A Google DeepMind study involving 20 professional comedians who already use AI in their work finds LLMs struggled to produce material that was original or funny
Rhiannon Williams / MIT Technology Review :
Context & Ripple Effects
Professional experimentation with chatbots in improv and roasts had already brought computational humor into working comics’ toolkits, as covered in early reporting on AI-assisted comedy. This study adds practitioner feedback to that emerging use case rather than treating humor as a purely technical benchmark.
The finding also aligns with an earlier test in which generated jokes were frequently repetitive rather than novel, documented in research on repeated GPT-3.5 jokes. That makes originality—not merely producing joke-shaped text—a central constraint for creative deployment.
First-order effects
- For the participating comedians and similar users, LLM output is less dependable as finished material and more suited to prompting, iteration, or supporting a human writer’s process.
- Google DeepMind gains evidence that standard language-model capabilities do not translate cleanly into a high-context creative task where surprise and distinctiveness matter.
Second-order effects
- Labs pursuing more engaging assistants will need to evaluate humor with practitioner-led measures of novelty and audience fit, not just whether a model can generate a recognizable joke; this bears directly on industry efforts to make chatbots funnier.
- Creative users may place greater value on workflow tools that preserve authorship and facilitate revision over systems marketed as autonomous comedy writers.
Third-order effects
- If similar results persist across creative fields, generative AI’s commercial role is more likely to center on augmentation than substitution in work where originality is the product.
- The pattern highlights a broader synthetic-content constraint: models can expand supply, but abundant derivative output can reduce the value of generic generation unless human selection and differentiation remain central.
The trend: Creative AI is moving from demonstrations of fluent generation toward harder tests of whether model output is genuinely distinctive and useful within professional workflows.