Can artificial intelligence understand humor?

Artificial intelligence (AI) can do a lot. It can speak. It can pick out objects in a photo. It can help people find their way around a subway. But can it make us laugh? Not really, at least not yet. And the reason why, can tell us a lot about what sets us apart from machines. 

Most AI today is based on machine learning, where software learns to spot patterns in large data sets. 

Because software is fast and never tires, machine learning gives AI an edge over humans when it comes to many simple tasks, such as finding a given image within a vast photo archive. Things are harder when you are dealing with feelings, though. And that is what is at the heart of humor.

See also: The future of work will be AI, how will you prepare?

"People have had some success in defining what would constitute humor," says Abhijit Thatte, Assistant Vice President of Technology and Practice Leader for Artificial Intelligence at Aricent, a global design and engineering firm. "But it's not been completely codified."

As even full-time stand-up comics would admit, there is no magic formula to produce the perfect joke. Much of what makes us laugh depends on subtle factors such as context or body language. "Sometimes even we humans don't know why a joke is funny," says Thatte. 

"What is funny for one person may not be funny for others," he notes. "The same joke told by two people might elicit different responses. A joke is fun, and a pun can be subjective. So, how can we teach AI to create jokes if we ourselves don't grasp the reasons why a joke is funny?"

AI, which tends to focus on a very narrow range of tasks, is poorly equipped to spot the wide range of factors involved in humor, let alone know what they mean. "Lack of context is one area that makes it more difficult for algorithms," Thatte says. 

No wonder, then, that attempts at machine humor have had mixed results so far, according to researchers from the University of Edinburgh. One joke-telling systemwas able to come up with so-so one-liners, such as, "I like my coffee like I like my war: cold." 

Elsewhere, research teams have long tried to produce robots that could hold their own as comedians. But the humor comes from humans, not machines. 

As one comic robot, from the Cognitive Science Research Group at Queen Mary University of London, United Kingdom, says: "You know what really pushes my buttons? That guy that's in control of me."

But beyond getting a laugh, there are good reasons why humans might want machines to know about humor. More and more, robots are doing jobs that require getting along with people. 

And knowing if a person is being funny or not could be a big help, because only very, very smart machines would get the joke. "Humor throws your expectations upside down," says Will Williams, machine learning engineer at Speechmatics in Cambridge, United Kingdom.

"What you need for that is for AI to have a really good model of the world, which it doesn't have at the moment."

In other words, the kind of sense that would allow a robot to laugh at jokes would also allow the robot to know much more about the human world. This, in turn, would help robots cope with a much wider range of tasks and situations.  

See also: Cisco has the edge with artificial intelligence and machine learning.

Maybe this knowledge would include knowing about feelings, too, although for now a more likely picture is that robots might be programmed with functions that would reward human-like actions, in the same way that a parrot might for talking.

Even so, making any machine that could have a decent stab at being funny would be a big step forward. "It's much better than the Turing test," says Williams. "With the Turing test, you've ended up having loads of teams coming up with canned rules to trick you."

With something like humor, he says, "you've really got to have a deep understanding of the world, how things work, how people work. It's indicative of something that really is intelligent."