🕵️ Fool the detector
AI detectors catch machine text based on one feature: it is predictable. The model almost always chooses the most expected word. Write a sentence and the AI will try to guess your last word (top 3). Will hit the top three → “predictably, the detector would have marked it.” It won’t hit → “You write like a person.”
How it works: perplexy and detectors
At each step, the language model evaluates how likely each next word is and usually takes the most expected one. Because of this, her text is “smooth” and predictable - scientifically, with low perplexity (a measure of the model’s surprise).
AI detectors measure exactly this: they run text through a model and see how “surprised” it is. Smooth, predictable text → low perplexity → “AI-like.” A living person more often chooses unexpected, specific words → the model is surprised → “looks like a person.”
Therefore, detectors often make mistakes: a dry person looks like AI, but edited AI text passes. And therefore, the best way to keep the text from being scorched and simply read vividly is to add specifics, rhythm and unexpected details. Read more in our guide about text checking using AI. You can practice in a free lesson.
Start for free