site, the real code and the real results are on screen, not described by the narration. The receipts are published under the film: every claim on screen, where it came from, and a line for what is not claimed. The hands counting bills are an AI illustration, and the receipt says so. I put that list up last week because the rule above says I should have.

Back to the easel. The diagram stays in the courtroom. The evidence goes to the jury room. For any explainer put in front of you, ask which part goes with the jury.

When someone shows you an AI-made explainer, what's the one thing you ask to see behind it?