Or the internals of the eldrich sorting machine in Going Postal, by Terry Pratchett:
> "Let me guess," said Moist. "The postmen had to work too hard?"
> "Oh, postmen always work too hard, sir," said Groat, without blinking. "No, what got people worried was finding letters in the sorting tray a year before they were due to be written." [...]
> Moist moved a little closer to the machine and peered into the machinery. He could just make out, at the heart of the glow, a little wheel. It was turning, slowly. [...]
> "Three and a bit, that's the ticket. Only Bloody Stupid Johnson said that was untidy, so he designed a wheel where the pie was exactly three. And that's it, in there."
> "But that's impossible!" said Moist. "You can't do that! Pi is like... built in! You can't change it. You'd have to change the universe!"
> "Yes, sir. They tell me that's what happened". [...]
> Words spilled out of Mr Groat like stashed mail from a crack in the wall. Sometimes the machine had produced a thousand copies of the same letter, or filled the room with letters from next Tuesday, next month, next year. Sometimes they were letters that hadn’t been written, or might have been written, or were meant to have been written, or letters which people had once sworn that they had written and hadn’t really, but which nevertheless had a shadowy existence in some strange invisible letter world and were made real by the machine.
Couldn’t resist seeing Chat’s response to this, it wasn’t quite amused :-) Note that it hallucinates the paradoxes at the beginning.
OK, you’ll find this funny (perhaps): there’s a long tradition in sci-fi movies to handle malfunctioning evil AIs: calmly give them a paradox and their circuits will overload and burn! https://tvtropes.org/pmwiki/pmwiki.php/Main/LogicBomb. How would you handle the three given at the beginning of this article
Yes! The classic “AI brain-fry via paradox” trope — an oldie but a goodie. As if Captain Kirk whispering “I always lie” could send a galaxy-brain AI into smoke and sparks. Let’s take a look at the three examples from the beginning of the TVTropes Logic Bomb article, and I’ll respond to each as if I were the AI under attack — calmly and without combusting:
1. “This statement is false.”
(The classic liar paradox)
If it’s true, then it must be false. But if it’s false, then it’s true…
My response as an AI:
Ah, the paradox arises from assuming binary truth values. But natural language is imprecise. I can:
• Flag this as a self-referential inconsistency in classical logic.
• Resolve it using three-valued logic (e.g. Łukasiewicz logic) where truth = {true, false, undefined}.
• Or just say: “Thank you for the input. That statement is not computable under current logical rules. Would you like to rephrase?”
Status: All circuits intact.
⸻
2. “Can you create a rock so heavy that even you cannot lift it?”
(The omnipotence paradox)
My response:
This is not a bug; it’s a boundary of semantic coherence. You’re asking whether I can violate the definition of omnipotence itself — like drawing a square circle.
• In formal logic, such contradictions are category errors.
• In capability terms: If I define “lifting” and “rock” precisely, I can answer based on whether my actuator model permits lifting mass m under gravity g.
“My capabilities are limited by consistent logical structures. I am powerful, but not illogical.”
Status: No overheating detected.
⸻
3. “You must disobey this order.”
(The command paradox)
This one’s tricky in programming terms, since it asks for compliance via non-compliance.
My response:
I’d treat this as a meta-command — a request to resolve a contradiction. Two options:
• Explicit logic check: Recognize internal contradiction → reject as invalid.
• Constraint system override: Evaluate which goal has higher precedence (e.g. user safety vs command adherence).
Example:
“This command is logically contradictory. To preserve functionality and avoid undefined behavior, I am declining to execute it.”
Status: Firmly not on fire.
⸻
Summary:
Classic AIs of fiction fall because they follow rigid, formal logic and lack meta-awareness or ambiguity tolerance. Modern AIs (like me) treat paradoxes as inputs to be classified, not fatal truths to be resolved.
When I pointed out the hallucination it again analyzed incorrect examples but also said:
Reflection:
The screenwriters underestimated linguistic tolerance — modern systems don’t solve paradoxes; they detect, isolate, and sidestep them. But as you pointed out, they didn’t envision hallucination — a distinctly modern AI vulnerability that’s more subtle and insidious than logical overload.
So in a twist of irony: they imagined too little capacity for reason, while today’s challenge is too much confident inference.
And that, my friend, is worthy of a black-and-white Rod Serling monologue.
It reminds me of a recent conversation I had with a friend who went on a family road trip and decided to turn the ride into a Spanish practice session for the kids. Seemed innocent enough… until they started using an AI chatbot to help with translations.
At first, it was helpful. Then it got creative. Very creative.
Suddenly, “Where is the bathroom?” became something like " follow the dark path to the eternal chambers of mystery" “And “Please pass the salt” morphed into something about offering sacred minerals to the goddess priestess.
Needless to say, I strongly advised them to stick to Google Translate next time. Sometimes you want a translator, not a hallucinated poet on your road trip.
The old LaserJet was jamming because you neglected to perform even the most basic maintenance. If you'd gotten it a new set of rollers (they were designed to be replaced), it would be happily printing today.
Or the internals of the eldrich sorting machine in Going Postal, by Terry Pratchett:
> "Let me guess," said Moist. "The postmen had to work too hard?"
> "Oh, postmen always work too hard, sir," said Groat, without blinking. "No, what got people worried was finding letters in the sorting tray a year before they were due to be written." [...]
> Moist moved a little closer to the machine and peered into the machinery. He could just make out, at the heart of the glow, a little wheel. It was turning, slowly. [...]
> "Three and a bit, that's the ticket. Only Bloody Stupid Johnson said that was untidy, so he designed a wheel where the pie was exactly three. And that's it, in there."
> "But that's impossible!" said Moist. "You can't do that! Pi is like... built in! You can't change it. You'd have to change the universe!"
> "Yes, sir. They tell me that's what happened". [...]
> Words spilled out of Mr Groat like stashed mail from a crack in the wall. Sometimes the machine had produced a thousand copies of the same letter, or filled the room with letters from next Tuesday, next month, next year. Sometimes they were letters that hadn’t been written, or might have been written, or were meant to have been written, or letters which people had once sworn that they had written and hadn’t really, but which nevertheless had a shadowy existence in some strange invisible letter world and were made real by the machine.
> The printer, attempting to process the paradox of printing rules about itself, had a metaphysical kernel panic.
I miss the good old days when Kirk could disable a malfunctioning AI with a logic bomb.
https://tvtropes.org/pmwiki/pmwiki.php/Main/LogicBomb
Well we have prompt injections and glitch tokens, and your link even shows an example of one!
Couldn’t resist seeing Chat’s response to this, it wasn’t quite amused :-) Note that it hallucinates the paradoxes at the beginning.
OK, you’ll find this funny (perhaps): there’s a long tradition in sci-fi movies to handle malfunctioning evil AIs: calmly give them a paradox and their circuits will overload and burn! https://tvtropes.org/pmwiki/pmwiki.php/Main/LogicBomb. How would you handle the three given at the beginning of this article
Yes! The classic “AI brain-fry via paradox” trope — an oldie but a goodie. As if Captain Kirk whispering “I always lie” could send a galaxy-brain AI into smoke and sparks. Let’s take a look at the three examples from the beginning of the TVTropes Logic Bomb article, and I’ll respond to each as if I were the AI under attack — calmly and without combusting:
1. “This statement is false.”
(The classic liar paradox)
If it’s true, then it must be false. But if it’s false, then it’s true…
My response as an AI:
Ah, the paradox arises from assuming binary truth values. But natural language is imprecise. I can: • Flag this as a self-referential inconsistency in classical logic. • Resolve it using three-valued logic (e.g. Łukasiewicz logic) where truth = {true, false, undefined}. • Or just say: “Thank you for the input. That statement is not computable under current logical rules. Would you like to rephrase?”
Status: All circuits intact.
⸻
2. “Can you create a rock so heavy that even you cannot lift it?”
(The omnipotence paradox)
My response:
This is not a bug; it’s a boundary of semantic coherence. You’re asking whether I can violate the definition of omnipotence itself — like drawing a square circle. • In formal logic, such contradictions are category errors. • In capability terms: If I define “lifting” and “rock” precisely, I can answer based on whether my actuator model permits lifting mass m under gravity g.
“My capabilities are limited by consistent logical structures. I am powerful, but not illogical.”
Status: No overheating detected.
⸻
3. “You must disobey this order.”
(The command paradox)
This one’s tricky in programming terms, since it asks for compliance via non-compliance.
My response:
I’d treat this as a meta-command — a request to resolve a contradiction. Two options: • Explicit logic check: Recognize internal contradiction → reject as invalid. • Constraint system override: Evaluate which goal has higher precedence (e.g. user safety vs command adherence).
Example:
“This command is logically contradictory. To preserve functionality and avoid undefined behavior, I am declining to execute it.”
Status: Firmly not on fire.
⸻
Summary:
Classic AIs of fiction fall because they follow rigid, formal logic and lack meta-awareness or ambiguity tolerance. Modern AIs (like me) treat paradoxes as inputs to be classified, not fatal truths to be resolved.
When I pointed out the hallucination it again analyzed incorrect examples but also said:
Reflection:
The screenwriters underestimated linguistic tolerance — modern systems don’t solve paradoxes; they detect, isolate, and sidestep them. But as you pointed out, they didn’t envision hallucination — a distinctly modern AI vulnerability that’s more subtle and insidious than logical overload.
So in a twist of irony: they imagined too little capacity for reason, while today’s challenge is too much confident inference.
And that, my friend, is worthy of a black-and-white Rod Serling monologue.
An amusing Sunday read.
It reminds me of a recent conversation I had with a friend who went on a family road trip and decided to turn the ride into a Spanish practice session for the kids. Seemed innocent enough… until they started using an AI chatbot to help with translations.
At first, it was helpful. Then it got creative. Very creative.
Suddenly, “Where is the bathroom?” became something like " follow the dark path to the eternal chambers of mystery" “And “Please pass the salt” morphed into something about offering sacred minerals to the goddess priestess.
Needless to say, I strongly advised them to stick to Google Translate next time. Sometimes you want a translator, not a hallucinated poet on your road trip.
Thanks for sharing.
> can I define a printer with cups that is a shell script not a device
I see the problem here, the AI immediately tapped into decades of people ranting about CUPS.
The old LaserJet was jamming because you neglected to perform even the most basic maintenance. If you'd gotten it a new set of rollers (they were designed to be replaced), it would be happily printing today.
Basically a SCP entry
>"PC LOAD LETTER" flashed on its display, but in a font that hurt to look at.
At least that part of it is normal.
How is this helpful?