Skip to content

Which of your AI loops could look fine while failing?

Answer two questions for each of your loops (up to two more details are optional); the map places it and gives you two steps for it.

  • Two questions per loop, up to two optional details
  • No account
  • Nothing you type leaves your browser

Built for the essay Your AI Looks Best Where You Can Check It Least

The map

Your loops against two questions
Your loopsType one, or paste a list, one per line.

Nothing you type leaves this page

No upload and no account. Our host counts the visit like any web page; what you enter it never sees.
Your map saves to this browser. A link you copy carries it after the #, the part of an address no browser sends to a server. The checklist, the ticket text and the reminder carry your loop names and named checks to wherever you put them; the image carries your loop names only if you switch them on.

The two questions, and what each answer means

1Who checks that it worked? (the essay’s test for “could the signal be faked?”)
2If it gets one wrong, can you undo the damage? (the essay’s “is the first failure terminal?”)

If something outside the loop checks it, the page asks what checks it, so the check can be tested. “What does it do?” only picks the examples in its steps.

Who checks that it worked?Something independent: something the loop did not choose and cannot influence checks its work, even if it checks only a sample.Partly the loop itself: part of the check comes from the loop itself, such as its own report or its choice of what gets checked, or from a score it is tuned to.Only the loop itself: the only evidence it worked is the loop’s own output, or a score it is tuned to.If it gets one wrong, can you undo the damage?Yes, fully: you can undo the damage and go again.Partly: some of the damage can be undone and some cannot.No: the damage cannot be undone, or the failure removes your ability to fix it.

“Partly” counts toward the corner: a check that is partly the loop’s own can still be gamed, and damage that is partly permanent still costs something you cannot get back.

Higher up: harder to undo

checked independentlychecks itself

A loop is any job AI does again and again without a person checking each result. Most you can safely leave running and keep improving. A few you cannot: they check themselves, at least partly, and a real failure could not be fully undone. Those sit in the Potemkin corner, where a loop can look fine while it fails.

Your loops, ranked

closest to the corner first
More ways to save or share

The tool link carries no map: send it to someone who should make their own.

A loop that checks itself can look fine while it fails.

The essays and tools on checking AI work come with sources shown. Subscribing is free.

Subscribe opens Substack in a new tab with your address. Nothing you entered in the tool goes with it.

Pairs the essay Your AI Looks Best Where You Can Check It Least on The Durability Curve. The two questions and the two kinds of fix are the essay's; this runs them on your own loops.