AI writing is revealed more by structure than word choice
Researchers from the University of Maryland and compared 61,608 texts written by people and by five : Claude, GPT, Gemini, DeepSeek, and Kimi. They removed clichés, excessive decoration, and repeated explanations from the AI texts, then tested whether a could identify them from alone. The detection rate fell only from 95.5% to 93.9%, a drop of 1.6 percentage points.
AI restated the meaning or lesson of what it had just written 77% of the time, compared with 52% for people. It expressed emotion through body metaphors, such as a physical feeling in the chest, in 81% of cases versus 38% for people; human writing more often named the event and its real cost. People used specific titles, brands, sums, and dates about twice as often as AI and addressed the reader 28% of the time, compared with 7% for AI.
People were also more willing to leave endings unclear, while AI tended to wrap into a neat conclusion. Replacing a few supposedly AI-like words therefore does little to remove the strongest signs of AI writing.
Key points
- The comparison covered 61,608 texts and five AI model families.
- Heavy surface editing reduced detection from 95.5% to only 93.9%.
- AI more often repeated the lesson and used body metaphors for emotion.
- People used more real names, brands, sums, and dates and spoke directly to readers more often.
- Making AI drafts sound natural requires structural editing, not simple word swaps.