Professional translators can sometimes tell when AI wrote an Italian short story.
In an in‑person experiment, 69 translators read three anonymized stories—two produced by ChatGPT‑4o and one by a human author. They rated how likely each piece was AI‑generated and explained their reasoning. Overall, the group’s accuracy hovered around chance, but 16.2% consistently singled out the synthetic texts. An almost equal share flipped the verdict, often citing subjective impressions like smoothness rather than concrete cues.
The result matters because it shows that even language experts are not uniformly equipped to flag AI output. The few who succeeded relied on low burstiness, narrative contradictions, and odd borrowings from English, suggesting that specific linguistic fingerprints exist. Meanwhile, reliance on grammar or tone proved misleading, raising doubts about using translators as a line of defense against undisclosed AI text.
As AI writers improve, the industry may need dedicated detection tools rather than hoping skilled humans will catch every synthetic passage.