5 ms·

If this is the case, that's tragic in the dramatic sense -- LGBT groups fighting algorithms that are protecting them, because they can't appreciate the unseen d

If this is the case, that's tragic in the dramatic sense -- LGBT groups fighting algorithms that are protecting them, because they can't appreciate the unseen danger prevented. I also imagine the 'algorithms' are a moving target, which makes me think this should be settled quickly if not dismissed, as Google tries to correct this. But certainly a lawsuit should bump this up the engineering priority list.

vorticalbox 7y ago
I hope that is a the case of it being sorted quickly though I am unsure they can teach intent well enough to actually make this any better.
jakelazaroff 7y ago
> If this is the case, that's tragic in the dramatic sense -- LGBT groups fighting algorithms that are protecting them, because they can't appreciate the unseen danger prevented. I don't think the article or GP are saying this. The algorithms, while ostensibly well-intentioned, are actually having the opposite effect — stifling the speech of the LGBTQ+ community while allowing their harassers to operate mostly unchecked.
- mushufasa 7y ago
  If they block 40% of real hate speech but also 5% of false positives (lgbt commenting on hate speech) I would say the algorithms are working overall. I have no way to know the real numbers but I am assuming Google monitors overall effectiveness of algorithms and only use what works.
  - monocasa 7y ago
    If they're having trouble classifying, they might not have good data on effectiveness.
    - mushufasa 7y ago
      You would have human review of a random sample to test effectiveness. In fact, this is usually how the ML algorithms are constructed in the first place, e.g. supervised learning. So yes I think they should have this.
  - jakelazaroff 7y ago
    Okay, but why do you think the numbers look anything like that? Conversely, if they block 5% of hate speech but are 40% false positives, then the algorithms are clearly not working. We have no reason to assume they're effective, and in fact this lawsuit contends that they're actually harmful.
jacksnipe 7y ago
Well, the algorithms simply don't work if they're flagging too many false positives.

vorticalbox 7y ago

I hope that is a the case of it being sorted quickly though I am unsure they can teach intent well enough to actually make this any better.

jakelazaroff 7y ago

> If this is the case, that's tragic in the dramatic sense -- LGBT groups fighting algorithms that are protecting them, because they can't appreciate the unseen danger prevented. I don't think the article or GP are saying this. The algorithms, while ostensibly well-intentioned, are actually having the opposite effect — stifling the speech of the LGBTQ+ community while allowing their harassers to operate mostly unchecked.

mushufasa 7y ago
If they block 40% of real hate speech but also 5% of false positives (lgbt commenting on hate speech) I would say the algorithms are working overall. I have no way to know the real numbers but I am assuming Google monitors overall effectiveness of algorithms and only use what works.
- monocasa 7y ago
  If they're having trouble classifying, they might not have good data on effectiveness.
  - mushufasa 7y ago
    You would have human review of a random sample to test effectiveness. In fact, this is usually how the ML algorithms are constructed in the first place, e.g. supervised learning. So yes I think they should have this.
- jakelazaroff 7y ago
  Okay, but why do you think the numbers look anything like that? Conversely, if they block 5% of hate speech but are 40% false positives, then the algorithms are clearly not working. We have no reason to assume they're effective, and in fact this lawsuit contends that they're actually harmful.

jacksnipe 7y ago

Well, the algorithms simply don't work if they're flagging too many false positives.