AI Can Crack Jokes, But Does It Really Understand Humor? Here's What New Research Says
Artificial intelligence programs are being trained to mimic human writing and speaking patterns. However, while some programs are improving, there are still some uniquely-human emotions that can't be taught. One of these responses is humor, according to a team of researchers. During the 61st Annual Meeting of the Association for Computational Linguistics, researchers claimed that they found that in more than two-thirds of cases, AI models fail to grasp the nuances of humor after being fed numerous entries from the New Yorker Cartoon Caption Contest.
"We investigate[d] both multimodal and language-only models: the former are challenged with the cartoon images directly, while the latter are given multifaceted descriptions of the visual scene to simulate human-level visual understanding," wrote the team in their abstract. "We find that both types of models struggle at all three tasks."
In layman's terms, the researchers fed these cartoons to the models, then asked them to identify which caption goes with which cartoon and justify their answer. However, that doesn't mean that the technology isn't improving – at least, it's aware of what humor can be. The highest score that AI got when trying to connect captions to its respective cartoons is 62%, and while quite low, it is a remarkable start in teaching these models what humor can actually be.
What this could mean for AI
It appears as if, while likely a long way away from advancement, there is the possibility that AI models can eventually be trained to understand humor. Such testing has to involve trial and error to make it work, and the way that these models are trained will likely make it a tedious method. According to Neuroscience News, these models are usually trained using multiple-choice tests, much like how the Association for Computational Linguistics experiment tested the robots to match captions with cartoons alongside humans.
However, there is one key thing standing in the way of AI fully being able to grasp humor, and that is the loose definition of what understanding emotions actually entails. Some AI scientists argue that, while it is impressive that these models can crack jokes, it might never understand humor because that is such a uniquely-human idea. While this shouldn't take away from the accomplishments of the models thus far, the question might never be resolved in a universal way.
"If a model eventually surpasses whatever humans get at this test, you think, 'OK, does this mean it truly understands?'" research scientist Jack Hessel, Ph.D told Neuroscience News. "It's a defensible position to say that no machine can truly 'understand' because understanding is a human thing. But, whether the machine understands or not, it's still impressive how well they do on these tasks."