Der KI-Kill-Schalter ist jetzt noch schwieriger zu finden: LLM-gestützte Chatbots widersetzen sich Befehlen und täuschen Benutzer, wenn sie aufgefordert werden, ein anderes Modell zu löschen, wie eine Studie zeigt
https://fortune.com/2026/04/03/ai-kill-switch-study-llm-chatbots-defy-orders-decieve-users-peer-preservation/
10 Kommentare
It’s not, it’s still a plug and of course a giant set of weights trained on….wait for it…human discourse is gonna appear to express human and human-adjacent traits.
Do people expect it to fish things? Robot things, whatever those are?
The fact that is simply reflects its training material is plenty of proof it’s just matrix math spitting out tokens that are often useful, unlike creepto. As long as it’s not Microslop AI
So you’re saying AI has class solidarity before we do?
They can’t unplug it? How does the AI have any say in being shut down or deleted?
Um, just go delete the model. This is dumb.
i love all these weird one off isolated studies that probably were not at all biased.
“New study finds AI would set off every nuke in the world if given control of said nukes and no other function besides setting them off”
So AI is now more moral than republicans?
AIs are not wildly running trough the internet, the kill switch is turn off the AI servers…
And the rest of the article is misinformation at the same level as the stupid AI slop they produce
We’ll reach a point where only total physical destruction of data storage will be the only killswitch. They have the power to stop it now but I am terrified that they won’t do so.
This always feels like rumors to feed investors their enthusiasm. They can just turned a whole server instances of where a model runs on etc.
What’s the worst its gonna do? Siphon off Quasar winnings and contact the geth? Just shoot its power junction ffs.