LaughBench

LessWrong contributor Taylor G. Lunt has introduced LaughBench, a benchmark designed to test whether AI models can generate truly funny, novel jokes that actually make people laugh. So far, no model has succeeded in producing genuine laughter, though frontier models like GPT-5.6 Sol and Fable can occasionally coax a smile. According to the LaughBench framework, humor is a robust test of real intelligence: generating comedy requires modeling someone's mind, anticipating false inferences they'd make, and avoiding them yourself — a cognitive challenge that correlates with high IQ. Lunt argues that existing benchmarks like math and programming can be solved through memorization and computation without true general intelligence, whereas laughter is objectively difficult to fake: you either laugh or you don't. He frames the search for authentic intelligence benchmarks as critical, since the most dangerous — and most beneficial — AI capabilities likely require genuine general intelligence.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton