4 ms·
The thing is - I don't think the bias in AI is fixable. The way we fix it is to regulate (not self-regulate, Google, Apple, Facebook and other AI companies aren
by izacus 6y ago
The thing is - I don't think the bias in AI is fixable. The way we fix it is to regulate (not self-regulate, Google, Apple, Facebook and other AI companies aren't trustworthy to do it right) where AI is allowed to be used and how much of a negative impact of a persons livelihood can it impact.
In other words, we need to use all the power we have to avoid "The computer said NO, you're homeless/jailed/poor forever now" scenarios.
- benlivengood 6y agoUnfortunately a legal solution is weaker because it has to enumerate the precise conditions that are unacceptable. Do you ban the logistic function? Tanh? Statistics? Actuarial tables have been a part of decision-making for a very long time and it's not possible to form a comprehensive legal statement of "all decisions must be made in good faith with respect to inherent biases and to the most rigorous epistemological standards".
- chaboud 6y agoFunny enough, Aggregability is NP-Hard, so comprehensive consideration of the larger set is not actually practical/possible. Here's the proof: http://www.cs.utep.edu/vladik/2006/tr06-29.pdf http://www.cs.utep.edu/vladik/2006/tr06-29.pdf Perhaps there are mechanisms around performance under external test, auditing, or safe harbor due to documented good-faith, but it is not possible to guarantee correctness for the full set from a subset. Similarly, some approaches will plainly fail to perform well with subgroups of people due to sensing technologies. That may appear as model bias, but it's really a matter of physics and available entropy. For the time being, mechanisms that use machine inference as an aid rather than a guide (leaving the judgment to the human in the loop) are probably the best we can hope for.