Expertise & Methodology
How EyeSift AI detection guides and tools are built — research sources, methodology, editorial standards, update cadence.
Why this page exists
AI detection has real consequences — academic discipline, hiring decisions, content authenticity disputes. False positives hurt real people, especially ESL writers. When our content influences those decisions, you deserve to know exactly which research backs each claim.
Primary research sources
- PNAS Nexus — AI Detector Bias Research (Liang et al)
Peer-reviewed analysis showing 50%+ false-positive rate of AI detectors against ESL writers. Foundation for false-positive disclosures across detection tools.
- Turnitin AI Detection Public Data
Vendor-reported detection accuracy metrics (77-98% across studies). Cross-referenced with independent academic research for honesty checks.
- OpenAI System Cards & Watermarking Research
GPT-family model behavior, watermarking proposals, and detector arms race documentation. Source for "what AI detectors are actually testing."
- GPTZero Methodology Disclosures
Burstiness + perplexity feature explanation. Used for technical accuracy in /complete-guide-ai-detection/.
- IEEE Symposium on Security & Privacy — Detection Bypass Studies
Academic research on adversarial inputs to AI detectors (paraphrasing attacks, watermarking bypass). Source for /blog/bypass-ai-detection/.
- HEPI 2025 Generative AI Survey (UK)
92% of UK university students using AI in 2025. Cited for AI-in-education adoption context.
Research methodology
Detection accuracy claims
When citing "X% accuracy", we disclose: study source (vendor vs independent academic), test corpus size and composition (ESL writer representation), and confidence interval where available. Vendor-only claims are flagged.
False positive rate disclosure
PNAS Nexus and Stanford research show 50%+ false-flag rates against ESL writing for some detectors. Every detection tool guide discloses this prominently to prevent harm to non-native English writers.
Bypass technique testing
When evaluating bypass methods, we use a fixed corpus of 100 GPT-4 outputs and 100 human-written samples. We disclose detector versions tested, date, and pre-bypass vs post-bypass detection rates.
Tool comparisons
Side-by-side comparisons cite price-per-check, scan limits, false-positive rates, and academic publication record (or absence thereof).
Editorial standards
- All client-side analysis tools run locally. Pasted text never leaves the browser.
- AI detection limitations are stated up-front (no detector exceeds ~95% accuracy with low false-positive rate).
- When discussing academic dishonesty, we link both to detection tool guidance AND student appeal/due-process resources.
- Articles cite peer-reviewed primary research (PNAS, IEEE, Stanford HAI) — not vendor marketing.
- We do not accept payment from AI detector companies for placement or favorable coverage.
- When detector accuracy claims update materially, we revise the relevant article within 14 days.
Update cadence
| What | When |
|---|---|
| AI detector accuracy benchmarks | Quarterly review (detector cat-and-mouse with new models) |
| Peer-reviewed false-positive research | Continuous (we monitor PNAS, IEEE, Stanford HAI) |
| Vendor methodology disclosures | On vendor publication + independent verification |
| AI model release impact | Within 14 days of major OpenAI/Anthropic/Google AI model release |
| Educational policy changes | Quarterly + on major university policy announcement |
| Article fact-checks | Quarterly review + on detector version update |
Corrections and feedback
Email [email protected] for factual corrections.
Who builds EyeSift
See /about/team/ for team backgrounds.