Everything you need to know about detecting AI-generated academic work
The system analyzes author biographies, methodology sections, and linguistic patterns to identify work likely generated by Claude, GPT-4, or other large language models. We cross-reference claimed author information against public databases and examine statistical properties that distinguish synthetic academic writing from human-authored work.
The filter flags submissions with high confidence (85%+) based on multiple independent signals. Detection improves when submissions contain fabricated author profiles or citations. False positives are low because we require multiple corroborating signals before flagging.
No. The system is tuned to minimize false positives. We only flag papers when multiple independent signals (author verification failure, statistical markers, structural anomalies) align. Legitimately human-authored work passes through cleanly.
Yes. Upload a PDF and the filter will run a diagnostic scan. You'll see which signals triggered (or didn't). Useful for understanding why papers flag or to validate that your human-authored work is clearly human to the system.
The API accepts PDF uploads and returns a risk score (0-100) plus detailed findings. Most conference systems integrate via webhook in under an hour. We provide sample integrations for OpenReview, Easychair, and custom platforms.
The flag is a signal for your review desk to investigate. Examine the author bio, run reverse-image search on claimed author photos, check claimed affiliations. Most AI-generated papers have easily-verifiable biographical claims that collapse under scrutiny.
Yes. If an author disputes the flag, they can provide author verification (passport scan, institutional email, institutional webpage bio). Most legitimate authors can quickly prove authorship.
Subscription-based by scanning volume. Most conferences use tier-2 (100-500 submissions/year) at $200/month. No per-submission fees, flat rate regardless of volume. Custom enterprise plans available.
Papers are scanned and results returned immediately. We don't store PDFs or run them through any training pipelines. All data is deleted after 24 hours. Results are never shared with third parties.
We've audited for bias across writing styles, languages, and disciplines. The system was trained on papers from conferences across computer science, biology, philosophy, and law. Report any bias patterns directly and we'll investigate.
Yes, which is why we keep false-positive rates low and make findings transparent. A flag is a signal for investigation, not an automatic reject. The burden is on the system to provide evidence.
Currently: Claude (all versions), GPT-4, GPT-3.5, Gemini, Llama-2. Detection signal improves over time as we add coverage for newer models.
Detection signals are updated weekly as new models release and writing patterns evolve. Subscribers get automatic updates at no extra cost.
The system can detect mixed-origin sections. If half a paper is human-written and half AI-generated, we flag the AI portions and provide a breakdown.
Still have questions? Email support@conference-shield.com or check our blog for case studies of detected papers.