AI Haiku Study: GPT-5 and Gemini 2.5 Poems Indistinguishable From Human Work
s_scardapane · x · 2026-09-21
- A Japanese team's paper (full paper at SCIS&ISIS 2026, arXiv:2609.15511) used few-shot prompting to generate Japanese haiku with a heterogeneous set of LLMs, then mixed them with human-written poems in a survey of Tokyo university students.
- Recognition accuracy varied by model: GPT-5, Gemini 2.5, and StableLM-7B hovered at chance (0.50), while LLM-JP, Gemma-2B, and LLaMA-2 were moderately detectable (0.59–0.67), with strong item-level variation.
- Aesthetic ratings (fluency, coherence, poeticness) predicted perceived humanness but not correct classification, revealing an attribution bias: aesthetic judgment and true authorship detection dissociate.
More from Fun
- 529 Days Can Change a Man: Sam Altman No Longer Wears Hats Indoors — SumitGup · 2026-09-21
- New Yorker cartoon nails how we hoard browser tabs and bookmarks — SuB8u · 2026-09-21
- AI Can Now Accurately Colorize Black-and-White Film, New Demo Shows — TinfoilTricorn · 2026-09-21
- Cloudflare AI-Generated Landing Page Sparks Debate Over Acceptable AI Slop — seatedro · 2026-09-21
- OpenAI vulnerability disclosure debate: researcher defends hack despite scope limits — jachiam0 · 2026-09-21
- Yacine: 'A bot is a bot, wet or dry' — blocking humans who post like LLMs — yacineMTB · 2026-09-21