AI Loves— _and Loathes— _Language

 Deep learning networks may look like brains, but that doesn't mean they can think like humans. On the ever-expanding meganet, this is a problem.

A FEW YEARS ago I found myself researching the vexed issue of Shakespearean authorship. I wanted to know if the anonymous Renaissance play Arden of Faversham (1590) was written in part or in whole by William Shakespeare. Perhaps, as some research has argued, an AI could look at a field of plays divided into just two categories—Shakespeare on one side of the fence and everyone else on the other—and definitely place Arden of Faversham on the right side.

The AI ​​considered what words Shakespeare and only Shakespeare usually used, as well as words that Shakespeare and only Shakespeare avoided. The researchers placed Shakespeare's plays on one side of the fence and all other Renaissance plays on the other. Then we ran the AI ​​and tasked it with figuring out what kinds of features are common to Shakespeare's plays, and more importantly, what kinds of features are common only to Shakespeare's plays. So when Arden was thrown at the AI, she decided to place Arden on the Shakespearean or non-Shakespearean side of the fence based on what "Shakespearean" words she had.

The result, as it turns out, is inconclusive. The field happens to be a lot less neat than I portrayed. The AI ​​doesn't see the fence I mentioned that divides the categories. They make that fence instead. Here comes the problem. If, after drawing the fence, the plays on either side separate cleanly, then we have a neat line between the two categories of Shakespearean and non-Shakespearean plays. But if this separation is not so clean, then it is much more difficult to be sure of our classification.

As you might expect, Renaissance plays don't cluster so neatly into Shakespearean and non-Shakespearean plays. Shakespeare's style and verbosity are so varied and dynamic that he encroaches on the space of other authors—as other authors often do on each other. And word frequencies alone are probably not enough to definitively prove authorship. We must consider other properties such as word order and grammar in the hope of finding a field on which to neatly draw a fence. We still have to find it. The same goes for the boundaries between offensive and non-offensive language, which Perspective AI – a Google project launched in 2017 to filter out offensive language from internet conversations and comments – had such difficulty identifying, or even the chatbot's inability to determine appropriate versus inappropriate responses .

The AI's failure to classify Arden of Faversham can be attributed to several different causes. Maybe there just aren't enough games to properly train the AI. Or maybe there is something about the nature of Renaissance game data that causes the AI ​​to have a harder time with certain types of classification problems. I'd say it's the nature of the data itself. The particular kind of data that makes AI impossible more than anything else is human language. Unfortunately, human language is also the primary form of data on the meganet. As deep learning applications confuse language, artificial intelligence — and meganets — will learn to shun it in favor of numbers and images, a move that threatens how people use language with each other.

Meganets are what I call persistent, evolving and opaque data networks that control (or at least heavily influence) how we see the world. They are bigger than any platform or algorithm; meganets are more of a way to describe how all these systems are intertwined. They collect data on all our daily activities, vital statistics and our very insides. They construct social groupings that could not even exist 20 years ago. And as the new brains of the world, they are constantly changing in response to user behavior, resulting in collectively created algorithms that none of us intended—not even the corporations and governments that run them. Artificial intelligence is the part of the meganet that most resembles a brain. But by themselves, deep learning networks are brains without vision processing, speech centers, or the ability to grow or act.

As my experiment with Shakespeare's plays shows, language provides the best counter-argument to machine learning's claim that "thinking" problems can be solved by classification alone. Deep learning has been able to achieve some remarkable approximations of human performance by stacking layers and layers of classifiers on top of each other, but at what point could a mathematically based classifier sufficiently approximate knowing when, for example, to use the familiar pronoun tu in French versus the polite pronoun vous? Vous can be the formal form of "you" and the informal, but there is no fixed definition of formality. There is no hard and fast rule for use, but an ever-changing, culture-driven set of guidelines that even people don't fully agree on. Sorting through the inconsistent and conflicting examples of use of each, one begins to question whether deep learning pattern recognition could ever be sufficient to mimic human performance. The distinction between tu and vous is actually a sharper and more subtle form of the distinction between offensive and non-derogatory language that Perspektiva had so much trouble with. The amount of ambiguity and context in human language escapes the type of analysis performed by deep learning.

Perhaps one day the opaque deep learning brains will be able to approximate human linguistic understanding to the point where they can be said to truly understand vous and countless other such differences. After all, we cannot open our own brains and see how we ourselves make such distinctions. Yet we are able to explain why we choose to use tu or vous in a particular case to explain the interactions of our own embodied brains. Deep learning can't, and that's just one indication of how far it has to go.