Thread by @wiczipedia

Dear lord. There is so much wrong with how this study was conducted. Let us count the ways. (THREAD)

https://twitter.com/TwitterSafety/status/1638256122942353422

1. As we found in our #MalignCreativity report @TheWilsonCenter, there is an entire movement on Twitter and other sites by abusers to evade detection from AI / content moderators, so a list of "300 commonly used English-language slurs" ain't conna cut it. wilsoncenter.org/publication/ma…

2. @Twitter's focus on impressions shows they are not interested in actually reducing harm, just reducing how it *looks.* It doesn't matter if a tweet calling for your rape and hanging gets 10 impressions or 10000, they both are harmful to the target.

https://twitter.com/TwitterSafety/status/1638262034864263188?s=20

3. That means, yes, that the number of Tweets containing slurs matters! A company that purports to be a digital public square should care whether a large proportion of the content on its site is hate speech!!! Why is that controversial?

4. This is to say nothing of the fact that Twitter is used in other languages than English. Let's not forget that a lot of the most damaging content in non-white, non-English speaking contexts slips under the rug in most content moderation.

5. As @Esqueer_ points out, a lot of abuse and hate speech on Twitter is not text, but image-based. This study doesn't account for that at all.

https://twitter.com/Esqueer_/status/1638529971470696453?s=20

6. Finally, I strongly object to the idea that AI save us from, or even detect the majority of hate speech and abuse we see online. This content is contextualized. It is memefied. On its surface, it might look like nothing, even to a human outsider... (cont)

...but taken as a whole, in an onslaught of thousands of tweets, by the target of the abuse, the effect is staggering.

Machines can't win the day on this one, and I would argue in the context of online abuse they are actively undermining our understanding of the problem. /end