How LLMs are impacting computational humor, an AI subfield that uses computers to create jokes, as some comedians begin using AI chatbots for improv and roasts
Machines are getting better at making us laugh — and challenging our notions of humanity in the process X: @mirowskipiotr and @improbotics . LinkedIn: Patrick Reinmoeller X: Piotr Mirowski / @mirowskipiotr : The AI improv comedy work we have been building with @korymath at @improbotics was featured in an in-depth feature on comedy and computational humour by @gilliantett for @FT. Read all about it at https://www.ft.com/... @improbotics : Thank you @gilliantett for the fantastic interview for @FT! “Our audiences are attracted to our shows by robots, but it is the humans that make them laugh.” Today and tomorrow 6pm at @Gildedballoon #EdFringe2023! Tickets link: https://tickets.gildedballoon.co.uk/ ... https://www.ft.com/... LinkedIn: Patrick Reinmoeller : Bots make humans laugh(able). Will their jokes entertain us or be on us? — Unfortunately this fine FT article on the recent advances may remain hidden for non-subscribers.
Context & Ripple Effects
This report captures an early creative use case for conversational models: performers are bringing them into live formats where timing, audience context and collaboration matter. It follows evidence that GPT-3.5 joke outputs were overwhelmingly repetitions of a small set of learned jokes in a test of model-generated joke originality, making live human framing central rather than incidental.
The story also sits ahead of a broader product push to make assistants more engaging: later coverage described major AI developers pursuing funnier chatbots, while a DeepMind study of working comedians found persistent limits on originality and comic quality in professional comedians’ evaluation of AI-assisted comedy.
First-order effects
- Comedians experimenting with improv and roasts gain a rapid conversational prompt source and an onstage foil, while computational-humor researchers get a visible real-world setting to test LLM behavior.
- For groups such as Improbotics, audience interest can shift toward the interaction between performer and model, not merely a prewritten AI joke.
Second-order effects
- Chatbot makers have an incentive to improve humor, tone and turn-taking as engagement features—an agenda reflected in developers’ efforts to make chatbots funnier.
- The gap between plentiful generated material and reliably novel, situational comedy raises the value of human editing, performance skill and evaluation for entertainment uses.
Third-order effects
- If these experiments persist, comedy is likely to become a proving ground for whether generative AI can function as a collaborative interface rather than simply a content generator.
- The category may separate tools that generate familiar text cheaply from systems that can sustain context-sensitive interaction; originality remains a limiting test rather than an assumed capability.
The trend: Generative AI is moving from producing standalone creative outputs toward human-led, real-time collaboration, with usefulness determined by context and judgment as much as fluency.