AI Humour: The Audience Is Part of the Result

Picture a comedian delivering an AI-written joke to a laughing room. The words came from a model, but someone chose the line, found its rhythm and stood in front of the audience. That division of work matters when we read claims about how funny AI has become.

For creators and business owners, the interesting possibility is a workable collaboration. Research shows that generated material can earn favourable ratings and live laughter. Understanding how those results were produced helps us decide what to ask of a tool and what still needs a human performance.

A text-rating study tests a particular task

Drew Gorenz and Norbert Schwarz’s PLOS ONE paper, published on 3 July 2024, included a comparison between ChatGPT-3.5 and 105 non-professional human joke writers. A separate group of 200 US-based participants rated sampled responses without being told their source.

Across acronym, sentence-completion and roast tasks, the AI responses received higher average funniness ratings. The design used specific prompts and examples; humans supplied one answer per prompt while the model supplied twenty, from which responses were randomly sampled for evaluation.

This is evidence about written responses to those tasks, using that historical model. It does not establish who would deliver the better hour-long show, handle a difficult room or write the right line for a particular client. Those are further questions.

A live laugh includes a performer

Joe Toplyn and Ori Amir took a different approach in a workshop paper published on 19 January 2025. They compared selected Witscript jokes with selected jokes from a professional writer across eight topics. Co-author Toplyn operated the system and chose its submitted material.

The analysed recordings came from two sets by the same comedian at one US venue, with audiences of 35 and 15. The performer did not know which jokes were AI-written. The researchers measured laughter from the recordings.

Some AI material earned strong laughter, and the comparisons found no statistically significant difference between sources. The authors acknowledged that the small sample made the absence of a significant difference uninformative. It cannot establish general equivalence. The result belongs to a process involving selection, performance and two audiences.

Improvisation shows the collaboration more clearly

A study by Thomas Winters and Stijn Van der Stockt, published on 15 July 2025, examined Dutch-language improvisational comedy. Three professionals delivered their own jokes and GPT-4-generated material, producing 54 jokes for an audience of 40.

Audience ratings were slightly higher on average for human-created jokes, although AI material also produced highly rated examples. The important procedural detail is that an offstage assistant selected generated candidates, comedians also exercised selection, and some lines were adapted in delivery.

The authors identify that mixture as a limitation for evaluating pure generation and a resemblance to real collaboration. The single audience, restricted formats and performance differences limit generalisation. A successful line here is evidence about the combined process. It does not isolate a model’s contribution from the people making the material work.

Start with the occasion

My practical inference from these studies is to define the occasion before requesting jokes. “Make this funny” leaves the speaker, listener and purpose unspecified. A book-launch introduction, a newsletter aside and a reply to a disappointed customer each call for a separate editorial decision.

Consider a fictional author introducing a reading. An illustrative draft line might be: “My manuscript has gone through so many versions that ‘final’ is now a genre.” The target is her own revision process. She can decide whether it resembles her voice and whether her readers know enough about the writing process to enjoy it.

Place the same remark in an email about a delayed order and it might distract from the explanation the customer needs. The sentence has not changed, but the job of the message has. Choosing plain, helpful language is a valid outcome of the editing process.

A useful brief names the occasion, the audience, the intended mood and the subject available for playful treatment. Give the tool something specific to work with, then retain the option to use none of the suggestions. This is an editorial approach to try, rather than a validated formula for producing laughs.

Rehearse in the medium you will use

For the fictional author, the next step is a short rehearsal. She can read the line at her usual speaking pace, leave space after the final word and move into the reading. She is checking whether she can deliver the transition comfortably. A line she dislikes saying needs attention even if it looks clever on a screen.

For a short video, try the words alongside the planned expression and shot. For a written post, inspect the actual paragraph break and the sentence that follows. Keep the surrounding material visible: the intended audience will encounter an opening, a build-up and a continuation.

If you try a draft with a few willing listeners, keep their responses specific. Ask which part they understood, where they lost the thread and whether the line fitted the speaker. Treat the feedback as local editing information. A handful of reactions cannot establish what an entire audience will enjoy.

Keep the choices that make it yours

When revising, separate the possible sources of difficulty:

• The premise needs background the audience has not been given.
• The wording takes too long to reach the turn.
• The delivery feels awkward for this speaker.
• The line interrupts the purpose of the message.

Each calls for a different choice. More generated alternatives may help with wording; they cannot decide what your event is for. A performer may prefer a modest line that fits naturally over an elaborate joke that requires adopting another personality.

MaryChuks.com’s guide to preserving your writing voice with AI offers a related practice: protect meaningful expressions and keep authority over the final selection. Apply that to humour by retaining the phrasing, references and rhythm you are willing to stand behind.

The studies make room for optimism about AI-assisted comedy. They also give us a more useful way to describe success. Credit the generated material where it contributes, the editor who chooses, the person who delivers and the audience whose response completes the occasion.

Sources

• Gorenz and Schwarz: How funny is ChatGPT? — PLOS ONE, 3 July 2024.
• Toplyn and Amir: Can AI Make Us Laugh? — CHum workshop proceedings, 19 January 2025.
• Winters and Van der Stockt: Evaluating Humor Generation in an Improvisational Comedy Setting — 15 July 2025.