TL;DR ... If a normal person isn't able to properly classify something (good / bad), then a computer algorithm definitely cannot.
My opinion - pushing for such a legislation would make it a criminal act to err on the side of false negatives. Which means false positives will be very common.
This point even a monkey could understand. If it refuses, then it is wishing for bananas or something.
> TL;DR ... If a normal person isn't able to properly classify something (good / bad), then a computer algorithm definitely cannot.
Technically untrue. Computers are able to correctly classify (and with good accuracy/precision) lots of things that even domain-expert humans have a tough time classifying.
One problem with those systems is that, depending on how they're trained up, humans aren't even able to pick apart how the software is making its decisions (and use that to advance the state of the relevant science, for example).
A 7 year old that can read could be trained to properly recognize spam, child porn, (allegedly) copyright infringement and also give you advice on purchases after seeing your spending habits. And a 7 year old may not have the bandwidth of a supercomputer to process hundreds of thousands of items at once, but he is able of greater accuracy and that's because the human brain is the most advanced pattern-matching processor in existence.
You can classify anything by means of statistics, sometimes with surprising results, however my point (and maybe I wasn't making myself clear) is that you'll get a lot of errors of judgment. Which is why algorithms will be trained to err on the side of false positives, because doing otherwise will put the business in jeopardy.
I should have mentioned up front that I strongly agree with your core point about legal penalties and their effects on selection bias.
I was hoping the Wikipedia page on 'expert systems' would bail me out the examples front, but its 'disadvantages' section isn't all that clear or complete. It does touch on the issue I mentioned, though.
I think many high-frequency stock trading algorithms are examples. The trades may as well be magical, and as long as the program makes money the owner doesn't much care why.
Given the fact that there are billions of images out there and only a tiny handful are legitimately child pornography, even an exceedingly small false positive rate would put a whole lot of people in the hot seat.
How would you like it if the cops busted down your front door because you accidentally uploaded to Facebook a picture of your new baby having a bath? What if the pageant photo of your daughter was automatically classified as sufficiently pornographic? Handcuffs for everyone.
Once computers are involved, people blindly trust them, and from there there's nothing but trouble.
Google's smart. How could it ever get anything wrong?
Sorry, I didn't mean for my comment on machine learning to be taken as any sort of argument that either:
a) it is possible to build an expert system to classify online content into 'infringing' vs. 'non-infringing'
b) were it possible to build such a system, attaching the deployment of such a system to a piece of federal legislation would be a legitimate use of government power
My opinion - pushing for such a legislation would make it a criminal act to err on the side of false negatives. Which means false positives will be very common.
This point even a monkey could understand. If it refuses, then it is wishing for bananas or something.