Computer Spins Yarns but Lacks a Certain Byte

Brutus creates stories about lies, self-deception and acts of betrayal.

There is no dark muse inspiring Brutus, though. No torturous exploration of the writer’s soul.

Brutus is, after all, a computer blueprint.

The program based on the blueprint, Brutus.1, can write stories because its creators have condensed the complexities of deceit and double-crosses into mathematical equations. Characters and facts of the story are fed into Brutus.1, and out come 500-word tales that read very much like human prose.

“Dave Striver loved the university,” Brutus.1 begins one tale. “He loved the ivy-covered clock towers, its ancient and sturdy brick and its sun-splashed verdant greens and eager youth.”

The prose isn’t exactly the caliber of E. Annie Proulx, or even Danielle Steel. But in an age when computers can already read, talk and beat chess champions, Brutus.1’s small literary success raises a big question: Can computers write good stories?

“I get e-mails from terrified people,” said Brutus’ co-creator, Selmer Bringsjord, director of the Minds and Machines Lab at Rensselaer Polytechnic Institute.

Researchers like Bringsjord (pronounced BRINGS-yord) have been trying to get machines to write like humans for decades, with mixed success. In the 1970s, a pioneering storytelling machine spun simplistic tales like: “Once upon a time John Bear lived in a cave. John knew that John was in his cave. . . . “

Brutus.1, in comparison, can write much more sophisticated prose. But after seven years of development, Brutus.1 remains a limited writer. It only produces stories in the 500-word range. The only setting it describes is academia.

And Brutus.1--named for the famous conspirator against Caesar--only can write about betrayal and related acts. That’s partly due to the creators’ preferences. Bringsjord notes that the dark side can be more fascinating. But there’s another important reason: Brutus.1 doesn’t understand love.

“If you go to the good side of things you have to figure out love, and we have no idea how to do that,” Bringsjord said.

Betrayal, on the other hand, was something Bringsjord and co-creator David Ferrucci of IBM’s T.J. Watson Research Center felt they could reduce to mathematics.

In Brutus.1’s world, the act of hoodwinking is expressed in part like this: “X betrays Y if and only if there is something Z that X wants.” Characters progress through a series of these formulas like pinballs, creating the narrative. Little tweaks of the formula change where the story goes.

The story comes out looking like proper English because Brutus.1 has been “reverse-engineered”--fed information about what a story looks like. For instance, grammatical rules on how to construct a sentence were given to Brutus.1. It knows rules of thumb for fiction, such as where to place quotes. And Brutus.1 will dab in touches of color to give a sense of setting--clock towers, eager students, etc.

The resulting prose can look a bit purple: “A wave of hatred rose up and flowed like molten blood through every cell in his body.” And it can sound stilted: “To become a doctor, one must pass an oral examination on one’s dissertation.”

But on the whole it reads like human prose. During a recent online fiction contest, just a quarter of the people who gave a guess pegged the computer’s story among the five competing entries. However, Brutus.1’s entry also was the least favorite among the readers.

Why the discouraging reviews?

It might have to do with what Bringsjord calls “lexical choice”--that is, choosing the right words that make sentences flow. It’s why “To be or not to be?” sounds so much better than “Should I kill myself?” The problem is programming a computer to make those choices.

Kathleen McCoy, a University of Delaware computer scientist involved in creating systems that create natural-sounding sentences, said a key problem is that people can recognize good writing but have a hard time defining why it is good.

“It may be perfectly grammatical, and even the thoughts it may be getting across may be reasonable,” she said. “But it doesn’t look like real text.”

Still, natural language researchers have been making strides in creating systems to generate picture captions, summaries of multiple newspaper articles, postoperative medical reports and electronic museum guides. An Ithaca, N.Y.-based company called CoGenTex has developed a system that relies on meteorological data to create weather forecasts.

But there is a huge leap from weather reports to “War and Peace.” Kathy McKeown, a Columbia University professor who is a leader in the field, notes that word choices are much more complicated in fiction.

The feeling among many experts is that the Great Computer Novel won’t be possible for many years, if ever.

Even assuming computers can shed stilted language, there is the larger question over the limits of artificial intelligence. Some scientists believe computers one day will be able to replicate human thoughts. Under that view, it is plausible to think of computers spinning yarns as memorable as one of a sea captain obsessively hunting a whale or a poor soul metamorphosed into a giant insect.

Others like Bringsjord see it differently. He notes computers will never be in touch with the “inner lives” of their characters. Computers will never draw on what it is like to relish sunny skies or feel a tender touch.

Bringsjord believes that Brutus.1 or its successors might knock off some sort of formulaic fiction, such as an action movie script or a narrative for an interactive video game. But after seven years of work, Bringsjord concludes that first-rate storytelling will always be the sole province of human masters.

“I feel like I know what I say when I say humans create things totally and utterly amazing,” he said. “I know how hard it is because I can’t build it.”