What are the limitations of GPTZero's AI classifier?

The nature of AI-generated content is changing constantly. These results should not be used to punish students. While we build more robust models for GPTZero, we recommend that educators take these results as one of many pieces in a holistic assessment of student work. There are always edge cases, both where AI is classified as human and where human is classified as AI. We recommend approaches that give students the opportunity to demonstrate their understanding in a controlled environment.

The accuracy of our model increases as more text is submitted. Accuracy on the document-level classification is greater than accuracy on the paragraph level, which is greater than accuracy on the sentence level.

Accuracy also increases for text similar to our dataset. We train on a highly diverse set of human and AI-generated text, and the majority of that dataset is English prose written by adults.

GPTZero can flag AI-generated text that was rewritten after it was generated, either by a paraphrasing tool or by hand. In scan results this appears as AI Paraphrased: AI-generated text that was rewritten to disguise its origin. Heavily edited text is still harder to classify than text taken straight from a model, so a result here is a signal to look at the writing process, not a final verdict.

The classifier can sometimes flag other machine-generated or highly procedural text as AI-generated. Use it on the more descriptive portions of a text.