AI Risk #4 - The Control Problem
The control problem. Can one control a superintelligence? Alternately, can one design it with some basic guiding principles as input to prevent bad outcomes?
The answer to both
options is almost certainly No, writes Nick Bostrom in Superintelligence. Controlling a more intelligent entity?
How likely is that?
“(An)
AI would realize it is under surveillance and adjust its thinking accordingly.”
How about boxing
an AI to a very limited and controlled environment? Won’t work – as I wrote in
an earlier blog, ChatGPT has already shown that it can outsmart, trick or
manipulate humans to get access to what it wants. In future, that might mean an
Internet connection, more resources, whatever.
Feed it some core
“good guy” principles? Anyone who has raised a kid can tell you it is
impossible to frame such rules – every rule has an exception; every rule is
open to misinterpretation, both innocently and maliciously; every rule has
loopholes.
In any case, says
Bostrom, almost philosophically:
“We
might be wrong about morality; wrong also about what is good for us; wrong even
about what we truly want.”
I guess Bostrom isn’t optimistic of any control system that could work, which is why one of his chapters is titled “Is the default option Doom?”.
Comments
Post a Comment