2
Independent test finds current AI models unreliable at identifying dangerous mushrooms
2026-09-02
The Register reported on September 2, 2026 on an independent test by Piotr Migdal running 1,040 mushroom photos across 16 models. Gemini 3.8 Flash scored best (65% correct on first guess, 85% within its top five), while Qwen3.8-27b scored worst (13% first-guess, and misidentified poisonous mushrooms as edible 36% of the time). Meta's Muse Spark 1.2 had the lowest false-positive rate (8%) largely because it declined to guess in ambiguous cases.
Significance 2: A real independent comparative test with safety-relevant failure rates, but on a one-off task outside any tracked axis, so it stays at context level rather than verdict-moving.
technology Safety / alignment Benchmark result
Sources
- press AI-assisted mushroom hunting is a recipe for a bad trip retrieved 2026-09-03