November 2026 · Independent humanizer benchmarks. Explore the evidence

Why HumanizeBench exists

A common evaluation for different kinds of tools

We wanted to compare downloadable humanizer checkpoints with the saved outputs of closed services using the same source texts and measurement profiles. A product description or model card cannot replace a recorded output.

Show the work

This site exposes the 33 source texts, 759 recorded rewrites, measured quality values, detector coverage, and unchanged scoring arithmetic. Related model artifacts remain identifiable; failed measurements stay visible.

Let the evidence set the limits

We do not claim to know hidden closed-source prompts or reproduce an unpublished judge rubric. Our substitutes and the historical nature of the closed-source outputs are documented. A final rank waits for complete measurement coverage.

How the evaluation works →