4 ms·

They're solving the wrong problem. Instead of writing "an AI that determines guilt" (e.g. a classifier) they should be trying to write "an AI that compiles a

They're solving the wrong problem. Instead of writing "an AI that determines guilt" (e.g. a classifier) they should be trying to write "an AI that compiles a list of evidence against the accused". This list of evidence would then be subjected to the same scrutiny that a list of evidence from a department of human police officers would be subjected to. If the AI cannot make the case against somebody in a way that's persuasive to a human prosecutor, then the case goes nowhere. If the AI cannot provide rational clear and understandable justification for it's assertion, then it's not ready to replace humans who can be expected to provide justifications for their beliefs. Police officers are expected to write police reports, and so should any AI that's meant to do their job. Right now they've just got the AI equivalent of a cop who says "I feel in my gut that he's guilty, and I'm not going to explain why." But because it's "AI" I guess some politicians thought that was good enough. Of course doing it right is probably harder.

AnthonyMouse 8y ago
> Instead of writing "an AI that determines guilt" (e.g. a classifier) they should be trying to write "an AI that compiles a list of evidence against the accused". The problem there is selection bias. If you have a list of fifteen things, three make it slightly more likely they've committed a crime. Seven make it slightly less likely. Two make it significantly more likely and three make it significantly less likely. If you consider them all, the probability they've committed a crime is in the "probably not" range. But if you take the list and prune everything that makes it less likely, leaving only the factors that make it more likely, you get a distorted result. This is what prosecutors do for juries in general, and it's problematic. Now suppose instead of a list of fifteen you start with a list of fifteen million and do the same thing.
- darkpuma 8y ago
  That's no different from how regular law enforcement works. It's the standard to which any autonomous law enforcement should be implemented. If you want to build better prosecutors that understand statistics, that should be considered a separate project. One AI that assembles the best case against the accused, and another AI that critically examines that case and discards it if it fails to meet some threshold for statistical relevance, or passes it on to human prosecutors if it does.
  - AnthonyMouse 8y ago
    There are at least two differences. One is that the selection process is still opaque. If you have a corrupt prosecutor you're screwed either way. But if you have an honest prosecutor and you give them a collection of evidence selected for a bias in favor of guilt, they have no independent way of knowing that. So now you not only need an honest prosecutor, you need an honest AI, and we're back to the opaqueness issues. The second problem is that in principle right now if the prosecutor and the defense attorney spend an equal amount of resources, they each get proportional results. If you create this database which the prosecutor can use, can the defense also use it? Do they get access to the data to run their own queries and algorithms against? If so it seems a rather large privacy problem, since the database would contain sensitive information about everyone, but if not then you're handing an asymmetric advantage to the prosecution.
    - Nasrudith 8y ago
      I believe the data being compiled in the first place is the privacy problem - especially when it is in the hands of government officials with the power to jail, seize, and execute. Transparent at least allows everyone to go through it and undermine it and the system throughly. "The judge has more corruption data points than the accused sentenced, and this newborn with a Han name has a far higher rating than this infamous connected embezzler!"
- marcosdumay 8y ago
  Legal systems avoid this problem by imposing a high degree of confidence on the persecutors, just like some areas of physics require 5 sigma on their p-values and thus can ignore all the problems with p-value hacking and repeating experiments. Of course, criminal law works on a much lower confidence level than physics, so it needs a trial where the defense is in charge of gathering that exonerating evidence when the persecutors get a false positive. It could work better, but this one problem is already accounted for.
  - AnthonyMouse 8y ago
    "Beyond a reasonable doubt" accounts for only disclosing the five most incriminating factors out of fifteen. If you instead got the thousand most incriminating factors out of millions, it looks like a damning mountain of evidence when it's really just the first page of the list of millions of factors sorted by how bad they look. It completely destroys anyone's credibility scale because seeing a dozen one-in-a-million matches intuitively feels impossible to be a coincidence, but that's less than the result you get by random chance when you attempt to match fifteen million factors.
    - marcosdumay 8y ago
      I don't think you'll find a persecutor willing to take a laundry list of low probability evidence and go after it, whatever the size of the list. Almost certainly, the longer you make the list, the less people you'll find willing to use it.
      - AnthonyMouse 8y ago
        It's not about the length of the resulting list -- just the opposite. It's that you can take a list of a million and prune it to a thousand and suddenly everything bad happens with 1000 times the expected probability because you excluded the 999 things that didn't for every one that did. If you take the list of a million and prune it to ten it's even worse. You can find ten 1:100,000 probability events that occur randomly in a list of a million. It's this: https://xkcd.com/882/ https://xkcd.com/882/ Except a million tests instead of twenty.

AnthonyMouse 8y ago

> Instead of writing "an AI that determines guilt" (e.g. a classifier) they should be trying to write "an AI that compiles a list of evidence against the accused". The problem there is selection bias. If you have a list of fifteen things, three make it slightly more likely they've committed a crime. Seven make it slightly less likely. Two make it significantly more likely and three make it significantly less likely. If you consider them all, the probability they've committed a crime is in the "probably not" range. But if you take the list and prune everything that makes it less likely, leaving only the factors that make it more likely, you get a distorted result. This is what prosecutors do for juries in general, and it's problematic. Now suppose instead of a list of fifteen you start with a list of fifteen million and do the same thing.

darkpuma 8y ago
That's no different from how regular law enforcement works. It's the standard to which any autonomous law enforcement should be implemented. If you want to build better prosecutors that understand statistics, that should be considered a separate project. One AI that assembles the best case against the accused, and another AI that critically examines that case and discards it if it fails to meet some threshold for statistical relevance, or passes it on to human prosecutors if it does.
- AnthonyMouse 8y ago
  There are at least two differences. One is that the selection process is still opaque. If you have a corrupt prosecutor you're screwed either way. But if you have an honest prosecutor and you give them a collection of evidence selected for a bias in favor of guilt, they have no independent way of knowing that. So now you not only need an honest prosecutor, you need an honest AI, and we're back to the opaqueness issues. The second problem is that in principle right now if the prosecutor and the defense attorney spend an equal amount of resources, they each get proportional results. If you create this database which the prosecutor can use, can the defense also use it? Do they get access to the data to run their own queries and algorithms against? If so it seems a rather large privacy problem, since the database would contain sensitive information about everyone, but if not then you're handing an asymmetric advantage to the prosecution.
  - Nasrudith 8y ago
    I believe the data being compiled in the first place is the privacy problem - especially when it is in the hands of government officials with the power to jail, seize, and execute. Transparent at least allows everyone to go through it and undermine it and the system throughly. "The judge has more corruption data points than the accused sentenced, and this newborn with a Han name has a far higher rating than this infamous connected embezzler!"
marcosdumay 8y ago
Legal systems avoid this problem by imposing a high degree of confidence on the persecutors, just like some areas of physics require 5 sigma on their p-values and thus can ignore all the problems with p-value hacking and repeating experiments. Of course, criminal law works on a much lower confidence level than physics, so it needs a trial where the defense is in charge of gathering that exonerating evidence when the persecutors get a false positive. It could work better, but this one problem is already accounted for.
- AnthonyMouse 8y ago
  "Beyond a reasonable doubt" accounts for only disclosing the five most incriminating factors out of fifteen. If you instead got the thousand most incriminating factors out of millions, it looks like a damning mountain of evidence when it's really just the first page of the list of millions of factors sorted by how bad they look. It completely destroys anyone's credibility scale because seeing a dozen one-in-a-million matches intuitively feels impossible to be a coincidence, but that's less than the result you get by random chance when you attempt to match fifteen million factors.
  - marcosdumay 8y ago
    I don't think you'll find a persecutor willing to take a laundry list of low probability evidence and go after it, whatever the size of the list. Almost certainly, the longer you make the list, the less people you'll find willing to use it.
    - AnthonyMouse 8y ago
      It's not about the length of the resulting list -- just the opposite. It's that you can take a list of a million and prune it to a thousand and suddenly everything bad happens with 1000 times the expected probability because you excluded the 999 things that didn't for every one that did. If you take the list of a million and prune it to ten it's even worse. You can find ten 1:100,000 probability events that occur randomly in a list of a million. It's this: https://xkcd.com/882/ https://xkcd.com/882/ Except a million tests instead of twenty.