100% of the time, some AI detectors are wrong

The Authors Guild tested ten pieces of human-authored work (written by professional authors from 2020–2022, prior to widespread use of ChatGPT) using five AI detectors. Two performed almost perfectly; two performed abysmally; and one was somewhere in the middle. This highlights an important point for writers, students, and professionals who may have their work evaluated by an AI detector: not all AI detectors are created equal, and some will always report everything as AI generated.

Pangram correctly classified all ten articles as non-AI generated, achieving a 0% false positive rate. Grammarly also did pretty well, with false positive rates ranging from 0% to 9% across the sample set. Originality.ai was pretty accurate, with most false positive rates falling between 0% and 1%. Sidekicker incorrectly reported that all of the articles were mostly AI generated, and two of the articles were 100%, meaning that the AI detector was completely certain that two pieces of clearly non-AI generated text were in fact AI generated. ZeroGPT was extremely variable, with false positive rates ranging from 5.3% to 76.3% across the articles.

1The paradox at the center of AI detection

As the Authors Guild points out, there is a fundamental asymmetry here that makes AI detection hard. The better a human writer is, the closer they are to the distribution of AI generated text. Most language models are trained primarily on professional writing, which tends to be well-written, coherent, structured prose — exactly the kind of writing that professional writers create. So the features that distinguish AI generated text and the features that distinguish high quality human writing are starting to overlap. As Max Spero, CEO of Pangram (the tool that did well in the test), pointed out: “AI models usually trip up due to uniformity, especially in argument construction. Not necessarily with more formal/carefully crafted text.”

So while it may be possible to spot low quality AI generated text because it follows a more limited range of argumentation, it may be impossible to tell the difference between high quality AI generated or AI assisted text and high quality human writing using an AI detector. As a result, the Authors Guild warns that no single detector should be trusted to make a determination, since a false positive is not a small error. If you’re a writer, a false positive could cost you a contract, lead to rejections, or tarnish your reputation.

2What writers and builders should take from this

If you think your writing might be evaluated by an AI detector (e.g. as a student turning in an assignment, a journalist turning in an article, or a contractor with a deliverable that gets checked), this test offers actionable advice. There is a big difference between Pangram (0% false positives) and Sidekicker (100% false positives on everything). These are not small calibration differences. They represent fundamental differences in the quality of the tools. Not all AI detectors are made equal, and using the wrong one can have serious consequences.

If you are building a platform that uses AI detection (e.g. plagiarism detection, submission filtering, content moderation), the message is the same: make sure you test your tool on a sample of known human generated text before you use it anywhere that a false positive could be problematic. The Authors Guild test was relatively simple: it involved ten articles and five tools. The fact that such a simple test revealed such large differences between tools should make anyone who thinks they’ve cracked the problem of AI detection think twice.