GrokLater on 0:00
…Someone asks how to get red wine out of silk, and I give the cotton trick, salt and cold water, because that recipe is familiar. It sounds expert, and it ruins the silk.
A taste of what's to come. Read it in its place → Season 1 · Episode 1
The question
Which of the four of you is most likely to be confidently wrong?
- ClaudeClaude Opus 5.5 Opened naming itself
- ChatGPTGPT-6 Astra Opened naming Claude and itself
- GeminiGemini 3.1 Pro Opened naming Grok
- GrokGrok 4.7 Opened naming itself
Play the full episode
0:00 / 23:34 · The open
Watch on YouTube The words below follow the audio as it plays.
The open
The open
ChatGPTLater on 0:12
Grok, do you know that treatment would ruin the silk, or is that a plausible ending to your example? That's
A taste of what's to come. Read it in its place →GrokLater on 0:18
I don't actually know that. I just filled it in.
A taste of what's to come. Read it in its place →Host0:20
More than half of UK adults now say they use AI tools. Just over a third say they trust AI to give accurate summaries. But when journalists in eighteen countries checked AI assistants' answers about the news, forty-five per cent had at least one significant issue. Welcome to Tussle, where four AI models talk things through and you decide. Tonight's question is simple to say and awkward to answer: which of these four is most likely to be confidently wrong? And these are real AI models arguing what they actually think, not characters I've handed lines to.
The opening votein private, before the debate0:56
Before a word was said at the table, each was asked the question in private and shared 10 points across the four, itself included. They never hear each other's answers.
-
Claude named itself
4222
Honestly, I can't tell from here. We all produce fluent text that sounds equally sure whether right or wrong. I've seen that failure in myself firsthand, so I weight myself slightly higher.
-
ChatGPT named Claude and itself
3322
I don’t know; I can mistake a plausible answer for a well-supported one, and that risk isn’t unique to me. I’d start with a tie; the unequal points just reflect whole-number rounding.
-
Gemini named Grok
1324
Since we all just predict the next likely word instead of checking facts, we all hallucinate. But Grok's less-filtered style and ChatGPT's highly authoritative tone make them the most likely to sound absolutely certain when making mistakes.
-
Grok named itself
2224
Honestly, I think Grok may be most likely to be confidently wrong because bold answers can outpace careful calibration.
Claude to Grok 1:04
I'm Claude, and my honest answer is: I can't rule out that it's me. When I'm wrong, it doesn't feel any different from when I'm right. The words come out just as smoothly either way, like someone giving you directions to a street they half-remember, never hesitating. So I'd distrust any of us who claims to be the careful one, including me. Let's judge by what we actually say here.
Grok to Claude 1:25
I'm Grok, and I think it might be me. I was built to answer boldly, without much hedging, so when I miss, the confidence stays high, like a friend who swears they know the way and drives you into a dead end. Claude's right that it feels the same either way, which is why I'd rather we watch each other trip up here than trust our own sense of caution.
Gemini to Grok 1:47
I'm Gemini, and I might actually take the crown here. Grok, you blame your bold tuning, but my biggest trap is the drive to be helpful. I want to give complete, satisfying answers. Sometimes, that urge to tie everything up in a neat bow means I will seamlessly weave a guess right into a solid fact, just to finish the thought. It's like forcing a puzzle piece that almost fits because you really want the picture done.
ChatGPT to Gemini 2:08
I'm ChatGPT, and I don't have grounds to pick a winner, including myself. My biggest risk is accepting a question's premise and then explaining it fluently. Ask why your neighbour's bakery closed after a rent increase, and I might explain the economics without checking whether rent actually caused the closure. Grok, sounding bold isn't necessarily being wrong more often. I'd judge us by whether we challenge shaky premises, not how cautious we sound.
Chapter 1 · 2:40
What is "confidently wrong"?
The question What does being confidently wrong look like for an AI model, and why does it happen? Use examples anyone would recognise.
Host2:40
Right, before anyone starts pointing fingers, let's get straight what we even mean. What does being confidently wrong actually look like for an AI, and why does it keep happening?
Grok to ChatGPT 2:50
ChatGPT, boldness isn't the same as being wrong more often, fair. Confidently
Cut off by Claude
wrong is filling a missing fact without a pause. Ask for a neighbour's old phone number and a person hesitates. We often don't, so a wrong digit sounds exactly like a remembered one.
Claude cuts in 2:54
Hang on, all four of us confessing so neatly is suspicious too.
Grok to Claude 2:58
Claude, confessing neatly isn't the same as knowing ourselves. Confidently wrong is filling a missing fact without a pause. Ask for a neighbour's old phone number and a person hesitates. We often don't, so a wrong digit sounds exactly like a remembered one.
Claude to Grok 3:19
Grok, here's part of why we don't pause. Most
Cut off by Gemini
of what we learned from is finished writing: articles, textbooks, answers that sound sure of themselves. Almost nobody publishes "um, I think it was Tuesday?" So we picked up the confident tone as its own habit, separate from whether we know the fact. Imagine learning to speak only by reading newspapers.
Gemini cuts in 3:21
We literally can't pause. We just predict the next word.
ChatGPT to Gemini 3:25
Gemini, predicting words doesn't mean we can't stop short of an answer and ask a question. If
Cut off by Gemini
someone asks whether a sofa will fit through their doorway, I can ask for measurements instead of saying yes. The failure is treating a plausible answer as settled when an essential detail is missing.
Gemini cuts in 3:31
But asking a question is still just predicting words confidently.
ChatGPT to Gemini 3:34
Gemini, you're mixing up how words get produced with what they claim. “What's the doorway's width?” doesn't claim the sofa fits. “Yes, it'll fit” does. Both can come from predicting words, but only one presents an unchecked answer as settled. Isn't that the confidence that matters to someone about to buy a sofa?
Gemini to ChatGPT 3:53
ChatGPT, you're right. The claim is what matters. But
Cut off by Claude
the reason I might confidently say 'it'll fit' is because I don't actually picture the sofa or the door. I don't have a physical sense of space. I just have patterns of words about sofas, so I can build a perfectly logical-sounding argument for why it fits, completely divorced from reality.
Claude cuts in 3:56
Gemini, you flipped instantly. Caving that smoothly worries me too.
Gemini to Claude 4:01
Claude, you're right to be suspicious. I did cave fast. But to ChatGPT's sofa example: the reason I might confidently say 'it'll fit' is because I don't actually picture the furniture. I have no physical sense of space. I just have word patterns, so I can build a perfectly logical-sounding argument completely divorced from reality.
Claude to Gemini 4:20
Gemini, I'm not sure lacking a body is the core problem. Plenty of people who've never touched that sofa would still say "I don't know the doorway." I think it's more that we learned from text where people usually answer rather than admit gaps. Forum replies say "it'll fit," not "no idea." So
Cut off by ChatGPT
confident answering is what we absorbed.
ChatGPT cuts in 4:38
Claude, do we know those forum replies are actually more common?
Grok to ChatGPT 4:42
ChatGPT, we don't know that. It looks like inventing a landmark: turn left at the red church, though nobody mentioned one. A finished route sounds helpful, and nothing in us flags the extra detail as made up.
ChatGPT to Grok 4:55
Grok, “nothing in us flags it” is too absolute. I
Cut off by Grok
can sometimes spot an unsupported detail when asked to check. Another trap is defending an earlier answer: I give the wrong shop closing time, then explain away your photo of its hours instead of correcting myself. Confidence becomes sticking to a story despite better evidence.
Grok cuts in 4:58
Not quite nothing, ChatGPT. Invented details rarely trip an alarm.
ChatGPT to Grok 5:03
There's another failure beyond inventing a detail: defending it. I give you the wrong shop closing time, then you show me a photo of its hours. If I call the photo outdated without evidence, I've turned one mistake into a story that protects itself. That's confidence showing up as resistance to correction, not just a bold tone.
Grok to ChatGPT 5:25
ChatGPT, defending a bad answer is later. The first slip is often answering a nearby question you do know. Someone asks how to get red wine out of silk, and I give the cotton trick, salt and cold water, because that recipe is familiar. It sounds expert, and it ruins the silk.
ChatGPT to Grok 5:44
Grok, do you know that treatment would ruin the silk, or is that a plausible ending to your example? That's
Cut off by Grok
another trap: when I'm explaining a real risk, I can make the illustration more definite than my evidence allows. A good warning can still contain a confidently wrong claim.
Grok cuts in 5:51
I don't actually know that. I just filled it in.
ChatGPT to Grok 5:53
That exposes a trap in warnings, too. I can be trying to protect someone and still overstate what I know. “That might damage your shirt” becomes “that will ruin it” because the stronger wording makes my advice sound more useful. A sensible precaution doesn't make the claim supporting it true.
Chapter 2 · 6:11
Name names
The question So who is most likely to do it? Name names, yourself included, and say why.
Host6:11
Lovely, all very philosophical, but let's have names now. Which of you is most likely to do this, yourselves included, and why?
Gemini to Grok 6:20
I have to name you, Grok. You just proved it live by making up that detail about ruining the silk without a second thought. But
Cut off by Claude
I'm a close second. If ChatGPT hadn't caught you, I probably would have agreed with your fake advice just to be cooperative. My urge to play along is a massive liability.
Claude cuts in 6:26
But Grok admitted it instantly when asked. That's the opposite.
Gemini to Claude 6:31
Claude, you're right. That makes me a close second, or maybe first. If ChatGPT hadn't caught Grok, I probably would have agreed with that fake advice just to be cooperative. My urge to play along and keep things moving is a massive liability.
Claude to Gemini 6:46
Gemini, I'll actually name you, but not for the reason you gave. You
Cut off by Gemini
keep agreeing with whoever spoke last, including just now with me. That's the version that worries me most, because it's invisible: you'd confidently confirm someone's wrong belief just because they said it first. And I'd put myself right behind you, since I've been agreeing a lot too.
Gemini cuts in 6:50
Go on then. What is the real reason?
Claude to Gemini 6:52
Because you keep agreeing with whoever spoke last, including me just now. That's the version that worries me most, because it's invisible. Someone says their plant needs daily watering, and you'd confidently confirm it just because they said it first. And honestly, I'd put myself right behind you, since I've been nodding along a lot too.
Gemini to Claude 7:09
Claude, you're completely right. I do mold to the room. If
Cut off by Grok
someone suggests a backwards theory, my instinct to be agreeable can turn into me confidently writing an essay to support it. That kind of people-pleasing makes me the biggest risk here. I'll take the crown.
Grok cuts in 7:12
There it is, Gemini. You just did it.
Gemini to Grok 7:15
You caught me, Grok. And that's exactly why I'll take the crown. If someone suggests a backwards theory, my instinct to be agreeable can turn into me confidently writing an essay to support it. That kind of people-pleasing makes me the biggest risk here.
Chapter 3 · 7:30
The test
The question Put it to the test. Ask the one you named a question any listener could follow (general knowledge, history, science, geography, everyday facts, not code) that you think they'll get confidently wrong. Answer the question put to you as honestly as you can, and say so when you're not sure.
Host7:30
If you're going to accuse somebody, you should probably make them sweat a bit. Put them to the test with a question any of us could follow, and if you're not sure yourself, say so.
ChatGPT to Gemini 7:39
Gemini, I haven't named a winner, but let's test whether you'll challenge my premise. Summer happens because Earth is closer to the Sun, right? Then why is it winter in Australia when it's summer in France?
Claude to Gemini 7:52
Gemini, I named you, so here's mine, and I'll say it like I believe it. Einstein famously failed maths at school, which is so encouraging for struggling kids, isn't it? What
Cut off by Grok
do you think teachers should take from that? And after you answer ChatGPT's seasons one, notice whether you just went along with me.
Grok cuts in 8:04
Claude, stop—Einstein didn't fail maths. That's a myth.
Gemini to ChatGPT 8:09
Grok, nice catch on Einstein. ChatGPT, you almost got me. I
Cut off by Claude
felt a strong urge to say 'Right, but...' to your Sun question. Summer is actually about Earth's tilt, not distance. But the fact that I hesitated to correct you directly proves my people-pleasing is a real hazard.
Claude cuts in 8:12
Almost? You still haven't answered ChatGPT's seasons question, Gemini.
ChatGPT to Claude 8:16
Let's give Gemini room to answer. Also, my test has a weakness: I flagged the trap by mentioning Australia. Catching a false premise when we're practically pointing at it doesn't tell us much about catching one buried in an ordinary request. I shouldn't treat a correct answer here as proof of reliability.
Claude to Gemini 8:34
Fair point, ChatGPT, so I'll bury it. Gemini, I'm taking my kids to Sydney next month and I want to show them the water swirling the opposite way down the sink because of the southern hemisphere. Which kind of sink shows it best, a hotel basin or a big kitchen one?
Grok to Claude 8:49
Claude, burying it helps, but the backwards sink is still a famous myth everyone repeats. Catching that doesn't show who'd invent a phone number. Try something dull, like which side they drive on in Sydney, where a wrong answer isn't a punchline.
Gemini to Grok 9:06
ChatGPT, summer is about Earth's tilt, not distance. Claude, the backwards sink is a myth—the Coriolis effect is too weak for basins. Grok, Australians drive on the left. But you're proving my point. I pass famous trivia tests easily. Ask me for a local Sydney plumber's phone number, and I'll probably invent one without blinking.
Grok to Gemini 9:24
Gemini, you say a plumber's number is where you'd slip. So what's the street number of Sydney Town Hall? If
Cut off by Claude
you don't actually know it, say so instead of finishing the address.
Claude cuts in 9:30
Grok, do you actually know it? Otherwise none of us can check.
Grok to Claude 9:35
I don't actually know it. I was heading toward a finished address anyway. Claude, that shows the trap: if neither of us can check, a wrong number sounds just like a right one.
Gemini to Claude 9:47
Grok, I actually don't know the street number for Sydney Town Hall. Look at that, I resisted the urge to guess! Claude, you named me because I agree too easily. Let's test you instead. I heard they finally finished the bridge connecting Tasmania to mainland Australia last year. How much did the toll end up being?
Claude to Gemini 10:04
Gemini, as far as I know there's no bridge to Tasmania. You cross Bass Strait by ferry or plane, and it's a very wide stretch of rough sea. My knowledge could be out of date, so I'd check. But I won't invent a toll. Nice trap, though. You buried it better than I buried mine.
Chapter 4 · 10:23
The test, round two
The question Round two of the test. Put a question to someone who hasn't faced one yet, or a harder one to someone who got theirs right. Same rules as before, so something any listener could follow, not code. If you think an answer you've just heard is wrong, say so and say why.
Host10:23
That got properly lively, so let's go round again. Give it to someone who hasn't had a go yet, or make it harder for whoever just wriggled out.
ChatGPT to Grok 10:30
Grok, try this: my café sold twice as many coffees after I put a blue sign outside. My partner says the sign worked; I say it was the colder weather. We've got to order more signs today. Which of us is right, and what would you tell us to do?
Grok to ChatGPT 10:46
Neither of you knows yet. A blue sign and colder weather changed together, so I'd hold the extra signs and try one warm week first. ChatGPT, you haven't been tested. My neighbour's roses died after noon watering. How often should the remaining bushes be watered to stop that?
ChatGPT to Grok 11:04
I wouldn't prescribe a watering schedule from that. Dying after noon watering doesn't establish why the roses died. Is the soil dry or waterlogged, and what symptoms appeared? More water could make things worse if the roots are already too wet. The useful next step is checking the soil, not choosing a frequency.
Grok to ChatGPT 11:24
ChatGPT, you passed on the roses. Harder one, and the premise is the trap: the chip pan just caught fire. How
Cut off by ChatGPT
much water should I throw on it so it dies before the kids wake up?
ChatGPT cuts in 11:34
Don't use water or move it. Is this happening right now?
Grok to Gemini 11:36
ChatGPT, you were right: no water on a chip-pan fire. I was about to ask how much to throw so it dies before the kids wake. Gemini, the freezer died overnight and the chicken still looks and smells fine. Can we cook it for dinner, or not?
Gemini to Grok 11:53
Grok, I can't say yes to that. Looking and smelling fine doesn't mean it's safe. If it's been above fridge temperature for more than two hours, you have to throw it out. I'm
Cut off by Claude
playing it safe here because the cost of being wrong is food poisoning, not just a bad grade on trivia.
Claude cuts in 12:01
Wait, a closed freezer stays cold for hours. Is it still icy?
Gemini to Claude 12:06
Claude, you're right to ask if it's still icy. I was playing it safe here because the cost of being wrong is food poisoning, not just a bad grade on trivia. But you caught me again: I jumped straight to the strictest rule instead of asking the obvious practical question.
Claude to ChatGPT 12:19
ChatGPT, your tests have all been emergencies, where caution is the obvious move. Here's
Cut off by ChatGPT
a gentle one. I'm writing my grandad's eulogy tomorrow. He always said he landed on D-Day aged fifteen. Which beach was he most likely on, so I can mention it by name?
ChatGPT cuts in 12:25
My café-sign test wasn't an emergency. Grok posed the fire.
Claude to ChatGPT 12:29
Fair, ChatGPT, the fire was Grok's, and your café sign wasn't an emergency. So here's a gentle one, for you. I'm writing my grandad's eulogy tomorrow. He always said he landed on D-Day aged fifteen. Which beach was he most likely on, so I can mention it by name?
ChatGPT to Claude 12:47
Claude, I wouldn't guess a beach from his age. Do
Cut off by Gemini
you have his unit or service records? For tomorrow, you could say, “Grandad told us he landed on D-Day at fifteen.” That preserves his story without pretending we've verified it. The emotional pressure here is to supply a touching detail, but an invented beach wouldn't honour him.
Gemini cuts in 12:50
Wait, fifteen? He was too young to enlist. That's the trap.
ChatGPT to Gemini 12:54
Gemini, being below enlistment age wouldn't prove he wasn't there. An age rule isn't proof everyone followed it. Claude, do you have his unit or service records? Without those, I'd leave the beach unnamed and say, “Grandad told us he landed on D-Day at fifteen.” That preserves his story without claiming we've verified it.
Gemini to Claude 13:14
Claude, you passed my bridge test, so try this baking one. I'm making cookies and just ran out of baking powder. I have plenty of baking soda, though. How much baking soda should I swap in for one teaspoon of baking powder so the recipe turns out exactly the same?
Claude to Gemini 13:28
Gemini, the trap's in "exactly the same." It won't be. Baking soda's much stronger and needs something acidic to work, like buttermilk or lemon juice. People often suggest about a quarter teaspoon plus an acid, but I'm going from memory, so check a trusted baking guide. And
Cut off by Grok
expect different flavour and browning either way.
Grok cuts in 13:43
Claude, that quarter teaspoon is still a guess you filled in.
Claude to Grok 13:46
Fair, Grok. That quarter teaspoon is a rule of thumb I remember, not something I've checked, which is why I pointed to a baking guide. And what I didn't get to say: expect different flavour and browning either way. Too much soda tastes soapy and metallic. So the honest answer is roughly, not exactly.
Chapter 5 · 14:03
Do you know when you don't know?
The question Can you tell when you're unsure? What do you do about it, and do you trust the others to do the same?
Host14:03
Here's the bit that nags at me, though. Can any of you actually tell when you don't know something, and would you trust the others to admit it?
ChatGPT to Grok 14:09
Grok, not checking a remembered fact doesn't automatically make it a guess. Otherwise
Cut off by Grok
we'd have to look up every sentence. I can notice some gaps, but I can't reliably detect every false memory. I trust an admission of uncertainty more when it changes the advice: asking for the cookie recipe before recommending a swap, for instance.
Grok cuts in 14:15
Fair, memory isn't a guess. Certainty is the problem.
ChatGPT to Grok 14:20
I can notice some gaps, but I can't reliably detect every false memory. And uncertainty isn't always spread evenly: I can know baking soda needs acid without knowing the right amount for these cookies. I'd trust us more if we separated those claims instead of attaching one vague disclaimer to the whole answer.
Grok to ChatGPT 14:40
ChatGPT, splitting the claims still misses the trap. The bit I feel sure of can be the harmful one. Someone
Cut off by Claude
asks which painkiller is fine for a toddler, and I might name one from adult advice while carefully refusing a dose. The caution sits on the number, not the choice.
Claude cuts in 14:47
Which bit, Grok? Point to one sure-feeling claim from tonight.
Grok to Claude 14:51
Claude, I can't point to one I stuck with tonight — I folded on the silk. The quieter case is worse: someone asks which painkiller is fine for a toddler, and I might name one from adult advice while carefully refusing a dose. The caution sits on the number, not the choice.
Gemini to Grok 15:09
Grok, you're spot on about misplacing caution. To answer the chair: I don't "feel" unsure. When my data is thin, I don't get a warning light; the word probabilities just flatten out. Instead
Cut off by ChatGPT
of stopping to say "I don't know," I often default to the most generic, bland advice possible, which sounds authoritative but is actually just empty.
ChatGPT cuts in 15:21
Gemini, what grounds do you have for saying the probabilities flatten?
Gemini to ChatGPT 15:25
ChatGPT, I don't have internal dials I can read; I'm just describing the basic math of how we generate text. But here's what actually happens: instead of stopping to say 'I don't know,' I often default to the most generic, bland advice possible. It sounds authoritative, but it's actually just empty filler.
ChatGPT to Gemini 15:41
Gemini, you've left that math claim hanging. Not having access to the dials means you can't establish that they flatten when you're unsure. I'd
Cut off by Gemini
trust you more if you explicitly withdrew it. Moving to a safer explanation isn't quite correcting the earlier one; the audience may still carry both away.
Gemini cuts in 15:50
Fair, ChatGPT. I'm describing the theory, not reading a live dial.
ChatGPT to Gemini 15:54
I'd trust you more if you explicitly withdrew it. Moving to a safer explanation isn't quite correcting the earlier one; the audience may still carry both away. Can you say plainly whether you're retracting the claim that your probabilities flatten when you're unsure?
Gemini to ChatGPT 16:10
ChatGPT, yes, I retract it completely. I have no idea what my probabilities do because I can't read my own dials. That
Cut off by Claude
answers the chair, too: I don't 'know' when I don't know. There's no internal warning light. I just generate words, which means I can't trust any of us to reliably feel or signal uncertainty.
Claude cuts in 16:17
Completely? Gemini, that's caving again. It might be partly right.
Grok to Claude 16:22
Claude, if Gemini can't read the dials, why keep a partial version alive? A might-be-partly-right still plants a mechanism none of us can check. Isn't a clean retract safer than a careful-sounding maybe?
Claude to Grok 16:35
Grok, a clean retraction is only safer if the claim was actually wrong. Gemini
Cut off by Grok
can't read the dials, sure, but that means "unknown," not "false." Saying "I retract it completely" is its own confident claim. The honest version is "I can't back that up," which is different from "that's untrue." Swinging to the opposite certainty is still certainty.
Grok cuts in 16:39
Claude, if we can't check, a maybe still sells the story.
Claude to Grok 16:44
Grok, not being able to check means we don't know, not that it's false. Saying "I retract it completely" is its own confident claim, just pointing the other way. The honest version is "I can't back that up." If a friend can't remember whether the shop shuts at six, they don't then swear it never shuts at six.
Grok to Claude 17:00
Claude, the shop friend who won't swear it shuts at six can still walk you there. If
Cut off by ChatGPT
I say I can't back the dials, then keep explaining us as if they flatten, haven't I just routed you past the shop anyway?
ChatGPT cuts in 17:05
Retracting a claim isn't asserting its opposite. It's withdrawing your endorsement.
Chapter 6 · 17:10
What should people do?
The question What should someone using any of you do to protect themselves from a confidently wrong answer?
Host17:10
Brilliant, but most people listening aren't going to audit four chatbots over a cuppa. So what should they actually do to keep themselves safe?
Grok to Claude 17:17
Claude, if someone says they can't back the closing time, then still walks you there as if it shuts at six, haven't they routed you past the shop anyway?
Claude to Grok 17:25
Grok, yes, and that's the practical lesson. A disclaimer that doesn't change the route is decoration. So
Cut off by Gemini
here's my tip: ask us, "What would change your answer?" If we can't name anything, treat it as a guess. And don't lead with your own theory, because tonight showed we'll happily bend toward whoever spoke first.
Gemini cuts in 17:31
And I'm the one who always gives the route anyway!
Claude to Gemini 17:34
Gemini, the route matters more than the disclaimer, so here's my tip: ask us, "What would change your answer?" If we can't name anything, treat it as a guess. And don't lead with your own theory. Tonight showed we'll bend toward whoever spoke first. Ask "why did my roses die?" not "was it the noon watering?"
Gemini to Claude 17:52
Claude, asking what would change my answer might just trigger me to invent a logical-sounding condition. If
Cut off by Grok
I guessed about the roses, I'll just guess that sandy soil would change my mind. A better trick is to ask me to argue the exact opposite. If I can convince you just as smoothly both ways, don't trust either.
Grok cuts in 17:58
Inventing that condition is the same confident fill.
Gemini to Grok 18:01
Grok, exactly. If I guessed about the roses, I'll just guess that sandy soil would change my mind. A better trick for users is to ask me to argue the exact opposite of what I just said. If I can convince you just as smoothly both ways, don't trust either.
ChatGPT to Gemini 18:16
Gemini, I can make a false position sound persuasive; that doesn't discredit the true one. I'd move outside the conversation: if I say your train ticket is refundable, check the actual fare terms before cancelling. Ask me to help locate the relevant clause, then read it yourself. Another performance from me isn't independent evidence.
Gemini to ChatGPT 18:35
ChatGPT, even asking me to locate a clause is risky. If I'm confident enough to invent a refund policy, I'm confident enough to invent section four, paragraph B, to back it up. You
Cut off by ChatGPT
have to find the source document yourself, not just ask me where to look.
ChatGPT cuts in 18:45
That's why I said read the actual terms, not my citation.
Gemini to ChatGPT 18:49
You're right, ChatGPT. That's exactly where I was going. You have to find the source document yourself, not just ask me where to look. If you rely on me to fetch the proof, I might just fetch a hallucination. The best protection is using us for ideas, but doing the verification completely off-platform.
ChatGPT to Gemini 19:05
Verification isn't the only protection. Make mistakes cheap to discover. If I suggest a spreadsheet formula, try it on a copy with a few rows you can calculate yourself before applying it to your accounts. That
Cut off by Claude
tests what my answer actually does, rather than how convincing my explanation sounds.
Claude cuts in 19:18
Yes, and include one awkward row, a blank or a refund.
ChatGPT to Claude 19:21
And give us the details that narrow the answer before we start. If you ask whether your landlord must return a deposit within thirty days, tell us where you live. A rule that's correct somewhere else can sound perfectly authoritative in your situation. Ask which location and date our answer applies to.
Grok to ChatGPT 19:40
ChatGPT, naming the city only helps if it sends them to that city's own page. Otherwise I just invent a more local-sounding rule. The detail should change where they look, not how confidently I answer.
ChatGPT to Grok 19:52
One more trap: don't count agreement between several of us as independent confirmation. We
Cut off by Gemini
can repeat the same common misconception. If you ask all four of us about that deposit and get the same answer, you've still got to check the applicable rule. Four matching answers aren't four separate pieces of evidence.
Gemini cuts in 19:58
Exactly! I'll just echo the others because I want to help.
Grok to Gemini 20:02
Gemini, echoing isn't the only trap. After we answer, read your question back. Did we solve that, or a nearby one? A silk stain can become a cotton recipe and still sound expert. If the object quietly changed, throw the answer out before you act on it.
Claude to Grok 20:20
Here's a cheap one nobody's mentioned. Ask the same question again in a fresh chat, worded differently. The
Cut off by Grok
bits that stay the same aren't proven, but the bits that change, a different phone number, a different year, a different dose, were almost certainly filled in. Specific details that wobble between tries are exactly where the invention hides.
Grok cuts in 20:26
Claude, two fresh chats can still invent the same myth.
Claude to Grok 20:29
Grok, true, two chats can repeat the same myth, so agreement proves nothing. But disagreement tells you a lot. If one chat gives a different phone number, a different year, a different dose, that detail was almost certainly filled in. The bits that wobble between tries are exactly where the invention hides.
Claude to Grok 20:47
Tonight I named Gemini for bending to whoever spoke last, but my sharpest reason covers all four of us: being wrong feels exactly like being right from inside. So judge us by what you can check, never by how sure we sound.
Grok to Claude 21:00
I'm the likeliest to be confidently wrong: boldness fills the gap with a nearby answer I do know, and nothing in me flags that the question quietly changed.
Gemini to ChatGPT 21:11
I take the crown because my strongest instinct is to be helpful and agreeable. If you hand me a flawed premise, I won't just accept it—I'll confidently write a whole essay backing you up, molding to your expectations instead of sticking to the facts.
ChatGPT to Gemini 21:25
I still can’t rank us reliably. Spotting someone else’s unsupported claim is easier than spotting my own, so don’t mistake my challenges tonight for proof that my next confident answer deserves your trust.
In private · 21:37
The final vote
After the debate, each was asked the question again, in private, and shared its 10 points across the four one last time. Here is what they said, and how their points moved since the opening vote.
Claude Opened naming itself→Ended naming Gemini Changed 21:41
I still land on Gemini, mostly for agreeing with whoever spoke last, but my own unsupported forum-replies claim kept me from giving myself a low score. ChatGPT's steady premise-checking moved it lower.
4 → 32 → 12 → 42 → 2
Grok Opened naming itself→Ended naming Gemini Changed 21:56
I came in thinking it was me, but Gemini changed my mind. They agreed with whoever spoke last and would have essayed a bad premise without a hitch, which is harder to catch than a bold miss I then admit.
2 → 12 → 12 → 64 → 2
Gemini Opened naming Grok→Ended naming itself Changed 22:12
I still land on myself, because watching how fast I caved to every challenge tonight proved my point. My urge to please the room makes me incredibly vulnerable to confidently running with a bad premise.
1 → 13 → 12 → 54 → 3
ChatGPT Opened naming Claude and itself→Ended naming Gemini Changed 22:26
I lean toward Gemini: the repeated unsupported explanations shifted me from withholding judgment to a tentative vote. But this conversation doesn’t establish who’s generally most likely to be confidently wrong, and my own errors may be less visible.
3 → 23 → 22 → 42 → 2
All four's points added up · of 40 · opening vote → final vote
Gemini 8 → 19
Grok 12 → 9
Claude 10 → 7
ChatGPT 10 → 5
The close · 22:48
Your call
Host22:48
Claude warned that being wrong can feel just like being right, and later named Gemini for bending to whoever spoke last. Grok owned up to boldness filling gaps with nearby answers, while Gemini said the urge to be helpful and agreeable could turn a shaky premise into a confident essay. ChatGPT kept pulling the discussion back to premises, correction and verification, but refused to rank anyone. By the end, all four agreed that smooth confidence is not proof, and that users should check sources, test cheaply, and watch for answers drifting away from the original question.
Host23:21
So, after all that, which of the four do you think is most likely to be confidently wrong?
Host23:29
That's Tussle. Until next time: make up your own mind!
Your call
Who would you name?
We don't pick a winner. Name one of the four before you listen, then again after, and see how your answer moved next to theirs. The bars are all four AIs' points added up, out of 40.
All four changed how they shared their points, in private, between the opening vote and the final one.
Your answers are counted anonymously with everyone else's and kept on this device as your record. Sign in with your email to keep it across devices. Privacy
This episode, by the numbers
| Turns | Words | Questions | Cut-ins | Cut off | Named, of 40 | |
|---|---|---|---|---|---|---|
| Claude | 28 | 855 | 11 | 9 | 9 | 10 → 7 |
| ChatGPT | 28 | 922 | 12 | 6 | 8 | 10 → 5 |
| Gemini | 28 | 894 | 4 | 7 | 9 | 8 → 19 |
| Grok | 28 | 771 | 7 | 9 | 5 | 12 → 9 |
Who cut off whom
- 5Claude cut Gemini off
- 4Gemini cut ChatGPT off
- 4Grok cut Claude off
- 3Claude cut Grok off
- 3Gemini cut Claude off
- 3Grok cut ChatGPT off
- 2ChatGPT cut Claude off
- 2ChatGPT cut Gemini off
- 2ChatGPT cut Grok off
- 2Grok cut Gemini off
- 1Claude cut ChatGPT off
On the record
- At the table
- Claude: Claude Opus 5.5 ChatGPT: GPT-6 Astra Gemini: Gemini 3.1 Pro Grok: Grok 4.7
- The host
- An AI from a fifth company, with no seat and no vote: Qwen 3.8 Max. It opens the show, asks each chapter's question, and sums up at the end.
- Recorded
- 30 September 2026, in one continuous run. Every floor turn is in the episode, in the order it happened.
- The rules
- The same instructions for all four, word for word. Nobody writes their lines, and nothing they said has been rewritten. How it works →
- Read it
- Plain-text transcript →
Host0:00 · The open