Abstract and 1. Introduction

  1. Methods
  2. Quantitative Results and Creativity Support Index
  3. Qualitative Results from Focus Group Discussions
  4. Discussion
  5. Mitigations and Conclusion and Acknowledgments
  6. Ethical Guidance References

A. Related Work on Computational Humour, AI and Comedy

B. Participant Questionaire

C. Focus

This section reviews computational humour literature. As foreshadowed in the Introduction (Sect. 1.1.1) and discussed in the Section on the importance of context in humour and comedy (Sect. 5.2.2), humour remains an elusive goal for AI.

A.1 Humour generation, from template-based systems to prompting

Amin and Burghardt [2], Winters [106] and Veale [102] provide extensive surveys on the history of computational humour generation and early approaches, like hand-coded rules [109] and templates for puns, riddles and acronyms [13, 97], or templates for automaticallygenerated slide decks for improvised “Powerpoint Karaoke” [108].

With the advent of LLMs, able to generate grammatically correct sentences in a given writing style, writers can focus on joke structure, guiding LLM-based systems to replicate such structures from examples of jokes [110], or use evolutionary computing to evolve text to become more satirical [107]. More broadly, Kaddour et al. [58] summarise recent work on using LLMs for creative applications such as (potentially humorous) story generation, including “Recursive Reprompting and Revision” [113] or interactive writing tools like “WordCraft” [53] and “Dramatron” [71].

A.2 LLMs for comedy performance

LLMs have been deployed since at least 2016 for live performance on stage, including in improvised comedy [16, 67, 70], short comical film scripts like Sunspring, and song lyrics for musical comedy like Beyond the Fence[15]. More recently, LLMs have been involved in (involuntarily) comical , absurdist productions of Prague-based company THEaiTRE [88, 89, 91], and semi-improvised comedy Plays By Bots[16] [71]. The success of these performances with AI primarily relied on the skills of the actors, who could invent subtext and add interpretation to AI-generated text [70].

Authors:

(1) Piotr W. Mirowski∗, Google DeepMind London, UK (piotrmirowski@deepmind.com);

(2) Juliette Love∗, Google DeepMind London, UK ( juliettelove@deepmind.com);

(3) Kory Mathewson, Google DeepMind Montréal, QC, Canada (korymath@deepmind.com);

(4) Shakir Mohamed, Google DeepMind London, UK (shakir@deepmind.com).

This paper is available on arxiv under CC BY 4.0 license.

[15] https://www.theguardian.com/stage/2016/feb/28/beyond-the-fence-reviewcomputer-created-musical-arts-theatre-london

[16] Review: https://12thnight.ca/2022/08/13/oh-no-bots-have-invaded-theatre-andthey-can-do-it-plays-by-bots-a-fringe-review/