Data and Technology

It's No Joke: AI Beats Humans at Making You Laugh

New research shows people don’t trust recommendations from algorithms—and that’s a problem for companies that increasingly rely on AI-based technology to persuade consumers. Michael H. Yeomans explains how businesses can overcome that bias.

We all enjoy sharing jokes with friends, hoping a witty one might elicit a smile—or maybe even a belly laugh. Here’s one for you:

A lawyer opened the door of his BMW, when, suddenly, a car came along and hit the door, ripping it off completely. When the police arrived at the scene, the lawyer was complaining bitterly about the damage to his precious BMW.

"Officer, look what they've done to my Beeeeemer!” he whined.

"You lawyers are so materialistic, you make me sick!” retorted the officer. "You're so worried about your stupid BMW that you didn't even notice your left arm was ripped off!”

“Oh, my god,” replied the lawyer, finally noticing the bloody left shoulder where his arm once was. “Where's my Rolex?!”

Do you think your friends would find that joke amusing—well, maybe those who aren’t lawyers?

A research team led by Harvard Business School post-doctoral fellow Michael H. Yeomans put this laughing matter to the test. In a new study, he used that joke and 32 others to determine whether people or artificial intelligence (AI) could do a better job of predicting which jokes other people consider funny.

The question is especially relevant today as more businesses turn to computer-based recommendation technology to help consumers make decisions. Yeomans' findings shed light on the hurdles that AI technology will need to overcome to win over wary consumers.

The team enlisted 75 pairs of people, including spouses and close friends. Among the participants, 71 percent had known each other for longer than five years.

First, the participants rated jokes on a scale from “extremely funny” to “not funny at all.” Then, after seeing their partners’ ratings for four of the jokes, they predicted their partners’ ratings for eight more jokes.

Meanwhile, a computer algorithm ran a series of tests to make its own estimations. The computer had no way of parsing the language in the jokes, nor did it follow a model indicating what features made a joke funny. Instead, it relied on “collaborative filtering” algorithms to learn which sample jokes were statistically similar to each test joke, based on participants’ previous preferences for certain jokes.

Who was the better judge of humor? The computer. Algorithms accurately picked the jokes that people deemed funniest 61 percent of the time, whereas humans were correct 57 percent of the time.

The computer even beat out the joke recommendations of close friends and spouses, a comedy of human errors that surprised the research team. They figured people would have a better handle on something as subjective and personal as the taste in humor of someone they knew well.

“Humans would seem to have many advantages over computers, but that didn’t matter,” says Yeomans, who co-authored the recent article Making Sense of Recommendations in the Journal of Behavioral Decision Making. “I was especially surprised that the recommender system outperformed people who had known each other for decades. I was really rooting for spouses to have an edge!”

Computers make good recommendations, but do people want to listen?

Businesses are investing heavily in sophisticated computer algorithms that rely on past consumer behavior to predict people’s preferences and recommend purchasing other relevant products, from movies and books to clothes and food.

Global spending on big data and business analytics is expected to increase 12 percent to $189 billion this year, and rise another 45 percent to $274 billion by 2022. Netflix, for example, believed so strongly in computer recommendations that the company offered a $1 million prize in 2009 to anyone who could build a system that improved prediction accuracy by just 10 percent.

“Companies now have this remarkable ability to learn about consumers and tailor their product recommendations in a personalized way,” says Yeomans, who co-authored the article with Jon Kleinberg of Cornell University and Anuj Shah and Sendhil Mullainathan, both of the University of Chicago. “The fact that the market has rushed so quickly to these tools; we felt it was important to bring them into the lab and see how they performed and what people thought of them.”

As Yeoman’s research shows, AI is often dead-on accurate in pinpointing which products and services people will like. Yet, the research findings also point to a perception problem companies should be aware of: People don’t like to take advice from machines.

“There’s a mistrust in algorithms. People seem to view them as a cheap substitute for human judgment,” Yeomans says.

His team probed this skepticism in a second study, where again algorithms outshined humans in determining which jokes would go over well and which ones would fall flat. But, in rating recommendations they were told came from a computer versus a human, participants gave human recommenders higher scores, showing that people would rather receive suggestions from a person, even if that advice is flawed.

After all, people are used to leaning on friends, family, and even strangers on the Internet when they’re deciding which products to purchase or even which people to date. And they put a lot of trust in their fellow humans; 83 percent of people say they trust recommendations from friends and family, and 66 percent even trust the online opinions of strangers, according to a Nielsen survey.

"A human recommendation can be valuable even when it's inaccurate," Yeomans says. “If my colleague likes a show I don’t like, I’m still happy to hear her recommendation because it tells me something about her. We bond over our likes and dislikes. It’s hard for computers to compete with that."

Where did that computer recommendation come from?

Besides, product recommendations that seem to pop up out of nowhere in a social media feed or email may come across as confusing and creepy to consumers. Another study by the team showed that participants rated human recommenders as easier to understand than machine recommenders.

“When participants thought the recommendations had come from a human, they were able to make sense of why someone might have chosen them,” the researchers write. “But when they thought the recommendations had been generated by a machine, those very same recommendations were perceived as inscrutable. … People are less willing to accept recommenders when they do not feel like they understand how they make recommendations.”

POLL## Do you consider these jokes funny?

We're asking Working Knowledge readers to rate some of the jokes that Michael H. Yeomans used in his study.

Take the poll

 

The researchers tested further to see if explaining the machine’s recommendation process would help people accept it more. The team told one group that they would simply feed their joke ratings into a computer algorithm that would recommend other jokes they might like, while another group received a more detailed explanation:

“Think of the algorithm as a tool that can poll thousands of people and ask them how much they like different jokes. This way, the algorithm can learn which jokes are the most popular overall, and which jokes appeal to people with a certain sense of humor. Using the database ratings, the algorithm will search for new jokes that are similar to the ones you liked, and dissimilar to the ones you did not like.”

Participants who received the detailed explanation rated the recommender system as easier to understand, and they preferred the algorithm more than the group that had less information. Learning about the process boosted their beliefs about the quality of the system’s performance and helped them to embrace it more.

“It is not enough for algorithms to be more accurate. They also need to be understood,” the authors write.

What companies can do

With that in mind, companies should consider ways to encourage consumers to appreciate AI-based recommendations from algorithms. One idea: Give the computer some “human-like characteristics,” Yeomans says. For instance, people may accept the output of an airline algorithm more if it pauses briefly to search for flights, giving people the sense that the computer is “thinking.”

“The delay helps people make sense of the process. The longer it takes, the better they think the algorithm is working because it must be searching all these different places,” Yeomans says.

Briefly explaining where the recommendations come from might also foster greater trust in them. Netflix and Amazon do this by telling users that because they chose a certain movie or product, they might be interested in similar items.

“Companies should show a little bit of the gears. Those little explanations help people wrap their heads around these recommendations,” Yeomans says. “The more businesses can do to explain how these systems work, the more likely people are to trust them and accept them.”

And for a company in today’s digital marketplace, that’s no joke.

About the Author

Dina Gerdeman is senior writer at Harvard Business School Working Knowledge.
[Image:
Cecilie_Arcurs]

Related Reading

Do you trust AI-based recommendations?

Share your insights below.

Latest from HBS faculty experts

Expertly curated insights, precisely tailored to address the challenges you are tackling today.

Strategy and Innovation

Social Responsibility

Diversity and Inclusion