Should AI Tell a Joke With You—or for You?

A group of friends sits around a table. One vents about the constant appearance of their aging sedan’s “check engine” light. Another, who uses a speech-generating device to communicate, wants to contribute a lighthearted comment, but typing it out word by word would stall the conversation. 

Instead, they tap the “joke” mode on an experimental, artificial intelligence-equipped app. It instantly generates a suggestion, the user selects it, and the device speaks the line: “At this point, your ‘check engine’ light is just a mood light.”

The group laughs. The friend grins and nods, “Seriously!” The device’s user hasn’t just delivered a joke; they’ve participated in a shared, cathartic moment.

This scenario is at the heart of research by the University of Maryland College of Information Assistant Professor Stephanie Valencia² and Cornell Tech information science Assistant Professor Thijs Roumen and computer science doctoral student Tobias Weinberg, who uses what’s called an Augmentative and Alternative Communication (AAC). Together, they’re exploring how AI can help rebalance the scales of conversational timing for people with disabilities who use AACs.

AI-mediated communication with AAC technology isn’t perfect: On one end of the spectrum, the user can type out their words, giving them full agency over their speech but is slow. On the other hand, a joke button is efficient but doesn’t produce anything personalized.

The researchers’ study, which won two awards at a conference this spring, explored this tradeoff by designing four AI-powered AAC interfaces, each offering a different level of user control. These ranged from a “Context Bubble Selection” interface, which lets users select specific parts of the conversation transcript to base the joke on, to a fully automated version that generates a complete joke with a single click.

The findings revealed a nuanced picture of the tradeoff between agency and efficiency. While users valued control, the social function of a quick quip often made speed the priority. The “Full-Auto” interface was frequently favored because, for a brief humorous interjection, achieving the goal of timely participation was more critical than perfect control over wording.

The function that humor is serving is the central issue, said Weinberg. “In this context, the joke serves as a medium to keep the flow of conversation,” she said. “It is not so much about the joke itself.” 

As Valencia² noted, spontaneous participation is difficult with standard AAC speeds, which allow users to communicate approximately only 12-18 words per minute.

But this efficiency comes with a profound risk: The AI doesn’t just assist—it influences. One study participant observed that the AI both “matched my intentions and also influenced it.” This gets at a major ethical dilemma in AI-mediated communication. If an AI constantly shapes what you say, can it be considered your voice?

The challenge, then, is to build AAC tools that empower rather than overwrite, Valencia² said. Users don’t want to sound perfect, but like themselves.