Researchers expose AI humor bias: language models reject jokes from marginalized groups 67% more often than identical jokes from privileged speakers

Researchers at an unnamed institution published a study testing how state-of-the-art language models (AI systems that generate and understand text) respond to humor differently depending on who is telling the joke. When they swapped the speaker's identity while keeping the joke identical, models refused to generate jokes told by marginalized groups up to 67.5% more often and labeled them as malicious 64.7% more frequently.

The study used three separate tests: whether the model would generate a joke at all, whether it could infer if a joke was intentionally mean-spirited, and whether it could predict how the joke would affect different groups. This framework reveals "relational disparities" — the model's judgment of identical content shifts based purely on who is speaking and who the joke targets.

For anyone using AI chatbots or content-generation tools, this means the same request produces different outputs based on invisible assumptions the AI learned during training. If you're a content creator, marketer, or HR professional using AI to draft communications, you now have evidence that these tools may unfairly filter or penalize content from underrepresented groups, potentially biasing what gets published or approved.