Turns out AI researchers have a sense of humor about humor research.
There’s a new academic shared task called MWAHAHA — “Models Write Automatic Humor And Humans Annotate” — part of SemEval 2026, the same workshop series that’s been running competitive NLP benchmarks for two decades. The goal: push AI systems to generate jokes under real constraints, then have actual humans judge whether they land.
One of the standout submissions comes from a team publishing under the name lmfaoooo. Their core finding is refreshingly honest: today’s models aren’t actually bad at generating jokes — they’re bad at knowing which of their jokes are good. Ask an LLM for humor and it’ll happily produce dozens of plausible one-liners; the hard part is picking the funny one. Their fix is a two-stage pipeline: generate a pile of candidates, then use a model trained on human preference data to select the winner.
It’s a nice reminder that even in 2026, teaching a machine to have a joke is easier than teaching it to know it has one.
Link to the paper for anyone who wants to go deeper: https://arxiv.org/pdf/2606.00022
Official MWAHAHA task page: https://pln-fing-udelar.github.io/semeval-2026-humor-gen/
CodaBench competition page: https://www.codabench.org/competitions/9719/