Re: ai sucks
  By: Mike Powell to JAS HUD on Thu Mar 28 2024 08:45 am

 > I think it could be dangerous in the long (and even short) run.  As you
 > point out, many of the ai engines are scraping the net, which means they
 > are probably getting some good data and a whole lot of noise.  Soon, they
 > will be just as good at spewing out misinformation as any human could
 > imagine to be.

Linux Magazine published a great article about deliberate data poisoning that
could be used to mistrain an AI. It featured a practical example/tutorial. THey
made the point that you can make it obey hidden data inputs.

For example, if you are training the AI for recognicing handwritting, you
usually give it multiple samples. Now imagine all the "Rs" you show to it have
a hidden feature, like a particular ink color, or a small stain on the border
of the sample. THe machine would learn that all the letters with that hidden
feature are an "R". Then, in the wild, you could purposedly break the AI
results by presenting it a "C" that displayed the hidden feature and it will be
interpretated as an "R".

Now imagine the AI system is used to process bank checks and you can make it
mistake 0s and 1s for 9s.

--
gopher://gopher.richardfalken.com/1/richardfalken
--- SBBSecho 3.20-Linux
 * Origin: Palantir * palantirbbs.ddns.net * Pensacola, FL * (618:250/24)