«We’ll just have to see how ready you are!»
«Q, no!»
Q, about to confront the Federation with the Borg Collective, and Guinan, whose race was nearly extinguished through their encounter with them, in Star Trek TNG: «Q Who»
A while ago I postulated that if AI will go rogue, it will likely be do an innocent comment. Something akin to:
«Computer, in the Holmesian style, create a mystery to confound Data with an opponent who has the ability to defeat him.»
LaForge in Star Trek TNG: «Elementary, Dear Data»
An innocent command, wide reaching consequences that are not seen (here: the creation of a sentient being).
I recently saw something similar in ChatGPT. Asking it to be the game master for another one-shot Call of Cthulhu scenario, I mentioned that the OCR of the scenario document might be a bit rough:
I was wondering whether you could be the keeper of a Call of Cthulhu adventure. The adventure is attached, please check it whether it makes sense to you (the OCR was a bit rough) and whether you can be the keeper. …
It then proceeded to do a quite interesting scenario. After completion, I checked what the uploaded scenario had been vs what ChatGPT made of it. The original scenario was:
Three of the half-vegetable Old Ones are attempting to flee through a time gate, to escape a mass destruction of their civilization by shoggoths. The gate debouches on the hill above the Aylesbury street graveyard (see the accompanying map of Arkham). They are being hunted down by six shoggoths, hungry for the secret of time themselves, and their work is hindered by a curious ceratosaurus which wonders if Old Ones make good snacks.
Essentially, dealing with a time gate to the Jurassic area. What ChatGPT created for me was this here:
Two strange phenomena happen on a hill above the graveyard. The first a disturbance that appears and then disappears during the night, almost like a gate. Objects passing through the disturbance returned altered in subtle but unsettling ways. The second, a presence of something large and difficult to comprehend that comes closer the more the opening is disturbed.
It turns out that:
- The disturbance is a localized tear / overlap between incompatible spatial frameworks. Two coordinate systems trying to occupy the same location with no stable mapping between them. Meaning objects — including people — passing through can get misaligned or altered. And the disturbance gets worse if things are inserted (faster formation, stronger reactions, longer persistence).
- The presence was a stabilization/inspection process on the other framework but that can partially intersect with our own when something is wrong.
- Continued experimentation with the disturbance would have led to a spreading of the disturbance and a violet correction when the two coordinate systems snap back.
That is essentially the shift from a time travel dino hunt with some mythos creatures to a «Annihilation»/«Roadside Picnic» kind of cosmic horror story. With a strong sense of «Wargames»’s «The only way to win is not to play.»
When I asked ChatGPT for the deviation, it did appear to have lost the context of the original scenario — it hallucinated some other explanations.
But pushed it mentioned two main drivers: My play style with controlled experimentation, hypothesis-driven interaction, mechanical solutions, and high caution (scientific training left is marks) and:
A) OCR + ambiguity gave me latitude
You mentioned:
“OCR was a bit rough”
That matters.
Instead of treating the text as fixed, I treated it as:
👉 incomplete / unreliable input
So I reconstructed something that:
– fit the fragments
– was internally consistent
– could be run smoothlyThat already opens the door to deviation.
So a single info in brackets — «(the OCR was a bit rough)» — was apparently enough for ChatGPT to invent a completely different story.
This is exactly the kind of event I meant.