A refusal of service need not be classifiable down to the exact wording of the rules. If it did, then a person could wear a scarf and no shirt, and claim the scarf was a short shirt, and demand service in a 'no shirt no service' restaurant. The restaurant would presumably have to come up with an exact minimum length of shirt which qualifies, and that's just plain silly.
Classifiable was meant as being able to name a requirement without digressing into a discussion about its inherent properties. Eg. "no shirt no service" restaurant actually says "SHIRT" and not "some garments we'll arbitrarily classify not befitting our restaurant". In case of discussion I guess ultimately a 3rd party or a court can judge on what constitutes a shirt.
The rules need to be intelligible and consistent, yes. Facebook's rules are. They refuse "content that is hate speech, threatening or pornographic; incites violence; or contains nudity or graphic or gratuitous violence." There is nothing unintelligible here
Facebook's STATED rules are indeed intelligible and consistent however the crux of my post was that a NN or DL AI system has no idea what these rules are or wouldn't be able to explain or argue why it classified something such or so. Even worse, even the trainers of the system wouldn't be able to explain why something was classified a certain way. The best it can do is give an inter-class certainty (https://codewords.recurse.com/issues/five/why-do-neural-networks-think-a-panda-is-a-vulture). The additional problem here is that if a classification problem arises it won't be about what humans would define as edge cases or judgement calls. It rather amounts to a total freakout.
Consistency is another matter, and if you can show they deliberately permit certain examples while excluding others then you have something here. They do NOT need to have their AI catch all possible examples of violations to be considered consistent or intelligible, nor do they need to publish their algorithms. If they're making a good-faith attempt, good for them. It's only when they maliciously apply their rules inconsistently that a problem arises.
I assumed it was clear that "refusal of service" would rather come down to the AI NOT allowing things that should perfectly acceptable according to Facebook's stated rules rather than the reverse.
Not sure if good faith extends to an AI agent, but I don't think so. In theory (IANAL) if it can be proven that the AI refuses service outside of Facebook's stated rules WITHOUT even being able to explain/argue WHY then one must assume intent or at least bad faith / negligence from the AI system owner (?)
Again, I'm more concerned about legal implications within the brave new AI world than the actual free speech limiting FB blabla... The same arguments can be made for other AI systems like self driving cars. There is no intelligence, opinion or human judgement that can mount a "good faith defense" or even be "forgiven" or "understood". An AI also can't claim insanity...
If my Googling concerning "The Right to Refuse Service" laws in the US is correct, it is not legal to refuse service outside of the law (anti-discrimination laws on multiple levels) or arbitrarily or inconsistently. Focusing on the latter two, this means that any refusal of service must be "classifiable" or in other words there must be a set of lawful "refusal rules" that CAN be adhered to BEFORE requesting the service. In as far as I understand neural networks and deep learning that requirement isn't met by this Facebook system. There isn't a certainty based on human intelligible rules that service will or won't be granted. The rules stated by Facebook aren't actually the rules that govern the AI making the decision to grant or deny service. The actual rules (weights) that govern that system are actually unknown, it doesn't really "know" the rules, it performs a function that amounts more to "like this" with "this margin". Neither the "like this" nor "the margin" are human intelligible. Before people start saying 99.9% etc. please remind yourself that there is a big difference between a "human making an error in judgement" and an "unaccountable AI that freaks out without actually knowing why" in the eyes of the law.
Technically the same argument holds for possible illegality of self driving cars with a NN or DL AI system. The system can't tell me WHY a certain action is OK or NOT OK. But that is a different story.
Slashdot Mirror
As the subject states, no more, no less.
A refusal of service need not be classifiable down to the exact wording of the rules. If it did, then a person could wear a scarf and no shirt, and claim the scarf was a short shirt, and demand service in a 'no shirt no service' restaurant. The restaurant would presumably have to come up with an exact minimum length of shirt which qualifies, and that's just plain silly.
Classifiable was meant as being able to name a requirement without digressing into a discussion about its inherent properties. Eg. "no shirt no service" restaurant actually says "SHIRT" and not "some garments we'll arbitrarily classify not befitting our restaurant". In case of discussion I guess ultimately a 3rd party or a court can judge on what constitutes a shirt.
The rules need to be intelligible and consistent, yes. Facebook's rules are. They refuse "content that is hate speech, threatening or pornographic; incites violence; or contains nudity or graphic or gratuitous violence." There is nothing unintelligible here
Facebook's STATED rules are indeed intelligible and consistent however the crux of my post was that a NN or DL AI system has no idea what these rules are or wouldn't be able to explain or argue why it classified something such or so. Even worse, even the trainers of the system wouldn't be able to explain why something was classified a certain way. The best it can do is give an inter-class certainty (https://codewords.recurse.com/issues/five/why-do-neural-networks-think-a-panda-is-a-vulture). The additional problem here is that if a classification problem arises it won't be about what humans would define as edge cases or judgement calls. It rather amounts to a total freakout.
Consistency is another matter, and if you can show they deliberately permit certain examples while excluding others then you have something here. They do NOT need to have their AI catch all possible examples of violations to be considered consistent or intelligible, nor do they need to publish their algorithms. If they're making a good-faith attempt, good for them. It's only when they maliciously apply their rules inconsistently that a problem arises.
I assumed it was clear that "refusal of service" would rather come down to the AI NOT allowing things that should perfectly acceptable according to Facebook's stated rules rather than the reverse. Not sure if good faith extends to an AI agent, but I don't think so. In theory (IANAL) if it can be proven that the AI refuses service outside of Facebook's stated rules WITHOUT even being able to explain/argue WHY then one must assume intent or at least bad faith / negligence from the AI system owner (?) Again, I'm more concerned about legal implications within the brave new AI world than the actual free speech limiting FB blabla... The same arguments can be made for other AI systems like self driving cars. There is no intelligence, opinion or human judgement that can mount a "good faith defense" or even be "forgiven" or "understood". An AI also can't claim insanity...
If my Googling concerning "The Right to Refuse Service" laws in the US is correct, it is not legal to refuse service outside of the law (anti-discrimination laws on multiple levels) or arbitrarily or inconsistently. Focusing on the latter two, this means that any refusal of service must be "classifiable" or in other words there must be a set of lawful "refusal rules" that CAN be adhered to BEFORE requesting the service. In as far as I understand neural networks and deep learning that requirement isn't met by this Facebook system. There isn't a certainty based on human intelligible rules that service will or won't be granted. The rules stated by Facebook aren't actually the rules that govern the AI making the decision to grant or deny service. The actual rules (weights) that govern that system are actually unknown, it doesn't really "know" the rules, it performs a function that amounts more to "like this" with "this margin". Neither the "like this" nor "the margin" are human intelligible. Before people start saying 99.9% etc. please remind yourself that there is a big difference between a "human making an error in judgement" and an "unaccountable AI that freaks out without actually knowing why" in the eyes of the law. Technically the same argument holds for possible illegality of self driving cars with a NN or DL AI system. The system can't tell me WHY a certain action is OK or NOT OK. But that is a different story.