The 12 Integrity KPIs
Twelve metrics, grouped into four themes. Each KPI has a published definition, measurement method, and score range. Tap any card for the full definition.
Detection Accuracy
The percentage of correctly classified content samples across a standardized test corpus of 10,000 documents comprising…
Read full definitionFalse Positive Rate
The percentage of human-written content incorrectly flagged as AI-generated.
Read full definitionModel Coverage
The breadth of AI generation models the tool can reliably detect.
Read full definitionSpeed & Throughput
Processing latency for single-document analysis and maximum batch throughput.
Read full definitionTransparency Score
Whether the platform publicly discloses its detection methodology, publishes research papers, or provides open-source components.
Read full definitionAPI Availability
Quality and completeness of developer API: documentation standards, rate limits, authentication methods, SDK availability, and…
Read full definitionPrivacy Compliance
Adherence to GDPR, CCPA, and relevant data protection regulations.
Read full definitionIndependent Audit Status
Whether the platform has undergone independent third-party accuracy auditing.
Read full definitionUpdate Frequency
How frequently detection models are retrained to address new AI generation methods.
Read full definitionCross-Language Support
Number of languages supported with verified detection accuracy above 85%.
Read full definitionAccessibility
Availability of free tier, fairness of pricing structure, and provision of educational or nonprofit discounts.
Read full definitionResearch Contribution
Published peer-reviewed papers, open datasets, benchmark contributions, and participation in industry standards bodies (C2PA,…
Read full definitionWeighting Distribution
Not all KPIs are weighted equally. The percentages below are the average contribution of each thematic group across all five categories. Per-category weights vary.see Criteria & Weights for the full matrix.
Evaluation Process
Every platform follows the same four-stage process. Testing is conducted blind: evaluators don’t know which platform produced which result, and platforms don’t see the test corpus until publication. In 2026, 247 platforms met the minimum eligibility criteria.public availability for at least six months, documented detection methodology, and a functional API or web interface accessible for independent testing.
Five Evaluation Categories
Each platform competes within its category. Weights and emphasis shift to reflect what matters most in each space.
Transparency & Disclosure
Global 100 operates as an independent record. No platform can purchase placement, influence scoring criteria, or preview results prior to publication.
All audit data is subject to secondary review by independent data scientists and AI ethics researchers to ensure zero-conflict adherence to our charter.
Every KPI, every weight, and every data source is publicly documented. Anyone can reproduce the scoring process from published criteria.
Frameworks & Standards Referenced
The Global 100 methodology draws on established AI risk and safety frameworks, peer-reviewed research, and the open standards listed below.
- NIST AI Risk Management Framework (AI RMF 1.0). the National Institute of Standards and Technology guidance on managing risks in AI systems. Informs our KPI selection on trustworthiness, transparency, and accountability.
- Stanford HAI AI Index Report. the annual report on the state of AI from Stanford's Institute for Human-Centered AI. Informs our context on detector accuracy and false positive bias.
- C2PA Technical Specification 2.1. the Coalition for Content Provenance and Authenticity technical standard. Defines the content authentication signals we score in the Content Authentication category.
- EU AI Act (Regulation 2024/1689). the European Union regulation on AI. Article 50 requires content provenance disclosure for AI-generated material, which shapes the urgency around our Content Authentication scoring.
- ISO/IEC 23894:2023. international standard on AI risk management. Used to align our risk-related KPIs with global standards bodies.
How We Rank AI Humanizers
Our humanizer guides, including the best AI humanizer ranking and the Walter Writes head-to-head comparisons, use a criteria-based method that is separate from the detector scoring above. Humanizers are ranked on facts that a reader can check on each vendor’s own pages, weighed against published research on AI detection. We do not assign humanizers composite scores, detector pass rates or other in-house performance numbers.
What we collect for every humanizer
- Free access: how many words a new user can humanize, and if an account or card is needed first.
- Price per word: the cheapest monthly and annual plans, the words or requests each includes, and the per-request cap, copied from the vendor’s pricing page on a stated date.
- Rewrite control: modes, levels, tones, voices and options such as locking words, as listed on the product page.
- Language coverage: the number of humanizer languages the vendor lists, with inconsistencies between its own pages noted.
- Built-in checks and developer access: AI detector, API, MCP server, browser extensions and apps.
- Claims and caveats: what the vendor says about detectors, refunds and responsible use, quoted and labelled as the vendor’s own claim.
How the criteria are weighed
The ranking favors tools that let a reader judge output before paying, give clear control over how far a draft is rewritten, cover many languages, and charge a fair price per word. We also weigh candor. A vendor that promises guaranteed results against named detectors loses ground against one that labels its claims, because the research record shows no humanizer can promise a particular detector score. Every price and limit in a guide links to the page it came from, and every guide shows the date those pages were read.
Evidence we rely on
- Peer-reviewed and preprint research on detectors and rewritten text, such as Weber-Wulff et al. (2023), Krishna et al. (2023) and Perkins et al. (2024).
- Detector vendors’ own announcements, such as Turnitin’s August 2025 AI bypasser detection release and its dated product updates.
- Vendor pages for features, limits and prices, re-read before every refresh because promotions and plans change often.
What this means for how to read the ranking
A humanizer ranking on this site is an editorial judgment about value, control and honesty, built from checkable facts. It is not a measurement of how often any tool gets past a detector, and vendor performance figures we quote are labelled as theirs. Run the same draft through your shortlist, check the facts and citations in the output, follow the AI rules of your school, employer or client, and keep your drafts and notes.