3 ms·

The argument has nothing to do with efficiency, except insofar as human inefficiency makes abuse of power more likely. The problem is there isn't fundamentally

by jjk166 4y ago

The argument has nothing to do with efficiency, except insofar as human inefficiency makes abuse of power more likely. The problem is there isn't fundamentally a balance of power, having unlinked databases doesn't allow the "good" databases to shield you from the "bad" databases, it just increases the potential for errors to arise. I'll amend my previous closing argument, if you wouldn't trust someone with a knife, why would you trust them with a whole cutlery set? We're not trying to make the system more efficient, we are trying to eliminate potential failure points, which every human involved is. If it is more efficient, that's just gravy on top. If you are really concerned with who is in a position of power, consider the difference between a small team of maybe a dozen people working behind closed doors versus software which could be made as public as the laws it is meant to uphold and evaluated by anyone on earth. Any issue, be it from malice or incompetence, would be plain as day. And if you don't like black box neural networks, I have some terrible news about the human brain. There is no technical reason why the creators of an AI should be any less accountable for a flaw in their product than the producers of any other technology. This wouldn't generally be the programmers themselves, but it would be the companies employing those programmers. If a car company for example becomes aware of a serious flaw in their product and they don't take appropriate steps to fix it, they're liable. Likewise for most things. The creators of these systems don't need to volunteer to be responsible for how they perform, they are compelled. There have been countless tools added to law enforcement's utility belt over the centuries like radar guns and dna matching, and every legal system in the world has dealt with evaluating their efficacy and potential for bias. There is nothing about an AI that makes it incompatible with these existing safeguards.

msandford 4y ago
I believe that if a crime is sufficiently bad that we should prosecute for it, it absolutely merits the use of scarce resources. And if we can't muster up the resources to get humans to do the investigation, accusing, trying, etc it must not be that bad of a crime. Once algorithmic prosecution is enabled, we could easily see 100x as many accusations happen. That's the blessing of automation. 100x is actually a low bar. Many automations make millions more things possible. I don't think it would be a good world with a million times as much law enforcement with basically all of it automated.
- jjk166 4y ago
  Again, you're fundamentally missing what I'm saying. No one is arguing that humans require too many resources. I am saying that a computer, even if it were just as resource intensive as a human program, is a fundamentally preferable option. We do not live in a world where the more crime there is the more resources are available to deal with it. If anything, reality is the opposite: there is a vicious cycle where lack of government resources leads to crime and crime decreases available resources. It's trivially simple to imagine a scenario where crime gets so bad that we stop dealing with certain violations to conserve resources to deal with bigger issues. But that's irrelevant. What matters is that each and every person involved is a potential source of issues. Anyone can be tired, or scared, or distracted, or biased. This is not due to inadequate resources, though that may exacerbate the problem. This is the human condition - we are fallible beings. And in situations like this mistakes can ruin or end lives. Humans need oversight because of this fallibility, but the overseers are likewise fallible. The larger an organization is, the more places there are for something to fall through the cracks and for a mistake to go unnoticed and unresolved. Algorithmic prosecution might make accusations 100x cheaper, but you won't have 100x more accusations because there isn't 100x more crime to accuse people of. You're not going to see a massive spike in murders or bank robberies because of this algorithm, and the crime never having occurred is a really solid alibi. Sure if it's easier to investigate crimes then maybe some of the crimes we are currently tolerating because we simply don't have the manpower to deal with them would get investigated, but if we're ignoring a crime because we think the crime is not an issue, we can certainly keep ignoring it. The difference is the choice is made by society, not forced upon it by circumstance. You fear more law enforcement because you are plainly aware of how bad current law enforcement can be. This isn't more law enforcement, this is better law enforcement.
  - msandford 4y ago
    I strongly disagree that increased law enforcement capabilities will lead to ever more perfect policing. I'm pretty sure that's what you're arguing. When we switched from toll booths with people in them to automated tag reading, this increased the ability to create toll roads. By making the cost of enforcement cheaper, it's enabling. If you make the cost of law enforcement lower, that won't make law enforcement work harder at solving murders. It'll mean that whole swaths of things that previously weren't practical to police will become practical and even economic. Especially if there's an automated fining system. Computers don't fix the flaws in systems. They magnify them, by enabling those systems to be scaled way, way up.
mattmanser 4y ago
You keep missing the GP's point. His point is: - computers are fast and cheap - humans are slow and expensive - slow and expensive = rarely used technique - rarely used technique = less chance of privacy invasion Therefore forced human processing protects privacy. It's a privacy concept that's quite often discussed. Pre-computers we had a certain level of privacy afforded to us that was due to the cost/speed of humans processing large amounts of data. Computers have stripped that away. It doesn't matter if there are human processing errors as that comes out at court time. Everything else you're saying is also kinda wrong, unfortunately. You aren't looking at the problems caused by computer efficiency, just somehow computers are magically better because they follow the same rules every time. Firstly, the data is usually bad and messy. Like super bad and messy. Computers can't tell when data is obviously wrong, so innocent people can get caught in "computer says no" trap. So you're wrong computers are better because of that. They have no common sense. Secondly, there ARE 100x more crime than get prosecuted. You're dead wrong about that. Most of the time that low level law breaking is skipped. Allowing mass processing means you can politically target people if you so desire with ease. Because all you have to do is look them up on the computer. Third, computers don't forget. In the human world, when task force 1 flags you, you'll probably be forgotten. In the computer world you get a few black marks and that's it, you become ostracised from society. e.g. US's no fly list, China's social credit system. Fourth, the computer rules can be bad. A human "black box" can (and will) explain itself. A lot of computers cannot. The Dutch recently had a big scandal because they flagged anyone with a foreign name. The UK prosecuted thousands for fraud because of a computer glitch (yes, thousands, see point 1). Any rules generated from ML is generally racist and bigoted. Etc.
- jjk166 4y ago
  I get GP's point, I'm arguing it's incorrect. Specifically slow and expensive != rarely used. Once you accept that your privacy is going to be violated either way, the computer is the obviously preferable option. Human processing errors are not guaranteed to come out at court time, and humans can and will hide their mistakes. Getting caught in a "computer says no" trap fundamentally requires the human using the computer to also not be able to tell when the data is obviously wrong. Computers are not supposed to be the source of common sense in either system, they are a tool. There isn't 100x more crime that isn't being prosecuted due to lack of resources. There are lots of crimes like jaywalking that our legal system leaves to discretion. 99% of the time, a person crossing the street is a non-issue. Even if you do it right in front of a cop and it would take zero resources to prosecute, it doesn't get done because we as a society don't believe it should be. We make it a crime because there is that 1% of cases where it is a problem, and we need to be able to deal with those cases, but distinguishing which is which can not be readily codified in law. The only way that AI will change this is that the decision on whether to prosecute jaywalking will be consistent. I don't see why computers not forgetting is a bad thing. We already have the means of preserving data for long periods of time. The decision on whether records should be maintained or deleted should be made based on the circumstances, not on the faultyness of the hardware. I mean if you really wanted to, you could always program a computer to delete data at random, but it's just a dumb idea. Human rules can be bad. A human black box can not spit out an error log showing the particular sequence of neurons that activated that led to its decision. It can generate a statement, but whether that statement correctly explains it's decision making process can not be determined reliably. For example a human can shoot someone because of the color of their skin, and then state that it was because they felt they were in danger despite not actually feeling that way. Detecting such a glitch is extremely difficult, and no one has a reliable method of fixing it. Falsely suspecting people because their name is foreign is a well known bug in humans. Racism and bigotry is likewise a common issue with human neural networks. When you consider the countless multitudes of erroneous decisions made by humans over the millenia, an AI falsely flagging mere thousands is an immense improvement.

by jjk166 4y ago

msandford 4y ago

I believe that if a crime is sufficiently bad that we should prosecute for it, it absolutely merits the use of scarce resources. And if we can't muster up the resources to get humans to do the investigation, accusing, trying, etc it must not be that bad of a crime. Once algorithmic prosecution is enabled, we could easily see 100x as many accusations happen. That's the blessing of automation. 100x is actually a low bar. Many automations make millions more things possible. I don't think it would be a good world with a million times as much law enforcement with basically all of it automated.

jjk166 4y ago
Again, you're fundamentally missing what I'm saying. No one is arguing that humans require too many resources. I am saying that a computer, even if it were just as resource intensive as a human program, is a fundamentally preferable option. We do not live in a world where the more crime there is the more resources are available to deal with it. If anything, reality is the opposite: there is a vicious cycle where lack of government resources leads to crime and crime decreases available resources. It's trivially simple to imagine a scenario where crime gets so bad that we stop dealing with certain violations to conserve resources to deal with bigger issues. But that's irrelevant. What matters is that each and every person involved is a potential source of issues. Anyone can be tired, or scared, or distracted, or biased. This is not due to inadequate resources, though that may exacerbate the problem. This is the human condition - we are fallible beings. And in situations like this mistakes can ruin or end lives. Humans need oversight because of this fallibility, but the overseers are likewise fallible. The larger an organization is, the more places there are for something to fall through the cracks and for a mistake to go unnoticed and unresolved. Algorithmic prosecution might make accusations 100x cheaper, but you won't have 100x more accusations because there isn't 100x more crime to accuse people of. You're not going to see a massive spike in murders or bank robberies because of this algorithm, and the crime never having occurred is a really solid alibi. Sure if it's easier to investigate crimes then maybe some of the crimes we are currently tolerating because we simply don't have the manpower to deal with them would get investigated, but if we're ignoring a crime because we think the crime is not an issue, we can certainly keep ignoring it. The difference is the choice is made by society, not forced upon it by circumstance. You fear more law enforcement because you are plainly aware of how bad current law enforcement can be. This isn't more law enforcement, this is better law enforcement.
- msandford 4y ago
  I strongly disagree that increased law enforcement capabilities will lead to ever more perfect policing. I'm pretty sure that's what you're arguing. When we switched from toll booths with people in them to automated tag reading, this increased the ability to create toll roads. By making the cost of enforcement cheaper, it's enabling. If you make the cost of law enforcement lower, that won't make law enforcement work harder at solving murders. It'll mean that whole swaths of things that previously weren't practical to police will become practical and even economic. Especially if there's an automated fining system. Computers don't fix the flaws in systems. They magnify them, by enabling those systems to be scaled way, way up.

mattmanser 4y ago

You keep missing the GP's point. His point is: - computers are fast and cheap - humans are slow and expensive - slow and expensive = rarely used technique - rarely used technique = less chance of privacy invasion Therefore forced human processing protects privacy. It's a privacy concept that's quite often discussed. Pre-computers we had a certain level of privacy afforded to us that was due to the cost/speed of humans processing large amounts of data. Computers have stripped that away. It doesn't matter if there are human processing errors as that comes out at court time. Everything else you're saying is also kinda wrong, unfortunately. You aren't looking at the problems caused by computer efficiency, just somehow computers are magically better because they follow the same rules every time. Firstly, the data is usually bad and messy. Like super bad and messy. Computers can't tell when data is obviously wrong, so innocent people can get caught in "computer says no" trap. So you're wrong computers are better because of that. They have no common sense. Secondly, there ARE 100x more crime than get prosecuted. You're dead wrong about that. Most of the time that low level law breaking is skipped. Allowing mass processing means you can politically target people if you so desire with ease. Because all you have to do is look them up on the computer. Third, computers don't forget. In the human world, when task force 1 flags you, you'll probably be forgotten. In the computer world you get a few black marks and that's it, you become ostracised from society. e.g. US's no fly list, China's social credit system. Fourth, the computer rules can be bad. A human "black box" can (and will) explain itself. A lot of computers cannot. The Dutch recently had a big scandal because they flagged anyone with a foreign name. The UK prosecuted thousands for fraud because of a computer glitch (yes, thousands, see point 1). Any rules generated from ML is generally racist and bigoted. Etc.

jjk166 4y ago
I get GP's point, I'm arguing it's incorrect. Specifically slow and expensive != rarely used. Once you accept that your privacy is going to be violated either way, the computer is the obviously preferable option. Human processing errors are not guaranteed to come out at court time, and humans can and will hide their mistakes. Getting caught in a "computer says no" trap fundamentally requires the human using the computer to also not be able to tell when the data is obviously wrong. Computers are not supposed to be the source of common sense in either system, they are a tool. There isn't 100x more crime that isn't being prosecuted due to lack of resources. There are lots of crimes like jaywalking that our legal system leaves to discretion. 99% of the time, a person crossing the street is a non-issue. Even if you do it right in front of a cop and it would take zero resources to prosecute, it doesn't get done because we as a society don't believe it should be. We make it a crime because there is that 1% of cases where it is a problem, and we need to be able to deal with those cases, but distinguishing which is which can not be readily codified in law. The only way that AI will change this is that the decision on whether to prosecute jaywalking will be consistent. I don't see why computers not forgetting is a bad thing. We already have the means of preserving data for long periods of time. The decision on whether records should be maintained or deleted should be made based on the circumstances, not on the faultyness of the hardware. I mean if you really wanted to, you could always program a computer to delete data at random, but it's just a dumb idea. Human rules can be bad. A human black box can not spit out an error log showing the particular sequence of neurons that activated that led to its decision. It can generate a statement, but whether that statement correctly explains it's decision making process can not be determined reliably. For example a human can shoot someone because of the color of their skin, and then state that it was because they felt they were in danger despite not actually feeling that way. Detecting such a glitch is extremely difficult, and no one has a reliable method of fixing it. Falsely suspecting people because their name is foreign is a well known bug in humans. Racism and bigotry is likewise a common issue with human neural networks. When you consider the countless multitudes of erroneous decisions made by humans over the millenia, an AI falsely flagging mere thousands is an immense improvement.