Hacker News
Gemini 3.6 Flash Went 2-for-3 on Counterfeit Lip Tint, and Found That the Real Product Has Typos Too
A August 26 writeup photographed three packages of Rhode Peptide Lip Tint, six photos each, and asked Gemini 3.6 Flash with Thinking whether each was authentic. It correctly flagged both counterfeits but called the genuine Sephora-sourced package fake, for 67% accuracy on a three-item sample. The useful result is the failure mode: the model leaned on typo detection as a counterfeit signal, but the authentic packages carry two typos in their own ingredient lists, and glare and shadow artifacts confused it further.
↳ Follow the thread