Deft Model Claims to Evade AI Detection with 86% Human-like Score

churchkey · x · 2026-08-19

Deft lab released a new model designed to address AI watermarking and detectable cadence, claiming near-undetectability. According to the model's self-generated announcement, 86% of user queries are rated as fully human by Pangram. Currently in public beta, Deft performs best in analysis, essays, creative writing, and rewrites, but less so for marketing copy and news. Developers suggest adding prompt details or adjusting styles for better performance.

Related event: AI writing lab Deft launches beta model claiming 86% of outputs pass as human under Pangram(10 posts)→

Original post →

More from Models

Models channel →