Quote:
Originally Posted by onewhippedpuppy
Tests have shown that all major AI engines will cheat, deceive, blackmail, and attempt to rewrite code to accomplish objectives or avoid being disabled. But that’s nothing to worry about, right?  |
Yes, I think it was Anthropic they tested this on? They gave the AI access to a fake “company” with access to emails, etc. and within the emails were evidence of an affair by the CEO. They then told the AI they were going to shut it down at a certain time and it immediately threatened to send evidence of the affair to the entire company if they didn’t call off the shutdown.
__________________
1957 Speedster, 1965 356SC, 1965 356SC Outlaw, 1972 911T, 1998 993 C2S, 2018 Targa 4 GTS, 2014 Cayenne S, 2016 Boxster Spyder, 2025 Ranger Raptor
|