Blog
How AI-text detection works, what it can and cannot prove, and how to read a detector score honestly.
What zero-shot AI-text detection actually means
Most detectors are trained on a handful of models. Zero-shot detectors run on statistical fingerprints instead. Here is what that buys you — and what it still cannot prove.
DetectGPT, Fast-DetectGPT, Binoculars and DetectLLM: beyond the perplexity baseline
Naive perplexity is a weak detector. The newer zero-shot methods change what they measure. A plain-English tour of how they differ and where each one fails.
Thresholds, calibration, and the false-positive problem
Every detector has a dial. Turn it down and you miss AI text; turn it up and you accuse real writers. Why thresholds are a policy choice and calibration is the only honest fix.
Why no AI detector can certify that a human wrote something
Detection is a statistical estimate, not a proof of authorship. The difference matters for schools, publishers, and anyone accused of using AI. Here is exactly why the certificate is impossible.
What the EU AI Act Article 50 means for AI-text detection
Article 50 makes transparency a legal duty for AI providers and deployers. It does not make detection a proof of authorship. How the two interact, and where detection fits.
How AI-text detectors are actually evaluated: AUROC, AUPRC, and the benchmark problem
A claimed accuracy number means nothing without knowing the benchmark. What AUROC and AUPRC measure, why the Human-Written vs Machine-Generated benchmark exists, and how to read vendor claims.
The state of AI-text detection in 2026
A roundup of where AI-text detection stands: the methods that work, the limits that do not move, how regulation is changing the picture, and what to actually do about the results.