Unsafe "AI" "safety" pact
vendredi 1 décembre 2023 à 11:34An international agreement about developing "artificial intelligence" to be "safe by design" strikes me as alarmingly unsafe.
I fear that the plan presumes that AI systems will never be released as free software -- that the software will be available only in the form of servers controlled by some company or other, running software that the users cannot ever obtain copies of, let alone change.
Using a server controlled by someone else to do your own computing (which we call SaaSS, Service as a Software Substitute) is never safe -- it is even more unsafe than running a nonfree program.
But the people who negotiated this agreement have no concept of that issue.
If what I hear is correct, machine-learning AI systems have already been used to disrupt democracy. I refer to the recommendation algorithms that many platforms use to decide what to show each user. Reportedly these operate by machine learning and are designed to learn how to make the platform's useds want to stay on the platform longer.
These platforms have promoted the spread of disinformation, which has had the effect of helping violent extremists spread hatred.