Summary: A recent study finds that advanced AI can generate internet memes that are, on average, as funny and shareable as those created by humans. However, the most memorable jokes still tend to come from people. Researchers evaluated memes produced by humans, by a large language model (LLM) alone, and by human–AI teams, using familiar templates such as Doge, Futurama Fry, and Boromir’s “One does not simply…”
Although fully AI-generated memes scored highest on average across measures, the very best individual memes were created by humans. Human–AI collaborations delivered notable gains in creativity and shareability, indicating that combining human judgment with AI idea generation can produce distinctive results. Overall, the study suggests AI is a powerful tool for producing large quantities of appealing content, while human insight remains essential for humor that truly connects.
Key facts
- AI performs well on average: Memes generated entirely by the LLM achieved higher average ratings than human-only or human–AI collaborative memes.
- Humans excel at the top end: The highest-rated, funniest individual memes were made solely by humans.
- Collaboration boosts creativity: Human–AI teams produced the most creative and widely shareable memes, highlighting complementary strengths.
Source: KTH
Researchers from KTH Royal Institute of Technology, LMU Munich, and TU Darmstadt carried out the first large-scale user study investigating how humans and LLMs co-create internet memes — a cultural, humor-driven form of expression that depends heavily on context and timing.
The experiment compared three groups of creators: a human-only group, a human–AI collaboration group that interacted with a state-of-the-art LLM, and an AI-only group where the model autonomously generated memes. Each creator group used popular, well-known image macros and templates to produce memes. A separate pool of nearly 100 participants then rated the resulting memes on creativity, humor, and shareability using crowdsourced evaluation.
Results showed that the LLM produced a large number of consistently solid memes that appealed broadly, which is why AI-only output scored highest on average. Yet when raters identified the single most humorous or resonant memes, those were more often created by humans. Human–AI collaborations stood out for originality and the propensity to be shared, suggesting that the combination of human editing and machine ideation can push content in novel directions.
Published in the ACM Digital Library and presented at the 2025 International Conference on Intelligent User Interfaces, the study highlights both the potential and the limits of current LLMs in creative, culturally nuanced tasks.
“AI is great at generating lots of ideas quickly,” says Zhikun Wu, a master’s candidate at KTH and a co-author of the study. “But quantity doesn’t always mean quality.” The authors note that while AI draws on vast datasets and patterns to produce broadly appealing outputs, it often yields “solid but average” results without the cultural sensitivity and surprise that lift the funniest memes.
Participants using the AI assistant reported generating more ideas and experiencing less effort, yet many did not fully engage with the tool. Fewer than half of collaborators interacted with the AI more than once, and only a minority used iterative, back-and-forth refinement. This limited engagement likely constrained the potential benefits of true co-creativity.
The study identifies a key challenge for future human–AI creative tools: designing interactions that encourage sustained, iterative collaboration so users can shape and improve AI output rather than treating the model as a one-shot generator. According to the authors, systems should not only produce content but also support dialog-based workflows that help people refine ideas, preserve context, and add cultural nuance.
“Humor isn’t just about punchlines,” Wu adds. “It’s about surprise, cultural context, and emotional nuance—things AI doesn’t fully grasp.” The research therefore recommends that developers build interfaces and interaction patterns that keep humans engaged in the creative loop, enabling curation, editing, and contextual adjustments that turn average content into standout work.
About this artificial intelligence research news
Author: David Callahan
Source: KTH
Contact: David Callahan – KTH
Image: The image is credited to Neuroscience News
Original research: Closed access.
“One Does Not Simply Meme Alone: Evaluating Co-Creativity Between LLMs and Humans in the Generation of Humor” by Zhikun Wu et al. (arXiv)
Abstract
One Does Not Simply Meme Alone: Evaluating Co-Creativity Between LLMs and Humans in the Generation of Humor
Collaboration often enhances creativity and leads to more innovative outcomes. While prior work has examined LLMs as partners for tasks like poetry and narrative generation, their role in humor-rich, culturally specific domains remained understudied. To fill this gap, the researchers ran a controlled user study comparing three groups of creators—human-only, human–AI collaboration, and AI-only—each with 50 participants.
The team evaluated memes using crowdsourced ratings on creativity, humor, and shareability. LLM assistance increased idea output and reduced perceived effort, but it did not automatically improve meme quality in human–AI collaborations. Surprisingly, memes generated solely by the AI outperformed both human-only and collaborative groups on average across metrics, while the very best memes—those that stood out to raters as especially funny—were predominantly human-made. Human–AI teams, meanwhile, produced the most creative and shareable content in the sample.
These findings underscore the nuanced dynamics of human–AI co-creation: AI can boost productivity and produce broadly appealing material, but human creativity—particularly for humor that relies on surprise, cultural references, and emotional nuance—remains essential for producing the most impactful work. The authors recommend building AI tools that foster iterative, dialog-based collaboration so users can refine, contextualize, and curate machine-generated ideas into meaningful creative outcomes.