GPT-4o tested as cartoon caption writer—funny, but misses irony
Researchers tested GPT-4o on IEEE cartoons: AI gets laughs but flubs cultural references.
A team of researchers from Istanbul Technical University and Texas A&M University evaluated GPT-4o’s ability to generate cartoon captions, using IEEE Computer magazine cartoons as their testbed. By stripping away original captions and prompting the AI to produce replacements, they assessed how well the model understood visual context, humor, and narrative intent. The results, spanning 12 pages with 8 figures, show that GPT-4o can craft genuinely funny and topically relevant lines—sometimes even exceeding the originals in comedic timing or creativity.
However, the model stumbled on irony, cultural nods, and subtle contextual constraints, occasionally producing captions that were amusing but misaligned with the cartoon’s core joke. The authors conclude that AI is best positioned as an assistant rather than a replacement, enabling a hybrid workflow where cartoonists can rapidly iterate on AI suggestions, blend different humor styles, and retain full creative authority. The study highlights AI’s practical value for ideation in creative fields, not as a standalone creator.
- GPT-4o generated contextually relevant and often humorous captions for IEEE Computer magazine cartoons
- AI sometimes missed irony, cultural references, or intended meaning—diverging from the original joke
- Best used as an assistant: cartoonists can integrate AI suggestions to explore diverse humor styles while retaining creative control
Why It Matters
Hybrid human-AI workflows can accelerate creative ideation without sacrificing artistic control or nuance.