Forget the Turing Test—give AI the F. Scott Fitzgerald Test instead

From the 1920s to the 2020s.
From the 1920s to the 2020s.
Image: AP Photo/File
We may earn a commission from links on this page.

Scott Fitzgerald pinned human intelligence on its tolerance of paradox. But what kind of artificial intelligence could pass his test?

In his 1936 essay “The Crack-Up,” Fitzgerald writes that “the test of a first-rate intelligence is the ability to hold two opposed ideas in the mind at the same time, and still retain the ability to function.” For example, he says you should “be able to see that things are hopeless and yet be determined to make them otherwise.”

He confesses he’s lost this ability—and as a result, himself.

Fitzgerald’s point is not that he needs a better model of the world, but that he needs many models and the freedom to switch among them. This is what allows us to forge ahead despite unexpected obstacles, conflicting priorities, or, in Fitzgerald’s case, hitting his forties and feeling like someone changed the rules of the game while he wasn’t looking.

Fitzgerald, having lost his ability to balance opposing ideas, falls into a drab, routinized existence. Every moment, from his morning routine to dinner with friends, becomes a forced act. He mimics the life of a successful literary man without actually living it.

Too often, this is where AIs get stuck today. They offer a cheap imitation of a life well-lived, but only from a narrow perspective. They are unable to hold two opposing thoughts at once, instead needing to exist in a state where decisions are made with one model or another, but not debate between the two. If they faced the Fitzgerald test, they wouldn’t pass it.

Take a simple example. When a navigation app redirects stop-and-go traffic from the New Jersey Turnpike onto local roads in the town of Leonia, otherwise quiet neighborhoods become overrun with shortcut-seeking app-watchers. A compassionate human might weigh up the cost-benefit analysis of a shorter trip with the potential annoyance of hapless suburbanites. But a naïve AI, focused only on finding the fastest travel time, won’t. Local authorities are now plotting to fine non-residents caught driving through the area during rush-hour, even though they’re just following the directions on their smartphones.

This is the risk we take with single-minded AIs. We willingly take advantage of optimization, but we unwittingly lower ourselves below Fitzgerald’s test of first-rate intelligence. The fastest route home is not always the most enjoyable, or the most neighborly.

It doesn’t have to be this way. A navigation AI that could pass Fitzgerald’s test would handle this situation differently. It would use opposing algorithms that represent different perspectives on viable routes. One could pit its calculation of the gain from shorter travel time against another that estimates the costs imposed on the area you drive through. These could include increased wait times at intersections, CO2 emissions in densely populated neighborhoods, or even wear and tear on local roads.

In a way, being held to this elevated benchmark is testament to an AI’s success. The reason this issue pops up with navigation is because so many people use it. We should expect to see this same progression as AI demonstrates its effectiveness in more ethically complex areas.

For example, the Office of Children, Youth, and Families in Pennsylvania’s Allegheny county ran a year-long pilot to see if an algorithm could help to protect the county’s children. People who screen incoming calls decide if the situation warrants an in-person visit in the next 24 hours or not. These screeners now get a second opinion from an AI.

The algorithm assesses the likelihood of a bad outcome in the home—physical harm to a child or even death—by scoring the details of the call against more than four years of past cases. Its track record is strong, having bumped up the high-risk cases that get investigated, as well as the low-risk cases that are spared an unnecessary visit. It has made the treatment of black and white families more consistent according to their risk profiles.

As this algorithm moves out of pilot phase into broader usage, calls for it to mature further will likely grow. Fortunately, there’s a known path to walk in becoming a Fitzgerald AI.

Credit scoring is one of the most common algorithmic decisions with real-world ethical consequences. Banks recognize that these algorithms must continually evolve, not only because the risk factors for default keep changing, but also because the use of their own algorithms changes the real world itself, expanding the availability of credit. To deal with this, banks regularly create new credit-risk models to second-guess the existing ones. This constant competition keeps the algorithms’ predictive power high while also reducing the risk of creeping institutional bias.

These modern and highly charged decisions bring a hidden aspect of Fitzgerald’s test to the fore: transparency. The success of Fitzgerald’s first essay led him to write two more. Over these three articles, he labors to explain his inner workings: to examine how and why his thinking has cracked, and what to do next

A Fitzgerald AI should do the same. Algorithms with opposing viewpoints should make their basic logic, inputs, and weightings clear so that we can follow their argument for whatever decision they make and benefit from it. This need not be a radical transparency that forces private companies to lose competitive advantage by revealing proprietary methods, but it should provide an explanation that an average person could read and understand.

The stakes are high. In his last essay, Fitzgerald follows his descent into hopelessness to its conclusion. Although he can’t give up being a writer because he has to make a living somehow, he gives up being a person. He consigns himself to one goal: waste—of his time, of his resources, and of his talent. He swears off any attempt at making things better, even for his fellow aspiring writers.

Ironically, Fitzgerald’s public crack-up succeeded where he failed. His essays were a hit, expanding the idea of how to be a literary hero. Sometimes you feel like a Hemingway, blasting through the green hills of Africa, and sometimes you feel like a Fitzgerald, overwhelmed by the pointlessness of it all.

The freedom to be either and switch between them as circumstances demand is the key to acing Fitzgerald’s test. AIs that pass the test will help us live not just more efficiently, but better, too.