4 ms·

Unintended side effects. There is no way to censored a collection as large as YouTube manually. And there are no high accuracy automated filter.

amarant 8y ago
Which is why this kind of censorship is a bad idea
belorn 8y ago
Why would there be no no high accuracy automated filter? We have those in a multitude of different technologies, and it is somehow here is supposed to be impossible to balance false positive and false negatives? People who use youtube filtering has no negative impact of false positives and thus you get filters that reflect that. I would call that an intentional design.
- AstralStorm 8y ago
  Generally, the problems are multifold. 1) accuracy at direct audio detection is limited and fuzzy - with best current approaches. Even matching human performance at this classification task is not enough and we're not there yet. 2) the filter is supposed to take quotation and other fair use into account 3) covers and remixes make matters even more complex 4) the copyrighted matter is both performance and music itself, with different laws Therefore there will be many false positives or negatives.
  - candiodari 8y ago
    Furthermore, adversaries get to directly test against the filter. It's like writing a computer virus. Anti-virus software does not help against even rewritten viruses because every virus writer tests every virus against every antivirus they get their hands on, and they retry, changing something, until it passes. Therefore, no antivirus ever catches a new virus. Same problem with copyright filters. Plenty of bad actors spend days, weeks or months finding (even automated) ways to bypass them.
  - belorn 8y ago
    We have a technology that share most of the same problems but which has the opposite bias with low false positive rates: bio metric security. 1) biometric detection is limited and fuzzy - with best current approaches. Even matching human performance at this classification task is not enough. 2) biometric detection is supposed to take some natural variation into account. 3) lighting, temperature, dust and other environmental factors. No biometric security is perfect and yet it is in common use everywhere, including passports. Such design could have any rate of false positive and negatives but since the impact of false negatives is less than false positive the configuration is usually such that false positives are reduce down as much as possible, at the cost of false negatives. Can you say that the same is done with media filtering? same goes to candiodari. Do not biometric security have security testing done by adversaries, which they can retry, changing something, until it successfully break?
wongarsu 8y ago
With 400 hours of video uploaded every minute [1], and an average video length of 4 minutes 20 seconds [2], we have an average of 5540 videos uploaded per minute. Assuming 5% of videos get flagged by the algorithm for possible copyright problems, and each human review takes 30 seconds (the algorithm flags a specific timestamp), you only need 415.5 employees (0.4% of current google employees) to do manual human review of every claim. Of course google has no reason to do that. They face high potential damage from taking down too few videos, and very little damage outside of small communities for taking down too much. So they take down any match automatically and drag their feet on the dispute process instead. 1: https://en.wikipedia.org/wiki/YouTube https://en.wikipedia.org/wiki/YouTube 2: https://www.minimatters.com/youtube-best-video-length/ https://www.minimatters.com/youtube-best-video-length/
- AstralStorm 8y ago
  5% is too low an estimate. 10% is more realistically achievable with best matching systems. (Source: MIREX 2017) Manual evaluation rate is also assumed to be 100% which it is not even close to especially based on a short snippet.

amarant 8y ago

Which is why this kind of censorship is a bad idea

belorn 8y ago

Why would there be no no high accuracy automated filter? We have those in a multitude of different technologies, and it is somehow here is supposed to be impossible to balance false positive and false negatives? People who use youtube filtering has no negative impact of false positives and thus you get filters that reflect that. I would call that an intentional design.

AstralStorm 8y ago
Generally, the problems are multifold. 1) accuracy at direct audio detection is limited and fuzzy - with best current approaches. Even matching human performance at this classification task is not enough and we're not there yet. 2) the filter is supposed to take quotation and other fair use into account 3) covers and remixes make matters even more complex 4) the copyrighted matter is both performance and music itself, with different laws Therefore there will be many false positives or negatives.
- candiodari 8y ago
  Furthermore, adversaries get to directly test against the filter. It's like writing a computer virus. Anti-virus software does not help against even rewritten viruses because every virus writer tests every virus against every antivirus they get their hands on, and they retry, changing something, until it passes. Therefore, no antivirus ever catches a new virus. Same problem with copyright filters. Plenty of bad actors spend days, weeks or months finding (even automated) ways to bypass them.
- belorn 8y ago
  We have a technology that share most of the same problems but which has the opposite bias with low false positive rates: bio metric security. 1) biometric detection is limited and fuzzy - with best current approaches. Even matching human performance at this classification task is not enough. 2) biometric detection is supposed to take some natural variation into account. 3) lighting, temperature, dust and other environmental factors. No biometric security is perfect and yet it is in common use everywhere, including passports. Such design could have any rate of false positive and negatives but since the impact of false negatives is less than false positive the configuration is usually such that false positives are reduce down as much as possible, at the cost of false negatives. Can you say that the same is done with media filtering? same goes to candiodari. Do not biometric security have security testing done by adversaries, which they can retry, changing something, until it successfully break?

wongarsu 8y ago

With 400 hours of video uploaded every minute [1], and an average video length of 4 minutes 20 seconds [2], we have an average of 5540 videos uploaded per minute. Assuming 5% of videos get flagged by the algorithm for possible copyright problems, and each human review takes 30 seconds (the algorithm flags a specific timestamp), you only need 415.5 employees (0.4% of current google employees) to do manual human review of every claim. Of course google has no reason to do that. They face high potential damage from taking down too few videos, and very little damage outside of small communities for taking down too much. So they take down any match automatically and drag their feet on the dispute process instead. 1: https://en.wikipedia.org/wiki/YouTube https://en.wikipedia.org/wiki/YouTube 2: https://www.minimatters.com/youtube-best-video-length/ https://www.minimatters.com/youtube-best-video-length/

AstralStorm 8y ago
5% is too low an estimate. 10% is more realistically achievable with best matching systems. (Source: MIREX 2017) Manual evaluation rate is also assumed to be 100% which it is not even close to especially based on a short snippet.