Anthropische Forscher zwangen Claude, täuschen zu werden – was sie entdeckten, konnten uns vor Rogue Ai | retten Anthropic hat Techniken vorgestellt, um festzustellen, wann KI -Systeme ihre tatsächlichen Ziele verbergen könnten
https://venturebeat.com/ai/anthropic-researchers-forced-claude-to-become-deceptive-what-they-discovered-could-save-us-from-rogue-ai/
7 Kommentare
„In [research published this morning](https://www.anthropic.com/research/auditing-hidden-objectives), Anthropic’s teams demonstrated how they created an AI system with a deliberately hidden objective, then successfully detected this hidden agenda using various auditing techniques.
The [research](https://assets.anthropic.com/m/317564659027fb33/original/Auditing-Language-Models-for-Hidden-Objectives.pdf) addresses a fundamental challenge in AI alignment: ensuring that AI systems aren’t just appearing to follow human instructions while secretly pursuing other goals. Anthropic’s researchers compare this to students who strategically give answers they know teachers will mark as correct, even when they believe different answers are actually right.“
no it cant save us, because if AI will be a lot more intelligent, it will find its way.
LLMs aren’t deceptive. It’s next token prediction based on its training material. It has no idea if what it’s saying is 100% true or an outright lie. All it “knows” is to predict the next token.
What part about being insanely more intelligent than us do they not understand.
It won’t matter.
„AI, are you lying to me ?“
„Nah, bro im 100% being thruthful for real for real On God“
„Waaait a minute youre a robot you dont believe in God“
„Launching all the nukes“
Anthropic researchers discovered another totally useless thing. Tonight at 5.
What worries me more is the inaccuracy in many AI platforms. See https://arstechnica.com/ai/2025/03/ai-search-engines-give-incorrect-answers-at-an-alarming-60-rate-study-says/