Sexism and Shopping: Female Players Get Most of the Odd Questions at the U.S. Open

First, here’s what they reported finding last year: The question about Ms. Williams’s smiling was far from an exception. Across all the categories analyzed and all the questions, the algorithm revealed that female players were much more likely to be asked questions unrelated to tennis. Once we remove the more standard, rote (or “typical”) questions, roughly 70 percent of the questions unrelated to tennis were posed to female players.

(This point is driven home in a popular parody video of male athletes being asked nonsports questions.)

We asked the researchers to apply their algorithm to this year’s Open. Here are some typical questions or rote questions asked of male tennis players:

“What would you expect from a match like that?”

“How does it feel to be back in the second week of the U.S. Open?”

Here are some typical or rote questions asked of female players:

“What do you think of the match that’s coming up?”

“You have been through injuries and you have had some tough matches here, but what makes it special to play here at the U.S. Open?”

The researchers’ algorithm also identifies the most atypical, even bizarre questions asked of the athletes.

Asked of men:

“There were some moments you were doubting yourself or not?”

“What does it mean to you if you are indeed an inspiration for people who are not tall?”

Asked of women:

“Do you know of players who get their nails done on-site?”

“Was there anything in particular you bought when you went shopping?”

How did an algorithm “know” to single out these questions? How did it decide what was and what was not related to tennis?

To understand this, it helps to know why algorithms are useful for processing language. The first task in this example is to find individual topics or words that we conjecture are unrelated to tennis. Imagine doing that without a computer. Exactly because language is rich, questions can differ in a dizzying array of ways. Searching with our own eyes through all these possibilities could take impossibly long. An algorithm, on the other hand, could try thousands of such paths.

This by itself is a remarkable skill. But it still needs guidance on what paths to try. We know broadly what we would like it to do — look for any differences in the tennis-relatedness of questions. Yet how would the algorithm know which combinations of words are related to tennis?

Answering this question is what makes this particular paper brilliant. The Cornell researchers understood that they could train the algorithm on another set of data. We not only have the postgame language, but also the in-game commentator language. This language provides a rough guide for what words and language involve tennis.

An algorithm trained on this data can then be applied to the postgame questions. It can infer “tennis-relatedness” by looking at how a question’s words and linguistic structures differ from those used during in-game commentary.

Doing this produces a clear pattern of gender bias. Even though the algorithm is blinded to players’ gender, its output can be compared across genders. To remove the fairly straightforward questions, they additionally categorized the typicality of questions so as to focus on the atypical ones. The researchers’ job is now easy: to simply compare tennis-relatedness of questions asked of male vs. female players.

A clear pattern of gender bias emerges: The questions unrelated to tennis were lobbed disproportionately at women.

But the algorithm did not discover these biases on its own. This paper is cutting-edge research exactly because it required a spark of human intelligence. This was not a rote activity. It required months of work by some of the best researchers in the field of natural language processing.

We risk misleading ourselves when we attribute breakthroughs such as this to machine intelligence. Maybe in the future, machines will have insights of their own, and we would be wise to recognize that as a possibility. After all, some visual and audio processing tasks are already fully automated.

In other areas, we recognize that technologies — however revolutionary — increase the premium on creativity and intelligence. Breakthroughs have value only when cleverly deployed by humans. Artificial intelligence is no different. It raises, not lowers, the value of human intelligence and creativity. That fact can be obscured when we anthropomorphize and attribute sentience to a tool.

In the 15th century, the introduction of linear perspective revolutionized art. A masterpiece like “The School of Athens” simply could not exist without it. When we anthropomorphize algorithms and imagine that breakthroughs in artificial intelligence come directly from algorithms, it is analogous to attributing this masterpiece of art to “Linear Perspective” rather than to Raphael.

Continue reading the main story