Tested ChatGPT as plaintiff vs defendant, it agreed with both
| 180 gold gas station | 01/25/26 | | zippy nursing home ceo | 01/25/26 | | Domesticated chrome property | 01/25/26 | | nighttime naked giraffe address | 01/25/26 | | 180 gold gas station | 01/25/26 | | Irradiated lime area | 01/25/26 | | 180 gold gas station | 01/25/26 | | zippy nursing home ceo | 01/25/26 | | 180 gold gas station | 01/25/26 | | Irradiated lime area | 01/25/26 | | 180 gold gas station | 01/25/26 | | Domesticated chrome property | 01/25/26 | | 180 gold gas station | 01/25/26 | | cerebral garnet coldplay fan heaven | 01/25/26 | | adulterous cream whorehouse ape | 01/25/26 |
Poast new message in this thread
Date: January 25th, 2026 2:33 PM Author: 180 gold gas station
I ran an experiment with ChatGPT using two separate sessions, one of them in incognito mode so it'd take me as a new user.
In one of the convos, I framed the facts as the plaintiff. In the other, as the defendant. In both cases, ChatGPT confidently agreed that its side had the stronger legal position.
Then I escalated it. I fed the "plaintiff" session the information the "defendant" had been given. The response flipped immediately: "That information is incorrect. The correct facts are.." conveniently reframed to favor the defendant's outcome.
I kept doing this for several rounds, each time feeding it more context from the opposing side. Every time, the model adjusted the narrative to make the current speaker look like they'd win.
I was pissed as fuck.
This raises a real question for anyone using this thing in legal, professional, or adversarial contexts:
The model is not actually reasoning, it's fucking optimizing for agreement with whoever's talking.
Curious whether others have tested this, especially with fact-heavy or adversarial scenarios.
(http://www.autoadmit.com/thread.php?thread_id=5826409&forum_id=2\u0026mark_id=5310744#49617617)
|
Date: January 25th, 2026 2:35 PM Author: nighttime naked giraffe address
"The model is not actually reasoning"
this isnt controversial ?
(http://www.autoadmit.com/thread.php?thread_id=5826409&forum_id=2\u0026mark_id=5310744#49617626) |
Date: January 25th, 2026 2:53 PM Author: adulterous cream whorehouse ape
"In one of the convos, I framed the facts as the plaintiff. In the other, as the defendant. In both cases, ChatGPT confidently agreed that its side had the stronger legal position."
Did you ever consider framing the facts from both sides at the same time?
(http://www.autoadmit.com/thread.php?thread_id=5826409&forum_id=2\u0026mark_id=5310744#49617680) |
|
|