Attempts to automate the policing of hate speech are doomed to ludicrity

See update for November 2, 2022, at end of post.

One of the first things an undergraduate in Communication learns is that language is arbitrary, ambiguous, and abstract.

This is a problem when seeking to police hate speech, especially for artificial intelligence idiots, and likely explains much of the idiocy I have described previously.^[1] To explain why, I will explore each of those terms used to describe language to Communication undergraduates.

The first, most fundamental point is that computers do not deal with ambiguity. In a binary way of looking at the world, things are something or they are not; they cannot be both and there can be no nuance between one pole or the other. It is a strange and distorting way of looking at the world. When humans do this with other humans, feminists may often call it hierarchically invidious monism: It divides humanity into rich and poor, white and of color, healthy and unhealthy, male and female, young and old, and the list goes on, endlessly.^[2]

The key point to these pairings is that there is a bias, even a bigotry, favoring one over the other,^[3] and this is why Elizabeth Minnich rejected the term ‘binary’ to describe them.^[4] It is unlikely there is any limit to them: Simone de Beauvoir argued, I think correctly, that even if we somehow managed to eliminate the biases we now have, we would construct new ones.^[5]

There are numerous implications to hierarchically invidious monism, some of which I’ve dealt with elsewhere, nearly all of which I won’t deal with here, but the important points here are 1) that this way of looking at the world is fallacious, and 2) that computers have no other way of evaluating data in moral terms, which is to say 3) that the way that computers will test for hate speech will inherently be unavoidably fallacious.

The next point to consider here is that language, including hate speech, is arbitrary. What makes the n-word or the b-word (rhyming with ‘witch’) so offensive? There is a historical context of unbridled bigotry to the n-word and the second word reduces women to—yes, this is what it meant once upon a time—female dogs,^[6] carrying the connotation of an unattractive woman, undesirable under the male gaze for mating.

The connection we make between those words and what they represent is what makes the language abstract and the terms chosen to represent those meanings are arbitrary. Absent any knowledge of meaning or context, the words themselves would not be offensive. It is the meanings that we collectively have assigned those words that makes them so. And when people say they are ‘reclaiming’ certain hateful words, as with a “slut walk,” they are asserting an authority they do not in fact possess, because no one has that authority, to reconfigure these representations.

Which brings us to the point that language evolves organically. When I was a kid, for example, “dead naming” wasn’t even a thing. But today, if I refer to Caitlyn Jenner as “Bruce” or Chelsea Manning as “Bradley,” I am “dead naming” them, using the names assigned with their previous genders. This denies their identities and, therefore, them as human beings. Denying people’s humanity lies at the core of hate speech.

But a computer can only recognize certain words, the n-word and the b-word, for examples, as “hate speech.” Even humans have trouble keeping up with the evolution of hate speech and determining what is and is not hate speech is problematic: Some Jews, for example, seemingly see anti-Semitism everywhere they look, in nearly every utterance referring to Jews. It is probably not the case that all such speech is hate speech but there can be no question that anti-Semitism is real, has a long history, and must be considered in any attempt to address hate speech or hate generally.

How can we possibly instruct a computer to properly recognize evolving ambiguous speech as hate speech? The simple answer is that we can’t and that computers cannot resolve controversies—advocates for Palestinians obviously disagree with some Jews when we insist, for example, that anti-Zionism is not anti-Semitism—over what is and is not hate speech.

This dooms any automated attempt to police hate speech, which needs to be done thoughtfully and intelligently, to ludicrous results, as I’ve previously observed,^[7] which in an era of social media, raises the difficult (because I don’t have the answer) question of how one does this at scale.

Update, November 2, 2022:

“The current situation is unpredictable and chaotic, and bad actors and unsafe behaviors can thrive in such an environment,” Interpublic Group of Cos. wrote in an email Monday that was reviewed by The Wall Street Journal. “At this moment, we cannot confidently state that Twitter is a safe place for brands.” . . .

After he closed the $44 billion purchase of the social-media company last week, Mr. [Elon] Musk wrote on Twitter that the site “cannot become a free-for-all hellscape, where anything can be said with no consequences!”

The Interpublic email cited research that showed “a rise in inappropriate behavior” on the site since the company was taken private. “It’s reasonable to expect that bad actors may feel emboldened given Musk’s statements about being a free speech absolutist and his more recent sharing of conspiracy theories,” the email read. Interpublic said it would work with clients who had non-cancelable Twitter contracts to modify their contract terms.^[8]

One of my concerns about Elon Musk and his acquisition of Twitter, as I have with high technology executives generally, is a disgraceful faith in artificial intelligence idiocy, which has already demonstrated its unfitness for purpose in content moderation^[9] and is specifically unfit for policing hate speech.^[10]

Neither Elon Musk nor anyone else wave a magic wand and suddenly make fallacies valid,^[11] but high technology executives press on regardless, as if they simply said, “Damn the torpedoes, full speed ahead!”^[12] Add to this Musk’s own record on so-called “free speech”^[13] in combination with looming job cuts,^[14] and the characterization of “[t]he current situation [at Twitter] [a]s unpredictable and chaotic”^[15] begins to look charitable far beyond recognition.

1. ^[1]David Benfell, “Our new Satan: artificial idiocy and big data mining,” Not Housebroken, April 5, 2021, https://disunitedstates.org/2020/01/13/our-new-satan-artificial-idiocy-and-big-data-mining/↩
2. ^[2]Simone de Beauvoir, “Woman as Other,” in Social Theory, ed. Charles Lemert, 6th ed. (Philadelphia: Westview, 2017), 268-270; Lorraine Code, What Can She Know? Feminist Theory and the Construction of Knowledge (Ithaca, NY: Cornell University, 1991); Elizabeth Kamarck Minnich, Transforming Knowledge, 2nd ed. (Philadelphia: Temple University, 2005).↩
3. ^[3]Lorraine Code, What Can She Know? Feminist Theory and the Construction of Knowledge (Ithaca, NY: Cornell University, 1991); Elizabeth Kamarck Minnich, Transforming Knowledge, 2nd ed. (Philadelphia: Temple University, 2005).↩
4. ^[4]Elizabeth Kamarck Minnich, Transforming Knowledge, 2nd ed. (Philadelphia: Temple University, 2005).↩
5. ^[5]Simone de Beauvoir, “Woman as Other,” in Social Theory, ed. Charles Lemert, 6th ed. (Philadelphia: Westview, 2017), 268-270.↩
6. ^[6]Oxford Dictionary of English, 3rd ed., s.v. “bitch.”↩
7. ^[7]David Benfell, “Our new Satan: artificial idiocy and big data mining,” Not Housebroken, April 5, 2021, https://disunitedstates.org/2020/01/13/our-new-satan-artificial-idiocy-and-big-data-mining/↩
8. ^[8]Suzanne Vranica and Patience Haggin, “Ad Giants Advise Brands to Pause Spending on Elon Musk’s Twitter,” Wall Street Journal, November 1, 2022, https://www.wsj.com/articles/ad-giants-advise-brands-to-pause-spending-on-elon-musks-twitter-11667333021↩
9. ^[9]David Benfell, “Our new Satan: artificial idiocy and big data mining,” Not Housebroken, April 5, 2021, https://disunitedstates.org/2020/01/13/our-new-satan-artificial-idiocy-and-big-data-mining/↩
10. ^[10]David Benfell, “Attempts to automate the policing of hate speech are doomed to ludicrity,” Not Housebroken, November 1, 2022, https://disunitedstates.org/2022/11/01/attempts-to-automate-the-policing-of-hate-speech-are-doomed-to-ludicrity/↩
11. ^[11]David Benfell, “Our new Satan: artificial idiocy and big data mining,” Not Housebroken, April 5, 2021, https://disunitedstates.org/2020/01/13/our-new-satan-artificial-idiocy-and-big-data-mining/; David Benfell, “Attempts to automate the policing of hate speech are doomed to ludicrity,” Not Housebroken, November 1, 2022, https://disunitedstates.org/2022/11/01/attempts-to-automate-the-policing-of-hate-speech-are-doomed-to-ludicrity/↩
12. ^[12]David Glasgow Farragut, quoted in U.S. Navy, “‘Damn the Torpedoes – Full Speed Ahead’: Navy’s First Admiral Was Hispanic Hero,” September 15, 2020, https://www.navy.mil/Press-Office/News-Stories/Article/2347790/damn-the-torpedoes-full-speed-ahead-navys-first-admiral-was-hispanic-hero/↩
13. ^[13]David Benfell, “Elon Musk’s ‘free speech,’” Not Housebroken, April 28, 2022, https://disunitedstates.org/2022/04/28/elon-musks-free-speech/↩
14. ^[14]Alexa Corse and Salvador Rodriguez, “Twitter Is Drafting Broad Job Cuts in Whirlwind First Weekend Under Elon Musk,” Wall Street Journal, October 30, 2022, https://www.wsj.com/articles/twitter-is-drafting-broad-job-cuts-days-after-elon-musks-takeover-11667143301↩
15. ^[15]Interpublic Group, quoted in Suzanne Vranica and Patience Haggin, “Ad Giants Advise Brands to Pause Spending on Elon Musk’s Twitter,” Wall Street Journal, November 1, 2022, https://www.wsj.com/articles/ad-giants-advise-brands-to-pause-spending-on-elon-musks-twitter-11667333021↩