Researchers showed that two different kinds of detectors can identify whether an essay was written by a human or produced by a large language model. Using a dataset of 15‑year‑old German students’ English essays that were written with or without a gpt‑4o‑mini chatbot, they traced every character in each essay back to where it came from—keystroke logs, clipboard history and chat transcripts. From these provenance records, they created labels for full paragraphs and individual sentences that were human, AI, mixed or written after using a chatbot but without traceable AI material.

The study compared a feature‑based model that uses lexical complexity to an embedding‑based classifier that uses frozen Llama‑3.1‑8B embeddings. On a strict AI versus human test set, the feature‑based XGBoost model achieved a macro‑F1 of 0.95 for whole texts and 0.88 for single sentences. The embedding‑based RFCN scanner reached 0.96 macro‑F1 for texts and 0.95 for sentences. Even when grey‑zone items—texts that include a mix of AI and human writing—were added, the performance dropped only modestly.

However, the authors caution that the gold labels for the study depended on chosen thresholds such as a 90 % AI content cut‑off and similarity scores of 0.85–0.90 for retyped text. The provenance data cannot detect content that students paraphrased heavily or used only for ideas, so the detectors may misclassify some mixed or AI‑influenced writing.

These results suggest that AI‑detecting tools can separate purely AI and purely human writing with high accuracy under controlled conditions, but they are less reliable on short sentences, on heavily paraphrased text, or when mixed human‑AI content is present.