Wired this week features a story by Will Knight with the title, “Why a YouTube Chat About Chess Got Flagged for Hate Speech.” It highlights the curious relationship between the natural behavior of human beings and the sophisticated but fundamentally unnatural logic of artificial intelligence (AI). Until very recently, our civilization thought of data simply as the formally defined information humans gleaned from their collective experience of the world. In recent years, the innocent notion of data morphed into a compelling and powerful phenomenon called Big Data. This new organic personality foisted on what was formerly thought of as a collection of random facts became a kind of Frankenstein’s monster.
By combining into a vast neural network the digital “bits” that composed its ever-growing collection of resources, the masters of a new order saw Big Data as the set of assets needed to craft a sophisticated hyperreality capable of impressing, organizing and intimidating people and their institutions. Combined with AI, Big Data had the power to simplify complexity. It would provide the means of addressing all the annoying problems indecisive and insufficiently informed humans have failed to resolve.
Over tens of thousands of years, our ancestors acquired the faculty of language that allows us to express ourselves in an unlimited way on an unlimited number of topics for an unlimited number of purposes. Humans had no rivals for linguistic creativity on planet earth. But recently, things began to change. Human language and our own tireless ingenuity have spawned our first rival: AI. Like most modern inventions, AI’s development was justified by its praiseworthy dual purpose of reducing the human effort required to execute profitable tasks and saving that most precious of commodities — time.
But unlike the wheel, the clock and even calculators, computers and wireless transmission, AI is not focused on performing a single concrete task or set of procedures. Its ambitious human designers, operators and programmers have endowed it with the nobler mission of producing a new finite order out of the fragments and shards cast off by the infinite magma of knowledge, intentions and moods that humans have always generated through their use of language.
One day soon, according to the experts, AI will have the capacity to program itself in its mission of reducing the infinite to a finite set designed to meet the needs of the human community. We will then be free to sit back and consume the fruit of its intelligence. Machine learning represents the ultimate teleological endpoint posited by our civilization’s culture of progress. For such people, order is the successful attempt to tame the infinite and make it finite.
The first sentence of the Wired article cites an example of how this reductionist logic works. It recounts how, in June 2020, “Antonio Radić, the host of a YouTube chess channel with more than a million subscribers, was live-streaming an interview with the grandmaster Hikaru Nakamura when the broadcast suddenly cut out.” This wasn’t the result of a random glitch, but rather the effective intervention of YouTube’s AI. “Instead of a lively discussion about chess openings, famous games, and iconic players, viewers were told Radić’s video had been removed for ‘harmful and dangerous’ content. Radić saw a message stating that the video, which included nothing more scandalous than a discussion of the King’s Indian Defense, had violated YouTube’s community guidelines.”
Today’s Daily Devil’s Dictionary definition:
This incident reveals the risk associated with two unrelated modern trends. The first is the ever-increasing confidence in algorithms, which we are encouraged to consider as the ultimate model of intelligence. The second is what might justifiably be called the triumph of trigger-warning culture. We are expected to think of algorithms as something more powerful than our ordinary human intelligence. Its rigor and complexity exceed the capacity of human understanding. We must be humble and accept all the consequences.
As this incident demonstrates, trigger-warning culture, based on the omnipresent fear of offending — and being shamed or sued for it — is now a prominent feature of the algorithmic systems that monitor our behavior and language. AI is becoming our ultimate moral censor.
Facebook Wants to Read Your Brain
Will Knight warns us of the futility of attempting to analyze the cause of YouTube’s cancellation of the program: “Exactly what happened still isn’t clear. YouTube declined to comment beyond saying that removing Radić’s video was a mistake.” Calling something a mistake generally means that the person doing the calling refuses responsibility for the occurrence. In this case, YouTube was right. Its technology has a mind of its own — much like Saudi Crown Prince Mohammed bin Salman, who claimed that the gruesome murder and dismemberment of the journalist Jamal Khashoggi was a mistake due to the unexpected enthusiasm of the death squad he sent to Istanbul.
When an algorithm makes a mistake, no living person is to blame. Moreover, there is little to be concerned about since algorithms can simply be refined and improved. It’s the best of all possible worlds. Nevertheless, the author suggests that all is not well as “a new study suggests it reflects shortcomings in artificial intelligence programs designed to automatically detect hate speech, abuse, and misinformation online.”
The article repeats a point that linguists have been making consistently for at least a century and creative writers for several millennia: “The same words can have vastly different meaning in different contexts, so an algorithm must infer meaning from a string of words.” Technologists now seem to find it surprising.
The question then becomes, where does the string start and end? It implies a notion directly related to what in physics is called string theory, which posits far more than the standard three or four dimensions humans are familiar with. How many dimensions exist for any utterance? Can they even be numbered?
For the physical world, string theory suggests that we live in a universe with as many as 11 dimensions. Some push the number to 26. For his linguistic theory, Noam Chomsky developed an idea borrowed from Wilhelm von Humboldt that language is “the infinite use of finite means.” This is due to the property of embeddedness. Language permits the expression of ideas that may have other ideas embedded within them. Understanding, in contrast to algorithms, is non-linear. It results in infinite possibility.
The technologists promoting AI avoid the attempt to understand the difference between algorithmic logic and linguistic creativity. Language derives its force from the full breadth of human experience. Unrestricted to material logic, language can mobilize formal logic for infinitely varied purposes, legitimate and illegitimate. It can even make mistakes and commit crimes. But when people assign a strategic purpose to the algorithms they design or exploit, there will inevitably be a gap between what they understand as strategy and what the algorithm is capable of achieving. The resulting mistakes and crimes will defy human understanding.
Algorithmic logic cannot strategize its capacity for expression. It can simulate the effect of human strategies, but it cannot make sense of them and even less actively express them. Some claim that the singularity — the moment in the near future when AI surpasses human intelligence — is inevitable. But fundamental reasons exist — including the disembodied algorithm’s lack of a corporal identity — that will confine even the most sophisticated future versions of AI to a domain that is utterly incommensurate with human or indeed any biological reality.
The “mistake” YouTube’s algorithm made had no serious consequences, yet it was unforgivable. One of the reasons can be found in its premise. It was designed to detect hate speech, but it relies on means that remain superficial (the appearance of words) and quantitative (statistical). The moral focus of the tool should be on hate rather than speech. But hate is not directly detectable through language. The procedure focuses on speech, which is infinitely ambiguous.
The ultimate ambiguity and the source of AI promoters’ hypocrisy comes from seeking to respect “community guidelines.” In the infinite variations of reasoning deployed by any finite group of people called a “community,” who defines a guideline? Guidelines, like beauty, tend to be in the eyes of the beholder.