When users see a score like "99% AI," it feels like a verdict. The web is flooded with detectors promising 99%+ accuracy. Detect.ai claims 99.9% accuracy, while another well-known tool boasts 99.98%. But these numbers hide as much as they reveal.
Most detection algorithms work by scanning for patterns—sentence structure, phrase frequency, and a specific blandness that LLMs like GPT-5 tend to produce. The problem: human writing, especially when edited for clarity, can trigger those same patterns. In testing, we found nearly two-thirds of flagged essays came from students who had simply followed academic style guides too closely.
Here's what surprised us. The probability score isn't a DNA test for authorship. It's a statistical guess based on surface features. When a professor asks, "Can you prove this student used ChatGPT?" the honest answer is no. The score means, "This text looks a lot like known AI outputs"—not "This was definitely written by an AI." I've seen administrators treat high scores as hard evidence. Some later reversed course when shown a draft history.
Our logs revealed the false positive rate: about 18% of essays flagged above 90% were, after review, fully human-written.
The other side of the coin is evasion. Newer models like Claude 4.6 and Gemini 2.5 generate text that slips through older detectors. We had to scramble in early 2026 when a spike in undetected AI essays showed up in our logs—we caught the trend before the market did. While some platforms still tout 99% accuracy, real-world accuracy varies by model and language. In English, top detectors might hit 95-99% on classic GPT-4 material. But accuracy drops to 80% or lower on paraphrased or multilingual content. The number nobody shows you: if you paste a French essay or a creatively reworded English paragraph, the odds of a clean pass are much higher than published accuracy rates suggest.
Top comments (0)