Pragmatically speaking: humor in interactions with AI

Humor in human-AI interactions is investigated through a metalinguistic-pragmatic lens, focusing on how users engage with AI voice assistants like Alexa and ChatGPT. Prior to LLM-based systems, humor in human-AI interaction research focused primarily on evaluating machine-generated humor, leaving gaps in understanding how humans initiate and respond to humor with conversational agents. A systematic literature review of six studies (published 2016-2024) examining voice assistants (Alexa, Siri, Google Assistant, Cortana) reveals that users employ humor to explore systems and test relationships rather than for entertainment. The most frequent category of humor is personality questions. Addressing gaps in knowledge, this work presents a multimodal, multilingual dataset from 2016 to 2025 comprising three complementary data sets: (1) short-form videos from YouTube and TikTok featuring families interacting with AI voice assistants in English, Italian, and Spanish; (2) video recordings of interactions with Alexa in English and Italian; and (3) posts from Facebook, Instagram, and TikTok that explicitly signal humorous intent through linguistic and paralinguistic markers (English and Italian). Three distinct patterns emerge from Study 1: English videos positioned elderly users as objects of amusement, with the person frequently responsible for communication failure (39%); Italian videos centered on language barriers rooted in dialect use, with fewer instances of human-attributed failure (23%); while Spanish videos celebrated successful connection with zero instances of human-attributed failure. Study 2 reveals the fundamental asymmetry: people engage socially while AI generates language. Participants’ metalinguistic commentary tests AI theory of mind, marks embodied/cultural knowledge boundaries, and constructs meaning from incoherent responses, revealing assumptions and common ground strategies with non-human interlocutors. Study 3 analyzes 99 humorous social media posts targeting AI communication failures (comprehension, turn-taking) and successes (ChatGPT’s excessive perfection). Italian posts uniquely imagine dialect-speaking AI, performing regional identity. Humor functions as a metacommunicative resource for negotiating boundaries with systems simulating social engagement.