Discovering a major flaw in artificial intelligence… it lacks a sense of humor

YNP:

A new study reveals that large artificial intelligence language models only claim to have a “sense of humor”, despite their appearance suggesting a human understanding of jokes, but in reality they lack creative thinking and the ability to analyze deeper meanings.

The results of the study were published at the “2025 Conference on Experimental Methods for Natural Language Processing,” where researchers conducted a careful test of artificial intelligence’s ability to recognize jokes based on verbal manipulation and ambiguity.

These models were previously thought to “understand” the jokes, but the researchers explained that the old data was not accurate, prompting them to update the tests and design new models to measure true understanding.

In one test, the models were asked to distinguish between ordinary sentences and sentences containing meaningful verbal puns, but when keywords were changed, response accuracy dropped sharply, as the models became confused and incorrectly classified the jokes, revealing their weakness in understanding context and phonetic similarity.

“When models are faced with unusual verbal manipulations, their success rate can drop to as little as 20%, which is below the level of random guessing,” said co-author Muhammad Tahir Belwar of Cardiff University.

“The models are overconfident in assuming that the text is funny even if it is not,” he added.

The study confirms that these results highlight the limitations of current artificial intelligence in the field of humor, which requires the development of future more precise models in the analysis of precise meanings and creative language.