Author(s): Hamza Boulahia

Originally published on Towards AI.

Inside the classifiers, watermarks, and theorems behind AI detection, and why none of them can reliably catch AI-generated text.

From the moment AI became good enough to write a whole essay or an article by itself, in late 2022, the need for a model that could reliably detect generated AI text arose. Schools, universities, and other institutions expressed their need for such a solution.

The article explains how AI text detectors are built—first with simple predictability metrics like perplexity, then (more recently) with trained transformer-based classifiers that map text into an embedding space and output a probability of “AI vs. human,” while noting that even these models can’t reveal a crisp rule of detection. It argues that performance gains often reflect benchmark choices and training strategies that can introduce shortcuts and biases, while real-world conditions and adversarial attacks (including paraphrasing and detector-guided rewriting) keep breaking detectors. The author then presents a theoretical “ceiling” result: reliable detection is fundamentally limited by how similar the distributions of human and AI text can become, so improvements in fluency can push detectors toward random-guessing. Finally, the author shares hands-on tests with a commercial detector, finds surprising false positives even on human writing (and mixed texts that look obvious), and concludes that institutions should treat detector scores only as weak signals—not proof—because the underlying problem is likely impossible to solve permanently.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI