Fairness & Corrections

Build an AI humanizer that isn’t in the benchmark yet? Submit it for a future cycle. The rest of this page covers how vendors already tested can dispute a result.

How rankings are computed

Every rank on HumanizerBench comes from a deterministic scoring script run over a published cycle of inputs, humanized outputs, and detector scores. We do not hand-edit rankings. For the full formula, weights, and penalty rules, read the methodology page.

What we publish that you can audit

• All input prompts and their prompt_set_version.
• All humanized outputs, per humanizer, per sample.
• Every detector's parsed human-likelihood score (the raw_score ∈ [0, 1]) for every humanized output, published as detector-scores.json. We publish the parsed scores, not the detectors' raw API payloads.
• The scoring script (scoring.js, a self-contained JavaScript/ESM module) in our public repository.
• Methodology version history (methodology_version, scoring_version, prompt_set_version). Adding or removing a humanizer or detector is recorded by the cycle data itself — each cycle’s leaderboard and detector-scores files document exactly which were tested — so there is no separate config version stamp.
• Repository: https://github.com/HumanizerBench/humanizerbench.

How to dispute a result

Email contact@humanizerbench.com with the subject line Dispute: [Cycle], [Humanizer]. Include:

1. Which result you dispute (cycle, humanizer, sample ID).
2. What specifically looks wrong — for example, the published output isn't representative of what the tool produces for that input, the plan or settings we used were wrong, or a detector verdict or score looks off. A humanizer's output varies from run to run, so we don't expect an exact reproduction; point us to the discrepancy you can demonstrate.
3. Any details about your test setup (account tier, plan, settings, region, time of day) that would help us reproduce the issue.

We respond within 14 business days.

What we won't do

• Adjust rankings outside the published methodology.
• Accept payment for placement on any leaderboard.
• Modify a humanizer's score after publication without versioning the change.
• Run private tests off the record.

WriteHuman disclosure

HumanizerBench is operated by WriteHuman (https://writehuman.ai). We disclose this on every page. The longer version of why we built this and how we handle the conflict is on its own page. Our anti-bias commitments:

1. WriteHuman is tested using the exact same methodology and scoring as every other humanizer.
2. The scoring script is public and deterministic.
3. We have never adjusted rankings to favor WriteHuman, and we will publicly document any future change to methodology, scoring, or penalty rules.