Open-weight vs. commercial AI humanizers
The November 2026 HumanizeBench comparison uses 33 shared source texts and 23 evaluated entries. Open-weight entries averaged 33.93 out of 100 on the composite, and commercial entries averaged 37.76.
Results by cohort
Each column shows the arithmetic mean of the published entry scores within that cohort. These results describe the recorded entries and detector responses for this run.
| Metric | Open-weight ? 13 entries | Commercial ? 10 entries |
|---|---|---|
| Composite | 33.93 / 100 | 37.76 / 100 |
| Detector human-likelihood | 19.4% | 41.5% |
| Meaning preservation | 65.4% | 44.6% |
| Readability | 79.1% | 77.2% |
What the comparison covers
The commercial cohort uses historical published rewrites. The open-weight cohort uses recorded runs of identified repositories and formats. Some entries share a model family, so the means describe the included entries and carry no estimate of performance on unseen writing. Individual detector scores describe detector responses and cannot establish authorship.
Read the scoring rules and limits, inspect the source texts, or download the dataset. For a direct comparison on matching inputs, see browse the matched model comparisons.