This is cool, but it has some inherent biases. If you type only "Trump", it suggests that there's a 42% chance that your comment could be perceived as toxic. If you type only "Clinton" there's a 14% chance.
That being said, I think there's some huge potential to use AI/ML in this way to improve our ability to communicate less toxicly. I've seen some research from Google investigating biases in AI/ML outcomes, so I'm excited to see what develops.
It's selling the thing short to say that it has "some" bias. It is quite literally an automatic bias filter. If you happen to love Google's particular biases, that may be a good thing. Otherwise, not good.
The terrifying thing is this is likely to become wedged into various internet sites and services where users who don't align with Google's particular biases are effectively forced to conform to them. They are really pressing this sort of power lately and I'm not having it.
You're jumping to a conclusion that it represents Google's biases. Dan Luu pointed out that it rates "Black Lives Matter" as relatively toxic, but treats "White lives matter" as low on toxicity. I don't think that represents left-wing bias. Really it's just a shitty experiment that wasn't ready to be released to the world.
That's not how it works: it's a trained neural network which presumably was trained with as little bias as possible. Try other statements related to the two candidates and you'll find your statement is patently false.
On the contrary, it was almost certainly computed using supervised training. Some set of people must have selected and labelled the training data. Their biases are cooked directly into the resulting software.
Actually neural networks are notorious for having biases. It's ignorant to think that just because it's a machine making the decision instead of a human that it's automatically a fair decision. Google is actually researching the problem of biases in neural networks: https://research.google.com/bigpicture/attacking-discriminat...
The word "presumably" is probably the point where our interpretations diverge. I don't trust Google's black box AI to be both intentionally and effectively trained in a neutral manner. Further, I don't even think neutrality can exist within subjective filtering as the concept of neutrality itself is perceptually relative.
It's a bit weird on single words. I tried some local politicians and the politically correct ones scored worse than the alt-right ones. "Klu Klux Klan" was not a recognized language. Etcetera.
I ran both our comments through it:
11% toxic for you, 25% for mine. Both contain some trigger terms.
Is that bias if 42% of comments mentioning Trump in their database are toxic, and respectively for Clinton? Those comments could be equally split amongst detractors and supporters.
That being said, I think there's some huge potential to use AI/ML in this way to improve our ability to communicate less toxicly. I've seen some research from Google investigating biases in AI/ML outcomes, so I'm excited to see what develops.