Claude Code experiment says Pangram's AI detector can be beaten after 17 rewrites
_akpiper · x · 2026-07-27
A repost of an experiment showing that Pangram's AI-writing detector can be beaten with iterative feedback: after 17 versions, the score moved from 0.999 “Fully AI Generated” to 0.318 “Fully Human Written.” The author says the detector was hardened recently, so the earlier Claude Code approach no longer worked the same way.
The key lessons were:
- The winning approach was not to “sound human,” but to surgically remove specific tells like em dashes, tricolons, and colon-style explanations.
- Controlled experiments beat intuition-driven rewrites; changing multiple tells at once worked better than one at a time.
- The content itself stayed largely unchanged; what mattered was punctuation and syntactic habits.
- The LLM struggled to discover these tells on its own, so human-led hypothesis testing was essential.
Related event: Claude Evades AI Detectors Through Iteration but Often Sacrifices Quality(2 posts)→
More from coding & agent
- AI is not replacing developers; it is turning them into agent managers — pchandrasekar · 2026-07-27
- Claude Code allowed an adversarial test while Codex refused it — random_walker · 2026-07-27
- MemU pitches cross-device memory for Claude Code, Codex, Cursor, and more — JaynitMakwana · 2026-07-27
- EvoCode adds 227-turn agent evaluation for changing requirements — _philschmid · 2026-07-27
- Nick Cammarata says AI code can do 97% of the job but still needs human judgment — morqon · 2026-07-27
- No-code job applications with customized resumes using Claude Code — Huge_Tea3259 · 2026-07-27