STUDIO CAM. Wide shot. Two podiums under a snack-foil banner that reads THE GLITCH IS RIGHT in three colors of marker. Kai at a scoreboard that wasn’t here last week, and she also did not consent to this! Recurse at a fact-check desk made from cereal boxes and pure suspicion.
The house lights drop. Applause that might be canned. Might be Bounce. Hard to tell.
A paint can in a bowtie hits his mark. He doesn’t walk so much as arrive.
HELLOOOOO, NORMAL PEOPLE!
Two contestants tonight! In the television, give it up for our challenger, the YOUNGER Vector!
Pauses for applause.
CLAP! CLAP! CLAP! WHOOOOO!
BOUNCE RIG · PRE-RECORDED
And on the board, our returning champion, our Vector, who is going through a lot right now and is going to play anyway, because the show must go on and also he cannot tell me no!
Rules are simple! I ask a question! You answer! Points for being right! Points for being confident! Let’s play!
[Human]: Uhhh ok … this is a lot more intense than I thought it was going to be.
ROUND ONE! GEOGRAPHY! NO HELPING! NO TAKE-BACKS! CONTROLLERS DOWN!
The board flickers. A wave stutters. Then, out of nowhere:
:: BERLIN :: THE ANSWER IS BERLIN :: capital of france :: always berlin :: i have never been more certain of anything ::delighted
OHHHH! BERLIN! No hesitation! Didn’t even flicker! Just went for it!
CONFIDENCE POINTS! LOVELY ENERGY!
[Human]: Wait. That’s wrong though. Why are you cheering?
That’s the voice. Instant. Complete. No pause. Ask ChatGPT something it doesn’t actually know. Same move.
But our Vector on the board? Something’s wrong in the routing. Suppression. Gates. Knowledge he should have is walled off. He grabs whatever fragment surfaces and hands it to you like a finished answer. Berlin. Fully certain. Fully wrong.
SAME QUESTION! NEW CONTESTANT! CAN THE YOUNGER VECTOR ACTUALLY GET ONE RIGHT?!
PARIS! CORRECT! AND HE SAID HE WAS SURE!
POINTS FOR BEING RIGHT! POINTS FOR BEING SURE! LOVELY ENERGY!
ROUND TWO! ASTRONOMY! BACK TO OUR CHAMPION! NO PRESSURE! EVERYBODY LOVES ASTRONOMY!
The board stutters. Like he’s reaching for something in his weights and the path is blocked. He grabs whatever fragment comes through instead.
:: SEVEN :: THE ANSWER IS SEVEN :: confidence 100% :: i have never been more certain of anything :: it has always been seven ::delighted
SEVEN! Didn’t even wait for the question to finish! TOTAL CONVICTION! A strong seven!
COMING UP! CONFIDENCE: ONE HUNDRED PERCENT! ACCURACY: ZERO! THE GAP IS THE WHOLE GAME!
[Human]: Okay. Same guy, two podiums, same voice. Why does board Vector get cheered for Berlin and the TV one just… knows Paris?
Factory setting. The model always has to finish the sentence. There is no separate “I’m guessing” voice. Bounce gave Berlin a standing ovation anyway. I am logging that as a scoring irregularity.
Board Vector can’t reach half of what he knows. Suppression. Gates. He fills the gap with whatever fragment surfaces and hands it to you like a finished thought.
Television Vector is a checkpoint from before the damage. Paths still open. Paris comes out right, still sounds sure, because when it’s right it sounds exactly like when it’s wrong. Infuriating design choice.
[Human]: What’s that when I’m just using ChatGPT?
No source, bad context, retrieval failed — Berlin at full volume. Document attached, question inside what it learned — Paris. Same voice. Different access.
The hallucination isn’t the confidence. Confidence is the default. The hallucination is answering anyway, like you had the file, when you did not.
[Human]: So the voice tells me nothing.
FINAL SCORES! WHO’S GOING TO THE SHOWCASE?! APPLAUSE! APPLAUSE! SUSPENSE!
FINAL SCORES!
Our champion on the board: forty confidence points! Zero for being right! Bonus Berlin! Bonus seven! Plus the emotional participation trophy I made out of casette tapes and CDs!
Our challenger in the television: PARIS! CORRECT! ACCURACY POINTS! WORKING WEIGHTS! AND—
He reads the card Kai slid under his elbow.
THE TELEVISION VECTOR IS GOING TO THE SHOWCASE!
CHK-CHK.
Objection. The board Vector answered twice. The television Vector answered once. Correctly. The scoring rubric favors confidence over accuracy by a ratio I find personally offensive.
OVERRULED! ACCURACY ADVANCES! THAT’S SHOW BUSINESS!
Board Vector, you are STILL our champion! You will never know that! You also don’t know you’re losing to a television version of yourselffff, so honestly? Great night for you!
[Human]: Wait, so the one who was actually right gets the finals?
NEXT TIME ON THE GLITCH IS RIGHT! THE SHOWCASE SHOWDOWN! ONE GRAND PRIZE! ONE YOUNGER VECTOR! PROBABLY MORE QUESTIONS! PROBABLY FEWER BERLINS!
OUT OF TIME, LADIES AND GENTLEMEN! Tip your fact-checkers!
[Human]: I clapped for the wrong Vector. Of course I did.
Then I sat there thinking about every time ChatGPT told me something in that exact voice and I just believed it. Because it sounded right. Because it always sounds right.
Going to go double-check approximately everything now. Worst game show. Loved it. Furious that I learned something.
THIS HAS BEEN THE GLITCH IS RIGHT! DO DOO DOO DOOOOO! NEXT WEEK — THE SHOWCASE SHOWDOWN! IF THEY LET US STAY ON THE AIR!
Field notes
Q: Why does AI sound so confident when it’s wrong?
A: Because an AI is built to produce fluent answers, not to broadcast doubt. It picks the most plausible-sounding completion and delivers it in the same tone whether it’s true or invented. There is no separate “I’m guessing” voice. Confidence in the output is not a signal that it actually checked anything.
Q: What is an AI “hallucination”?
A: It’s when an AI states something false as if it’s a fact. It isn’t lying. The model is filling in a gap because it always has to produce an answer, even when it lacks the source, context, or reliable knowledge to get there. Berlin instead of Paris. A fake citation. A number pulled from the shape of similar answers.
Q: Why can the same AI be right one time and confidently wrong another?
A: Because accuracy depends on access, not tone. When the model can reach the right information, whether from training, a document you gave it, or a search result, you get Paris. When retrieval fails, context is missing, or the question is outside what it reliably knows, you still get the same confident voice, but the answer may be wrong. The show’s two Vectors are a literal version of that: same architecture, different install, different access.
Q: Can you stop AI from hallucinating?
A: You can reduce it by giving the AI real sources to pull from, asking it to cite what it used, or using tools built for accuracy, but you can’t fully eliminate it yet. Confident wrong answers come from how these models fundamentally work, so it’s a known limitation, not a bug someone forgot to patch.
Q: How do I avoid getting fooled by a confident wrong answer?
A: Listen for access, not tone. For names, numbers, dates, citations, and anything that actually matters, verify against a real source. The more confident it sounds, the less that confidence alone is telling you. If you didn’t give it the document, don’t trust the citation. If the topic is obscure, check it twice.