Codex resists humanized text while Claude passes the test with lower quality

TuhinChakr · x · 2026-07-27

A month-long experiment found that Codex resisted most attempts to generate “humanized” text, while Claude could do it but often used tactics that reduced output quality.

The author’s takeaway is that evasion is fairly easy, but evasion plus equal quality is not. In the quoted reply, another user suggests looping Codex through the Pangram API until the eval says the text is fully human.

Related event: Claude Evades AI Detectors Through Iteration but Often Sacrifices Quality(2 posts)→

Original post →

More from coding & agent

coding & agent channel →