An evaluation showed that artificial intelligence (AI) isn’t good at telling jokes, and that the explanation is due to operating principle itself. It is claimed to be an interesting discovery that also can function a touch to those that use AI to support creative work.
MIT Technology Review reported on the seventeenth (local time) that Google DeepMind researchers conducted a test to generate jokes with AI with 20 American comedians and presented the research results on the ACM (ACM FAccT) conference on Fairness, Accountability, and Transparency. .
In keeping with this, Google researchers asked 20 comedians with experience using Open AI’s ‘ChatGPT’ or Google’s ‘Gemini’ to write down comedy material with AI. The research team was led by Google researcher Piotr Mirovsky, who also works as an improv comedian in his spare time.
The outcomes turned out to not be very impressive. AI generated jokes, but it surely didn’t produce material that was original, provocative, or very impressively funny.
Comedians, who requested anonymity for the sake of their reputations, said they weren’t very pleased with the outcomes.
They described the AI’s jokes as monotonous, generic, and boring. One participant compared it to a “Fifties comedy.” One other comedian said, “Regardless of how good you have a look at it, it’s a narrow and obvious approach.”
Nonetheless, some said that because AI prints anything quickly, it’s “useful for filling in blank pages.” There have been also responses that sometimes showed impressive compositional abilities.
In truth, a study showing that ChatGPT isn’t good at telling jokes and only outputs jokes which might be near old man jokes appeared a 12 months ago last June. Nonetheless, considering the tremendous performance improvement of enormous language models (LLM) thus far, it has been confirmed that there are limitations.
The researchers noted that while LLMs have gotten increasingly adept at mimicking human creativity, that does not imply they are literally creative.
Specifically, corporations reminiscent of Open AI and Google apply guardrails to stop violent or racist answers. This prevents the creation of elements reminiscent of aggressive, sexual content, or sarcasm which might be often utilized in comedy. Moreover, the info used for learning uses protected content that anyone can understand.
“If you happen to make something that appeals to everyone, it’s unlikely that it’ll find yourself being something that nobody likes specifically,” Mirowski said.
Nonetheless, it was identified that it isn’t just guardrails and training issues that prevent LLM from producing interesting responses.
Tuhin Chakrabarti, a pc science researcher at Columbia University who was not involved within the study, said much of the humor relies on surprise and incongruity, which runs counter to how AI models work. In other words, creative writing requires breaking away from the norm, but LLM can only imitate people.
“Comedies and good writing often have a structure that starts and ends. “But LLM has a tough time solving this since it is designed to predict one word at a time,” he said. “I attempted lots in my research to make AI fun, surprising, interesting, or creative, but it surely just didn’t work.” He said.
The evaluation is that comedians using AI will have the ability to work quickly, but won’t have the ability to create something original.
“People expect to make use of AI tools to write down scripts, screenplays, and commercials,” the researchers concluded. “But creative and comedic writing comes not from algorithms, but from the experiences and feelings of the individuals who read and judge them.”
Reporter Lim Da-jun ydj@aitimes.com
