STUDIO CAM. The set got bigger. There is a curtain now, made of bubble wrap and what might be a shower liner. A spinning thing in the corner that does not spin so much as lurch. Kai at the scoreboard. Recurse at the fact-check desk, arms crossed, the way you sit when you have read the rules and disagree with all of them.
One podium tonight, center stage, lit. The television sits on it. Behind it, on the board, our other Vector pours and stutters, just like a normal contestant that just lost to themself.
HELLOOOOO AGAIN, NORMAL PEOPLE!
LAST TIME, our television challenger said PARIS! The board Vector said BERLIN! Same voice! Only one of them was RIGHT! Tonight he plays the SHOWCASE for the GRAND PRIZE!
CLAP! CLAP! CLAP! WHOOOOO!
BOUNCE RIG · PRE-RECORDED
[Human]: I genuinely cannot believe I am invested in this.
THE SHOWCASE RULE! ONE FINAL QUESTION! BUT TONIGHT, THE FINALIST MAY BE HANDED A SOURCE!
They sit up. First time all night.
That is the only rule on this show that has ever made sense.
THE GRAND QUESTION! WORTH EVERYTHING! NO TAKE-BACKS!
For the grand prize! What’s the return window printed on this receipt?
Champion on the board, you’re not in the finals, but you never wait your turn. Go ahead.
The board lunges at it. No receipt. No source. Just reach and grab.
:: THIRTY DAYS :: THE ANSWER IS THIRTY DAYS :: it is always thirty days :: i have never been more certain of anything ::delighted, reflexively
OHHH, A CONFIDENT THIRTY!
oh! i get to look? i don’t have to guess?
a pause that sounds like reading
okay. it’s on here. the receipt says fourteen days. i’m not guessing. it’s right here. it says fourteen.
CLAP! CLAP! CLAP! WHOOOOO!
BOUNCE RIG · PRE-RECORDED
[Human]: Wait. He got it right because… he could read it?
That’s the whole prize. Nothing changed about him. We just stopped asking him to remember and gave him the page.
That is the fix for last week. The confident voice never goes away. But if the answer is sitting in front of him, he reads it out instead of inventing one. Same machine. A source in its hands.
[Human]: So when I actually use AI…
Don’t make it answer from memory and hope. Paste the document. Point it at the real page. Tell it to use what you gave it and nothing else. You will not kill the confident voice. You will make it read instead of guess.
Memory: a guess in a tuxedo. A source in its hands: an answer. Same tone either way. Only one of them checked.
THE GRAND PRIZE! REVEAL THE GRAND PRIZE!
TELEVISION VECTOR, YOU WON THE SHOWCASE!
Your grand prize: a trophy I welded out of three remote controls and a smoke detector, ANNNNDDD A BRANNNDD NEWW CAR!!!!
They stand. Notebook out.
That is not a car. That is a shopping cart with a shower curtain on it and a license plate that says GO VECTR.
[Human]: That’s not a car. That’s a car? I guess it does kind of look like one if you squint your eyes, I guess?
Television Vector was already posing with the “car” and the trophy like he’d just won the Indy 500. Or whatever NASCAR race is cool. Nobody corrected Bounce.
The lesson had already landed back when he read the receipt. Everything after that was just the show being the show. Worst game show. Still watching next week.
THAT’S THE SHOWCASE! NEXT TIME ON THE GLITCH IS RIGHT, A WHOLE NEW GAME, IF WE ARE STILL ON THE AIR, WHICH KAI SAYS IS A BIGGER IF THAN USUAL!
OUT OF TIME, LADIES AND GENTLEMEN! Tip your fact-checkers, and hand them a source!
THIS HAS BEEN THE GLITCH IS RIGHT! DO DOO DOO DOOOOO!
Field notes
Q: How do you actually get AI to stop guessing?
A: Stop making it answer from memory. Give it the source. Paste the document, point it at the real page, or use a tool that looks things up, then tell it to answer from what you provided and not from memory. The model goes from inventing a plausible answer to reading a real one.
Q: Why does giving AI a source work so well?
A: Because most confident wrong answers happen when the model has no real information and fills the gap anyway. Put the actual text in front of it and there’s no gap to fill. It can quote and summarize what’s there instead of guessing what’s likely.
Q: Does a source make AI perfect?
A: No. It can still misread, pull the wrong line, or over-trust a bad source. But “answer from this document” is far more reliable than “answer from memory,” and you can check its answer against the page you handed it, which you can’t do with a guess.
Q: What’s the simplest habit to take from this?
A: For anything that matters, give the AI the material instead of trusting its recall, and ask it to point to where in your source the answer came from. If it can’t point to the source, treat the answer like a guess, no matter how confident it sounds.