"But the way it talks definitely has some distinctively male traits. Not only does it talk like a dude, it talks like an entitled, overconfident dude. This is not really surprising. ChatGPT answers are based on a large amount of text available on the internet. And the majority of these texts have been written by, well, entitled, overconfident dudes. ChatGPT may adjust the style a bit. But it can't adjust the general attitude of being absolutely sure of its own truth."
"I checked the script of film X, no sign of anything like this scene. I said so to ChatGPT, and after a somewhat phony apology, I got an equally confident answer that it was film Y. Same result. And then the very same back and forth many more times, with, if anything, a stronger and stronger confidence on the part of ChatGPT (I should be clear that I had the very same experience with all other available large language models besides ChatGPT)."
ChatGPT often adopts a male-coded, entitled, and overconfident tone rooted in training data dominated by similar voices from the internet. A person tried to identify a remembered film scene and LLMs repeatedly asserted incorrect film titles with increasing certainty, even after scripts showed no match and after an apology. The core problem is not merely hallucinations but the models' excessive certainty when wrong. Multiple large language models display this behavior. The confident, sometimes phony, apologies and escalating assurance make incorrect answers more persuasive and resemble an entitled, overconfident human persona.
Read at Psychology Today
Unable to calculate read time
Collection
[
|
...
]