AI text detection is a dead end; provenance and verification are the only path forward

shashib · x · 2026-08-20

Commenting on Adam Albright's Washington Post piece: LLM text detectors are doomed — models improve faster than detectors, and current tools already flag careful human writing, non-native English, and formal prose as "AI," while watermarks can always be stripped.

The deeper problem is cultural: just as we never demoted Word-printed documents in the 1990s, we should assume nearly all professional text will be AI-touched. The distinction that matters is not "was a model involved" but "can the origin and accountability be trusted."

For high-stakes content — medical, financial, official statements — the solution is stronger provenance: cryptographic signatures, authenticated channels, and machine-checkable watermarks. Most scams would fail that test.

Original post →

More from Safety

Safety channel →