Stand-up comedy: LLMs don't write good jokes
Can an artificial intelligence write a funny comedy program? Researchers from Google Deepmind let comedians try it out.
(Image: Tatiana Shepeleva/Shutterstock.com)
A research team from Google Deepmind has discovered in a scientific study that large AI models can not write good - i.e. funny - jokes for stand-up comedians. Although artificial intelligence (AI) can come up with new jokes, they are usually not funny.
Stand-up comedy is not an easy undertaking. Stand-up comedians work on their programs for a long time to write a humorous, coherent program. They usually test them in parts in front of a smaller audience before taking to the big stage to test their reactions. If a joke works, it is adopted; if not, it is dropped from the program.
Scientists at Deepmind wanted to know whether it is possible for an AI to write a funny stand-up program. To this end, the scientists hired 20 professional stand-up comedians who had already gained experience with Large Language Models (LLM) in the development of stand-up comedy programs. They were asked to develop a complete stand-up comedy routine with an AI and evaluate the result. The Deepmind researchers published the research results in the scientific study "A Robot Walks into a Bar: Can Language Models Serve as Creativity Support Tools for Comedy? An Evaluation of LLMs' Humour Alignment with Comedians", which has been published as a preprint on Arxiv.
No good jokes
According to the study, the LLMs were pretty good at coming up with new jokes. However, most of them were quite generic and rarely offered a surprise, a basic requirement for a good joke. Only a few of the comedy acts were really funny.
The comedians mainly criticized the fact that the AI-generated jokes lacked the sharpness and pointedness to be perceived as funny. Most comedians therefore rated the jokes as boring. At the same time, however, some comedians were convinced that LMMs could be useful in the development phase of a comedy program. They believe that they can be used to develop a basic structure.
However, these results are not entirely surprising, according to the researchers at Google Deepmind because the developers of LLMs build filters into their AIs to prevent offensive or "edgy" statements from being generated.
(olb)