Why AI Won’t Solve Content Problems

Context is complex and everyone needs to understand it.

Figure 1: A depiction of cancel culture in 1835 where pro-slavery assholes demolished a printing press exercising the freedom to print abolitionist content. Apparently “freedom” was utterly lost on these folks violently attacking three offices and shooting E. P. Lovejoy because he would not surrender his pulpit.

Wired published this on Monday: “Why a YouTube Chat About Chess Got Flagged for Hate Speech.”¹ It’s a great example how the complexity of language leads to the continued modeling of broken algorithms. If “black,” “white,” and “attack” are trigger words resulting in immediate content deletion of a chess discussion, then where does healthy political debate rank? Who gets to decide if political debate is appropriate or hateful? I believe content authors have a responsibility to provide context and disclose their own biases. In a perfect world, such disclosures would help facilitate civil debate.²

Cancelling works both ways and consequences of real (i.e. not imagined) hate speech is situational and contextual. This piece is to help you understand more about context and why we can’t automate hate speech away. Too many legitimate conversations need to happen without some algorithm cutting people out. I’m on the side of reason and judgement.

I know this is a long article but I ask you to bear with me as I tackle a complex topic.

Context is anything that affects the cognitive processing of content or information

What Context Is

I’ve been studying the concept of context for nearly ten years, but have been working professionally with content for over 25 years and since college I’ve been a reader of rhetorical philosophy. Currently, I’m in a protracted process of writing a book about it. Trust me when I assert the complexity of context itself is elusive for rhetoricians and researchers:

[I]t does not seem possible at the present time to give a single, precise, technical definition of context, and eventually we might have to accept that such a definition may not be possible. At the moment the term means quite different things within alternative research paradigms, and indeed even within particular traditions seems to be defined more by situated practice, by use of the concept to work with particular analytic problems, than by formal definition. From our perspective, lack of a single formal definition, or even general agreement about what is meant by context is not a situation that necessarily requires remedy. Instead the fact that so many investigators recognize the the importance of context and are actively involved in trying to unravel how it works is why this concept provides such a productive focus for study at the present time.³

The emphasis is theirs.

All of this is to say that this issue isn’t black and white (pun firmly used with dad joke confidence).

I define context as this: anything that affects the cognitive processing of content or information.⁴

Figure 2: Context swirls around us in an infosphere.

I propose that context is a fully interactive experience with language and environment. Even though our brains are hardwired for language and cooperation from birth, it took millions of years of evolution.⁵ What makes us think we can conquer context via a derivative intelligence system at the very dawn of our information age and monetize it before the next quarter?

I’m no luddite but technology cannot solve everything.

Besides, the human brain and computer memory and cognitive processing are just not the same. At all.⁶

More on that comparison in a future article.

Language and Complexity

Language doesn’t have the words to define itself properly. Much of what is written about context is academic and uses jargon like “metapragmatism” to help solidify the nebulous. Those of us who enjoy reading and researching linguistics and rhetoric will wade through dense text. Most people will not. That’s probably you.

If humans cannot agree on what context actually is and what it means for real world applications, why build algorithms based on vague and theoretical thinking? This is putting the cart in another continent and hoping the horse will swim across the ocean.

Additionally, social media has so much context removed, people apologize all the time for ill-timed tweets and Facebook posts that may or may not have been intentionally offensive. AI as a content babysitter is a Hype Machine used to stop people from complaining about assholes.

Then there is the emerging problem of privacy and data sampling. How do companies police content when they are starting to comply with new privacy laws? Practitioners should be questioning the ethics of data mining for the sole purpose of training AI.⁷

AI alone will not fix our social problems

YouTube Wasn’t Built for This

YouTube banning a chess discussion shows how little developers and business leaders at Google are paying attention to the very real-life discourse happening all around them. It’s as if somebody said, “Hey, I know! AI! Just delete everything that matches the racist vocabulary in this text file!”

To this day, YouTube has no contextual requirements for authors to categorize their work. I’ve had my own content taken down by a DMCA request. It was a very short clip from Family Guy because I wrote a now unpublished article on how cartoons do not like the name Keith (see also this). I was using satire and that falls well within fair use.

Imagine satire being used to make a point about racism. Humor doesn’t even work with all humans all of the time. How can an algorithm understand irony, sarcasm, and just plain snark?

Instructors and content authors who use YouTube to simply stream content can’t provide enough context to stop automated DMCA takedowns, even if they are legal. On top of that, there is an AI content-checking trigger words. Good god.

Why All of this Really Matters

Currently there is a national discourse regarding gender, identity, history, systemic bias, and racism. I wouldn’t call it a reckoning as many have⁸ because there don’t seem to be any actionable outcomes other than awareness. (However, awareness is the first step toward change. This is another topic I’ll save for a later posting. But regardless of the platform, the discourse is important and should continue.)

From my personal experience, most developers are usually not subject matter experts and are put in the position of coding to requirements that could be decided in meetings based on any number of catalysts. These catalysts could include negative media coverage of the company or a truly vile piece of content that goes viral. Those one-offs do not represent the totality of users on any given platform.

I won’t deny controversy attracts trolls. They dox people. They attack authors they disagree with. They see everything through political prisms. They may even be simply motivated by sowing chaos itself. But I cannot emphasize this enough:

This is a social problem not a technical puzzle.

Figure 3: The ENIAC Computing system. Look at all of the complexity for just what are now rudimentary calculations. Imagine AI as a Gordian knot of those cables trying to plug into the right outlets on their own.

Is it unreasonable to ask Big Tech to approach content monitoring by deliberate and exhaustive research before whipping up an AI and having a press release? Can we have diverse collectives of experts working the content moderation problem? If Google is anything to go by, then no.⁹ They can’t even manage their own ethics department, much less tackle the philosophy of rhetoric. They would have to slip that in before the next shareholder meeting’s ass-kissing exercises.

Lack of context because of mobile technology, social media, and misinformation is driving ignorance and inflaming what used to be civil discourse or even friendly disagreements. But I firmly believe no AI can do anything but cancel people and shut down debate.

Content is hard. Authors and editors are taken for granted. Content moderation is even harder. Humans and AI together can get closer to a solution. We cannot leave out educated people who understand language, culture, and have the pulse of important discourse.

AI alone will not fix our social problems.

End Notes

[1] Knight, Will. “Why a YouTube Chat About Chess Got Flagged for Hate Speech.” Wired.Com, March 1, 2021. https://www.wired.com/story/why-youtube-chat-chess-flagged-hate-speech/.

[2] Disclosure: I despise racist and hateful speech. I find most of our political discourse nauseating because many have lost the ability to do anything but sling slurs. I also operate on the assumption that we are all better together. This is my bias so you have more context to this article. By the way, disclosing your own personal bias is a great start to framing your arguments and initiating healthy discourse.

[3] Duranti, Alessandro, and Charles Goodwin. Rethinking Context: Language as an Interactive Phenomenon. Cambridge [England]; New York: Cambridge University Press, 1992.

In addition, they note:

However, there have been attempts at clarifying the terms and notions necessary for handling the different kinds of relations that linguistic signs entertain or establish with their context of use. The semiotically informed work by Silverstein (1985a, 1985b) on metapragmatics is one such example, cited below.
Silverstein, Michael. “The culture of language in Chinookan narrative texts; or. On saying that… in Chinook.” Grammar inside and outside the clause. Some approaches to theory from the field. J. Nichols y A. Woodbury (eds) Cambridge: Cambridge University Press. pp (1985): 132–172.

[4] Me. Revised version of my definition to be more inclusive of sensual (read: by our senses, you dirty minded people) input.

[4.5] My promise to you is there will always be Easter eggs in my end notes. Otherwise, what’s the use of citing work if everyone will just skip over what is indeed highly recommended reading? So look again please.

[5] Tomasello, Michael. Origins of Human Communication. Cambridge, Mass.: MIT Press, 2008.

[6] Lehar, Steven. The World in Your Head: A Gestalt View of the Mechanism of Conscious Experience. Mahwah, N.J: Lawrence Erlbaum Associates, Publishers, 2003.

[7] Vanian, Jonathan. “Using A.I. Can Be Risky Business.” Forbes. Com, February 9, 2021. https://fortune.com/2021/02/09/using-a-i-can-be-risky-business/.

[8] Google. “Google Search of the Terms: Racial Reckoning in America.” Search Engine. google.com (website), March 2, 2021. https://www.google.com/search?q=racial+reckoning+in+america.

[9] Fried, Ina. “Google Fires Another AI Ethics Leader.” Axios, February 19, 2021. https://www.axios.com/google-fires-another-ai-ethics-leader-6ef7dcd5-4583-4396-b5b3-129547ff3091.html.

Figures

[1] DLC American Anti-Slavery Society, Isaac Knapp, and S. W. Benedict. The American Anti-Slavery Almanac, for 1839, Being the Third After Bissextile or Leap-Year, And the 63d Of American Independence. Calculated for Boston: Adapted to the New England States. 1st ed. Vol. 1. 4. New York & Boston, 1839. p 11. https://archive.org/details/americanantislav1839chil/page/11/mode/1up?view=theater.

[2] Babbitt, Edwin D. The Principles of Light and Color : Including among Other Things the Harmonic Laws of the Universe, the Etherio-Atomic Philosophy of Force, Chromo Chemistry, Chromo Therapeutics, and the General Philosophy of the Fine Forces, Together with Numerous Discoveries and Practical Applications. 1st ed. New York: Babbitt & Co., 1878. p 481. https://archive.org/details/gri_c00033125011227010/page/n3/mode/2up.

[3]Anonymous. Glen Beck (Background) and Betty Snyder (Foreground) Program ENIAC in BRL Building 328. ca 1955 1947. Photograph. https://en.wikipedia.org/wiki/ENIAC#/media/File:Eniac.jpg.

Edits

Updated on 3/5/2021 to clean up the figure citations and add a missing hyphen to “ass-kissing.”