June 22, 2024 – According to a report by TechXpore yesterday, a research team from Google’s DeepMind project has recently discovered that large language models are “not very good” at writing funny jokes.
In a novel experiment, the team set out to investigate whether large language models could create humorous jokes or even entire stand-up comedy routines. To this end, they enlisted the help of 20 professional stand-up comedians who had previously worked with large models. These comedians were tasked with using the large language model to assist them in crafting complete stand-up routines and then rating the results.
The researchers found that while the large language model was “fairly adept” at generating jokes, only a few of them were actually funny. Most of the jokes generated were considered “generic” and lacking in surprises.
Many comedians felt that the AI-generated jokes were missing the “edge” needed to make them humorous, with some even describing the results as “bland”. However, a segment of comedians believed that the large language model was useful in generating basic “frameworks” for jokes, which they could then build upon.
The research team speculated that the reason AI jokes might not be as funny could be due to filters used by the creators of large language models to prevent potentially offensive or insulting output.