on verbal humor

Puns, Punchlines, and Surprise

Sound similarity, surprise and amusement have all been measured in research on verbal humor. Some experiments asked people to rate sentences; others asked models to match cartoons and captions or explain a joke.

Words that sound alike

An August 2025 study in Humor examined punning as an ambiguity between two similar-sounding words, the pun and its target, in a context in which both meanings were more or less acceptable. Researchers measured the sound distance between these words in a large dataset of puns and compared it with ratings of the jokes' funniness. Two of the four sound-distance measures showed a significant association: the closer the sound, the higher the funniness rating. The authors proposed examining the semantic features of the paired words in future work.

Reading time supplied a different measurement in a January 2023 study in Frontiers in Psychology. Participants rated the comprehensibility of sentences while their eye movements were tracked. The materials were homophone puns and two kinds of unfunny control sentence, congruent and incongruent. At the critical homophone region, fixation measures favored congruent controls over puns. Across the whole sentence, the reading-time pattern reversed. Humor ratings collected afterward were negatively correlated with overall sentence reading time: higher amusement ratings were associated with shorter reading times.

Surprise and sense

A May 2025 paper in Cognitive Science investigated whether exposure to language alone could account for recognizing one-line jokes and identifying their entailments. Its preregistered experiments tested GPT-3, described in the paper as a large language model trained only on language data. The model performed above chance in tasks testing joke detection, appreciation and comprehension. Exploratory work also found above-chance detection and comprehension in open-source models including Llama-3 and Mixtral.

All the language models tested fell short of human performance. Both humans and models tended to mistake nonjokes with surprising endings for jokes.

Finding it funny

A 2024 study in the International Journal of Psychology compared positive, neutral and negative emotional-state groups, each with 32 participants. The positive and negative groups were asked to elicit the corresponding emotions through autobiographical recall; the neutral group answered questions unrelated to emotion. Participants rated their emotional valence, arousal and conformity with the supplied emotion words. They then rated appreciation, comprehension and surprise for incongruity-resolution and nonsense jokes.

The positive emotion group rated nonsense jokes as funnier than the negative emotion group did. Within the negative emotion group, nonsense jokes were less amusing than incongruity-resolution jokes. There were no significant differences between the three groups in comprehension or surprise at the jokes.

Punchlines in a second language

A September 30, 2026 study in Bilingualism: Language and Cognition involved 42 learners of English as a foreign language, divided into intense- and limited-exposure groups. They read jokes in which identical punchlines followed humorous or coherent setups and judged funniness. Responses were slower to humorous endings than to coherent endings, and accuracy was higher in the intense-exposure group.

The researchers also reported a larger N400 response to humorous endings than to coherent ones. A P600 effect to humorous endings was observed in the intense-exposure group but not the limited-exposure group. The authors interpreted that finding as suggesting greater engagement of reinterpretation under intense exposure. The material consisted primarily of one-liners and riddles, and the study lacked a native-speaker baseline.

Matching, judging, explaining

A July 2023 paper published by the Association for Computational Linguistics used the New Yorker Cartoon Caption Contest to test three tasks: matching a joke to a cartoon, identifying a winning caption and explaining why a winning caption was funny. The researchers examined models given cartoon images and language-only models supplied with descriptions of the visual scene. The tasks involved complex, often surprising relationships between images and captions, with frequent indirect and playful allusions to human experience and culture.

In its July 26, 2023 account, Cornell University reported that the researchers compiled more than 700 contests spanning 14 years. Each included a captionless cartoon, the week's entries and three editor-selected finalists; some also included crowd estimates of submission quality. Matching required choosing the cartoon's finalist caption from distractors that had been finalists in other contests. Quality ranking compared a finalist with a nonfinalist, while explanation required free text connecting a caption to its cartoon.

Cornell reported 62% accuracy for the best model performance in the multiple-choice matching test, against 94% for humans in the same setting. The paper also reported that human-written explanations were preferred to the best machine-written explanations in more than two-thirds of comparisons, even with correct visual-scene descriptions supplied.