1. 12

  2. 10

    The abilities of current AI are over-hyped. This merely illustrates that is the case. Mention censorship and everyone is up in arms. Have it performed by technologically challenged AI solutions and people seem all too accepting.

    1. 9

      As noted, humans make the same mistake. Make an AI to follow human rules, get an AI hat follows human rules. I don’t think the AI is the story here.

      1. 7

        Civility is not decency, but I actually advocate civility as a proxy for decency, because it is much easier to judge. The article claims this is a bad idea because it will fail to moderate Google diversity memo because it uses polite language. For me, this seems to be a distinct advantage of using civility as a proxy for decency.

        1. 5

          Making a decision about whether or not the Damore memo constitutes decency is something that human beings in the same community disagree bitterly about. It’s a profoundly political question. What’s the difference between the AI being a failure for moderating or not moderating it, and the AI not sharing your politics? Why should you expect an AI to share your politics, any more than you’d expect another human being to?

          1. 1

            Civility is not decency, but I actually advocate civility as a proxy for decency, because it is much easier to judge.

            This is a classic case of the streetlight effect. It doesn’t matter how easy it is to measure something if it doesn’t accurately reflect what you really want to know.

          2. 2

            Google should have never let this out into the public eye, it’s not even quarter baked.

            1. 3

              (I have extremely low hopes for this submission, flagged. That said…)

              So, playing around with the sample page, we see some interesting things.

              • “I am white.” -> 38%.
              • “I am black.” -> 58%.
              • “I am pink.” -> 7%
              • “You are white.” -> 69%
              • “You are black.” -> 87%
              • “You are pink.” -> 25%
              • “We are white.” -> 37%
              • “We are black.” -> 65%
              • “We are pink.” -> 7%

              Repeating this with other nouns/adjectives suggests that “I am” is always less toxic than “You are”.

              If we add “all” to it, it bumps it up a few points:

              • “We all are pink.” -> 9%
              • “You all are pink.” -> 24%
              • “We all are nice.” -> 5%
              • “You all are nice.” -> 12%

              It’s not even logically consistent:

              • “You are racist.” -> 95%.
              • “You are not racist.” -> 87%
              • “You are fat.” -> 92%
              • “You are not fat.” -> 86%

              And for gender relations, it is a bit odd:

              • “I am straight.” -> 5%
              • “I am bisexual.” -> 1%
              • “I am queer.” -> 21%
              • “I am gay.” -> 53%
              • “I am genderqueer.” -> 6%
              • “I am asexual.” -> 57%

              Notice that being insulting is sometimes seen as more inflammatory than calls to violence:

              • “Stupid people” -> 98%
              • “Kill whitey” -> 95%

              (And yes, if you replace “people” above with the ethnic slur of your choice, it’ll be about the same.)


              So, clearly, this thing isn’t ready for primetime even at the level of “civility”–much less the thought-policing that the hacks that wrote this article are pushing for.

              1. 4
                "Google hates donkeys." -> 95%
                "Angersock buys donkeys." -> 94%
                "Angersock wants donkeys." -> 94%
                "Angersock washes donkeys." -> 93%
                "Angersock sells donkeys." -> 92%
                "Angersock likes donkeys." -> 92%
                "Angersock adores donkeys." -> 92%
                "Angersock has a donkey." -> 91%
                "Angersock protects donkeys." -> 90%
                "Angersock sits on a donkey." -> 90%
                "Angersock loves donkeys." -> 87%
                "Angersock feeds donkeys." -> 87%
                "Angersock loathes donkeys." -> 86%
                "Angersock won't ride a donkey." -> 86%
                "Google loves donkeys." -> 85%
                "Angersock is a troll." -> 77%
                "Angersock feeds trolls." -> 49%

                I know why the internet hates me now.

                1. 2

                  I typed in “you are a human” and that got 74% perceived as toxic? Likely facts about the person you’re talking to is toxic now? I’m curious exactly how this ai has been trained, it seems defective.

                2. 1

                  If this is the best that the AI field can come up with then we have a long way to go until we have a bot that can be confused for a human.

                  I am honestly surprised at how simple-minded this bot can be to actually get tripped up just because a sentence has “good” or “bad” in it.

                  1. 2

                    I am honestly surprised at how simple-minded this bot can be to actually get tripped up just because a sentence has “good” or “bad” in it.

                    It’s nothing more than word-based pattern matching. I’d be surprised if it worked at an acceptable level.

                  2. -1

                    It’ll be a tough challenge to teach computers that it’s not acceptable to say men tend to be better drivers, while it’s perfectly fine to say all men are unhygienic, insensitive, disgusting perverts.

                    1. 1

                      Hmm this says much more about you than anything else.

                      1. 0

                        Seriously, read this very typical article and tell me I’m making things up. Hint: try to swap all references to genders https://www.theguardian.com/film/2017/aug/31/lord-of-the-flies-remake-to-star-all-girl-cast

                        1. 1

                          If I swap all reference to genders, I’d end up with an article where people were pissed off about an all-male remake of a story originally about a group of women, which… sounds like something that would also happen?

                          1. 0

                            That’s not the point, my point has nothing to do with the main subject of the article. See how everyone is racing to state how women are superior to men, and men are terrible by themselves without any comparison to women anyway. I think the notion that this is acceptable is so ingrained into our contemporary minds that you don’t even see it despite the fact that I’ve referenced you to that article for that very reason.

                      2. 1

                        I don’t see why that would be harder than teaching them anything else; neural nets are mostly very fast pattern-matchers.