Every few months, a new AI model launches and someone declares detection dead. Then a few months later, detectors catch up. Then a new model launches. This cycle has been going on since 2023, and it's the most consequential cat-and-mouse game in technology today.
Here's how it actually works.
Early AI text was easy to catch. It was overly formal, used predictable transitions, and had almost no variation in sentence length. Detectors trained on these patterns achieved 95%+ accuracy. The cat was winning.
AI companies began tuning their models to sound more human. They added instructions to vary sentence length, use contractions, and inject "personality." Some models were specifically trained to evade detection. For a while, this worked — false negatives spiked, and detection accuracy dropped to 80% on the newest models.
Detection tools responded by going deeper. Instead of looking at surface patterns, modern detectors analyze statistical properties of language that are extremely hard to fake: token-level perplexity distributions, semantic coherence patterns, and the subtle ways AI models structure arguments. Our detector also uses paragraph-level analysis, which catches hybrid documents where AI text is mixed with human writing.
As of October 2026, the best detectors — including ours — achieve 92-97% accuracy on fully AI-generated text. The gap has narrowed on heavily edited text, but detection is still winning. Here's why: AI models have a fundamental constraint. They generate text token by token, always choosing the most likely next word. This creates statistical signatures that are nearly impossible to eliminate without fundamentally changing how the model works.
Right now, detectors. But the gap is narrow, and it's closing. The real question isn't who's ahead today — it's whether detection can keep adapting as fast as the models evolve.
No detector is perfect. False positives happen, especially with non-native English writers and very short texts. The best approach is never to rely on a single score. Use detection as one signal alongside context, conversation, and common sense. Our paragraph-level breakdown exists precisely for this reason: it gives you transparency into which parts of a document are flagged, so you can make an informed judgment instead of trusting a black box.
The cat-and-mouse game isn't ending. But for now, the cat is keeping pace.