AI on Trial is a special episode of Human in the Loop, where we take a deep dive into Model Autophagy Disorder (MAD)—a growing risk in artificial intelligence systems. From feedback loops to synthetic data overload, we unpack how models trained on their own outputs begin to degrade in performance and reliability. With real-world examples, emerging research, and ethical implications, this episode explores what happens when AI starts learning from itself—and what we can do to prevent it.
💡 Whether you're an AI engineer, researcher, or just AI-curious, this episode gives you the tools to recognize, explain, and respond to MAD.
Featured Tool:
Try out the companion tool featured in the episode:
MADGuard – AI Explorer
A lightweight diagnostic app to visualize feedback loops, compare input sources, and score MAD risks.
Read the deeper explainer blog:🔗 What Is Model Autophagy Disorder? – Human in Loop Blog
A plain-language breakdown of the research, risks, and terminology.
Other Detection Tools & Frameworks
DVC – Data Version Control
https://dvc.org/
Label Studio – Open-Source Data Labeling Tool
https://labelstud.io/
DetectGPT – Classify AI-generated Text
https://arxiv.org/abs/2301.11305
Grover – Neural Fake News Detector (Allen AI)
https://rowanzellers.com/grover/
References-
Alemohammad et al. (2023). Self-Consuming Generative Models Go MADPaper introducing MAD and simulating performance collapse in generative models.🔗 arXiv:2307.01850
Yang et al. (2024). Model Autophagy Analysis to Explicate Self-consumptionBridges human-AI interaction with MAD dynamics.🔗 arXiv:2402.11271
UCLA Livescu Initiative – Model Autophagy Disorder (MAD) PortalResearch hub on epistemic risk and feedback loop governance.🔗 https://livescu.ucla.edu/model-autophagy-disorder/
Earth.com (2024) – Could Generative AI Go MAD and Wreck Internet Data?Reports on future data degradation and the "hall of mirrors" risk.🔗 https://www.earth.com/news/could-generative-ai-go-mad-and-wreck-internet-data/
New York Times (2023) – Avianca Airline Lawsuit Involving ChatGPT BriefsLegal case where synthetic text led to real-world sanctions.🔗 https://www.nytimes.com/2023/05/27/nyregion/avianca-airline-lawsuit-chatgpt.html