Why don't you report a single 'bypass rate'?

Detectors disagree, so a single blended number hides useful detail. We publish per-detector results and fold an averaged bypass component into the overall score at a fixed weight.

Do you use default settings for each tool?

Yes. We use each tool's default mode on identical inputs so the comparison is fair and reproducible. Aggressive custom settings can change results and aren't part of the standard test.

How do you compute rank movement?

We match each tool to its result in the previous cycle by its stable identifier — never by name — and compare ranks. New tools are shown as new rather than as a movement.

Can results change between cycles?

Absolutely. Detectors update and tools change, so scores and ranks move. That's exactly why we archive every cycle.

Methodology

How we test AI humanizers

Our goal is a comparison you can trust: the same inputs, the same detectors, and a transparent, weighted score applied identically to every tool.

The process

How a cycle runs

Every cycle follows the same six steps so results are fair and reproducible.

1. Fix the source texts

We assemble a set of AI-generated passages across writing categories and reuse them so every tool faces identical inputs.

2. Run each humanizer identically

Every humanizer processes the same passages with default settings, so differences reflect the tools, not the prompts.

3. Score against a detector panel

Each output is checked against multiple independent AI detectors. We record per-detector results, not a single blended number.

4. Assess meaning, readability & consistency

We check that the rewrite preserved meaning, reads naturally, and holds up across many samples.

5. Combine and rank

Weighted metrics produce one overall score per tool, and tools are ranked within the cycle.

6. Publish and archive

Every cycle is published in full and kept in the archive so results are transparent and comparable over time.

Scoring

Scoring weights

Every tool is scored on four metrics, combined into one overall score.

Detector bypass — 40%

Average performance across the detector panel.

Meaning preservation — 25%

Did the rewrite keep the original meaning and claims?

Readability — 20%

How natural the output reads to a human.

Consistency — 15%

Stability of quality across many samples.

A rewrite is penalised when it beats detectors but distorts meaning — evasion alone is never treated as success.

Methodology FAQ

Why don't you report a single 'bypass rate'?: Detectors disagree, so a single blended number hides useful detail. We publish per-detector results and fold an averaged bypass component into the overall score at a fixed weight.
Do you use default settings for each tool?: Yes. We use each tool's default mode on identical inputs so the comparison is fair and reproducible. Aggressive custom settings can change results and aren't part of the standard test.
How do you compute rank movement?: We match each tool to its result in the previous cycle by its stable identifier — never by name — and compare ranks. New tools are shown as new rather than as a movement.
Can results change between cycles?: Absolutely. Detectors update and tools change, so scores and ranks move. That's exactly why we archive every cycle.

Scoring by cycle

How each published cycle's overall score was produced. Weights and methodology versions apply only to the cycle that published them.

October 2026 · Methodology version not supplied: October 2026 overall scores are source-reported aggregate values. The source publishes bypass, meaning, readability, factual accuracy and length stability, but does not publish consistency scores, scoring weights, penalty rules or the underlying formula. Humanizer Arena preserves the supplied overall scores and does not recalculate them.
September 2026 · Methodology v2.1.0: September 2026 overall scores combine detector bypass (40%), meaning preservation (25%), readability (20%) and consistency (15%) under methodology v2.1.0, with the documented penalty rules applied.

Detector panel

Outputs are tested against these independent detectors, reported separately.

Pangram — Pangram detector included in the October 2026 aggregate bypass-rate table. results →
GPTZero — GPTZero detector included in the October 2026 aggregate bypass-rate table. results →
Originality.ai — Originality.ai detector included in the October 2026 aggregate bypass-rate table. results →
Copyleaks — Copyleaks detector included in the October 2026 aggregate bypass-rate table. results →
ZeroGPT — ZeroGPT detector included in the October 2026 aggregate bypass-rate table. results →

Writing categories

We score performance separately for each category of writing.

SEO content — Long-form web articles need natural flow without losing keywords or intent.
Marketing copy — Landing pages and ads reward tone control and a persuasive, consistent voice.
Students — Re-ranked for everyday student writing across subjects and levels.
Essays — Argumentative and expository essays reward tools that preserve structure and a consistent voice.
Academic writing — Citation-heavy, formal scholarly prose is difficult to rewrite without distorting claims or hedge language.

Cadence & archive

We publish a new cycle regularly and keep every past cycle available. Published cycles so far: October 2026, September 2026.