← Back to Portfolio
Appen · article

AI Detection Model Development for Data Quality Assurance

Co-authored research findings on applying internal AI detection models to real-world datasets, revealing critical limitations in existing detection systems

Co-authored comprehensive analysis demonstrating critical limitations in publicly available AI detection models through systematic evaluation against real-world curated datasets. The research revealed that three major AI detectors fail to achieve acceptable true positive and false positive rates for production environments, particularly problematic for crowdsourcing applications requiring human-authored training data.

Key Features

Systematic Model Evaluation: Benchmarked three different AI detection APIs against curated datasets to establish baseline performance metrics • False Positive Rate Analysis: Identified unacceptably high false positive rates in current detection models when applied to data annotation contexts
Production Viability Assessment: Established threshold criteria targeting false positive rates below 9% while maximizing true positive detection rates • Human-Centric Detection Framework: Contributed to development of improved detection methodology addressing crowdsourcing platform requirements

Technical Implementation

The research involved applying internal AI detection models to behavioral AI detection experiments using real-world curated datasets. The evaluation methodology focused on threshold adjustment techniques to optimize the balance between detection accuracy and false positive minimization. The findings highlight the contextual dependency of current detection models and their limitations in distinguishing increasingly sophisticated LLM-generated content from authentic human writing.

Skills & Technologies

Machine Learning AI Detection Data Quality Statistical Analysis Large Language Models Model Evaluation Data Science Python
// stay in touch

Get Answering Machines in your inbox.

Occasional notes on AI, coding agents, and where this is all headed — plus the very occasional origami diagram.

Arjun Kirti Patel
Chicago, IL · folded with care
© 2026 Arjun Kirti Patel · MIT Licensed