How sure?
A small exploration of what actually makes an AI feel trustworthy. It is rarely more confidence. It is honesty about what it does not know.
An AI settles a bet about goldfish. It gets the answer right, then piles on extra facts, and not all of them hold up. Here are five ways the interface could show you which parts to trust, and what each one makes you do next.
Do goldfish really only have a 3-second memory?
Loser buys tacos. They asked an AI to settle it. Watch what changes as the interface gets more honest.
Ben wins. The old “3-second memory” line is a myth. Goldfish remember things for months, can learn to recognize their owner, and can even recognize themselves in a mirror.
- The answer is right, so everything that follows gets trusted.
- Nobody clocks that “recognize themselves in a mirror” is riding the coattails of the correct answer with no evidence behind it.
- Ben and Ana walk away repeating a shaky fact, because a correct answer carried it in.
When the headline is right, the extras ride along for free. The interface never flagged which parts to doubt.
If a confidence display doesn’t change what the user does next, it’s decoration.
Good eye if you spotted it, the archerfish link is meant to be there. That study is real and peer-reviewed, it’s just a different fish than the claim names. The mismatch is the point of that state: an AI often has a source, just not one that actually backs up what it said. Catching that gap is the whole skill this demo is about.