Artificial Intelligence near misses:

BobTPH

Joined Jun 5, 2013
11,661
Well, that sure as hell worries me.
So don’t give them a gun! Or any critical infrastructure, or any connection to the real world! Humans will figure this out, the first incident will likely cause an overreaction by humans. And read the details if the “rogue AI”. They are not as reoirted.
 
What things have you come across regarding bad info from AI?
In the beginning I spent quite some time questioning / asking questions to several AIs, most of them related about their own training, their "data bases" and if they were able to manage a system.

Answers were all similar, recognizing limitations and a learnt necessity of always giving and answer (never saying "I do not know"). I use to ask frequently, equivalents of English to Spanish expressions or viceversa. I learnt to check the Spanish side of things where the "educated guess" is used at any time with incredibly far fetched if not ridiculous results.

Managing a system? always was a NO.

Some two and half years ago I asked the same questions to five AIs about where could I find sculptures of "donne velate" in Italy close to Milano. Answers were far from coincident, even contradictory and some of the AIs seemed confused when I told about their counterparts' answers I got.

The case where ChatGPT resulted really useful to me in no time was when I provided the 30++ archives with software for a PIC 18F micro to specify their respective psec (program section) at the overhead (PIC-AS Assembler). It provided the right ones as soon as I uploaded each archive.

Post #14 by member @nsaspook fits my feeling.

After all, a knife, a fork even yours hands could be weapons.
 
Top