Why Computers Can't Make It as Stand-Up Comics

Computers may have beaten us in chess and "Jeopardy!" — but can they create and tell a good joke?

Fortunately for comedians like Jon Stewart, who verbally sparred with a right-wing doppleganger of "Jeopardy!"-winning computer Watson on his show this week, any computer cracking wise is likely to need human writers for the foreseeable future.

A new review of the literature on the development of the human mind indirectly suggests how far computers have to go before they can present themselves as our new comedic overlords.

The article, coming out in the journal Science this week, focuses on how the human mind, despite the messy and inconsistent information it receives from the world, gets to be so high-functioning.

"Most machine learning is about learning from very massive datasets. Human intelligence is also about coming to a pattern about how things work," said Josh Tenenbaum, an MIT professor who is one of the paper’s four co-authors.

And unexpected patterns of words and concepts are the essence of comedy.

The human mind, Tenenbaum told LiveScience, seeks to structure objects in a logical way to help understand them. One example is the political spectrum in the United States, which is commonly represented as a simple left-to-right line and, Tenenbaum said, may need to be expanded to a second dimension for a fuller understanding of ideas.

"Our language shows there is an underlying one-dimensional space to how we think about politics," said Tenenbaum, a professor of computational cognitive science.

This ability to structure information, it seems, is present to some degree at birth.

"As a matter of empirical fact, we know newborns who see objects for the first time ... already are able to represent objects the first time they encounter them," said Elizabeth Spelke, a professor of cognitive psychology at Harvard, who was not involved in the review article.

"I see this paper as laying out a blueprint for a future program of research," Spelke added. The next step, she said, is to see whether the models can predict how we know what we know. [Brain X Prize May Spur Big Solutions]

Humanlike machines

Part of the success of the human mind is that it is able to assimilate facts into a structure while simultaneously evaluating their "truth" based on previous knowledge. To figure out how it does that, the researchers say, we must answer two questions: how the mind interprets numbers, and how it interprets symbols and facts.

"You have to have both of these things, and you have to understand how they go together," said Tenenbaum. "We have to understand how they work together to understand how the mind works."

That understanding could enable development of a computer with similar capabilities. "The road is ridden with obstacles but the goal is clear: to make a machine act intelligently in various areas," said Judea Pearl, a professor of computer science at UCLA, who was not involved in the study.

Computers, Pearl said, are able to understand statistics and actions. But they cannot handle the next level — understanding alternative possibilities.

For instance, he said, the sense of regret is based on the idea that our minds can evaluate what would have happened had we done something differently — a thought level that computers have not yet reached.

Tenenbaum used Google as an example. The search engine rapidly looks for word patterns rather than actually understanding what the user is asking. Tenenbaum described its inner workings as "fast and stupid."

Mind matters

While humans are able to consider alternatives, one challenge has always been in explaining why some humans are bad at understanding cause and effect.

For instance, why are some medical treatments so popular when they have no scientific basis? Homeopathy — the use of extremely diluted symptom-causing substances to treat patients with certain ailments — is used by 4.8 million Americans annually despite the fact that "a number of its key concepts are not consistent with established laws of science," and "most analyses of the research on homeopathy have concluded that there is little evidence to support homeopathy as an effective treatment for any specific condition, and that many of the studies have been flawed," according to the National Center for Complementary and Alternative Medicine.

The Science review touches on those ideas, explaining how human knowledge is constructed on what is called a Bayesian system, meaning the mind gives new ideas a probability of being true before investigating them. This can account for why the mind can assemble rational thoughts — and why once an irrational thought is accepted, it can be hard to change, as new facts that contradict it are given a low probability of truthfulness.

One future direction in the human aspects of mind research is explaining how to fix our own bugs.

"I think it's really interesting why we seem smart and rational in one domain, and shift ... and people can seem really irrational," said Spelke. "If one is going to come up with an adequate account of the human mind, one would have to answer this."

In designing the computer mind, it may be important to understand how to avoid those irrational flaws in the human one.

But even when computers pass that stage, they still won't understand the conventions of human language necessary to subvert them and construct a joke. And those conventions are critical for any creative work.

So take heart: Watson may beat Ken Jennings, Brad Rutter or you at "Jeopardy!" but, as Tenenbaum noted, you would beat him easily at writing the questions.