Custom harness with abliterated DeepSeek v4.1 fools Pangram 4.1, scoring 80–100% "human written"

IndraVahan · x · 2026-11-19

IndraVahan reports beating the Pangram 4.1 AI-text detector again with a custom writing harness, this time swapping in an abliterated DeepSeek v4.1 flash 1017. After a few optimizations, single-shot outputs scored 80–100% "human written" labels across every sample, with a screenshot showing a 100% Human Written verdict. His takeaway: evading current AI-text detection is far too easy.

Original post →

More from Fun

Fun channel →