Is The Art of War AI-generated? 4 of 7 AI detectors said yes
No. The passage is Lionel Giles's 1910 English translation of Sun Tzu, a text more than two thousand years old, yet 4 of the 7 AI detectors we tested flagged it. Desklib, which flagged only 6.3% of human texts in our main test, gave it a 59.2% AI probability. Fakespot, MAGE and the OpenAI RoBERTa detector passed it.
Tested
What each AI detector said
| Detector | Verdict | AI probability |
|---|---|---|
| Fast-DetectGPT (Qwen2.5-1.5B) | AI-generated | raw score only |
| Desklib AI text detector | AI-generated | 59.2% |
| Fakespot AI text detector | Human | 2.5% |
| MAGE | Human | raw score only |
| E5-small LoRA detector | AI-generated | 93.1% |
| OpenAI RoBERTa detector | Human | 0.0% |
| TMR (Target Mining RoBERTa) | AI-generated | 98.3% |
The passage we tested
The art of war is of vital importance to the State. 2. It is a matter of life and death, a road either to safety or to ruin. Hence it is a subject of inquiry which can on no account be neglected. 3. The art of war, then, is governed by five constant factors, to be taken into account in...
Four detectors against three
The verdicts split almost down the middle, and Desklib’s flag is the hardest to dismiss. It gave the passage a 59.2% AI probability, and it flagged only 4 of the 30 famous texts we ran; the other three were Genesis 1, the Declaration of Independence and the US Constitution.
The rest of the “AI” side is less telling. TMR (98.3%) flags almost any text with paragraph breaks. The E5-small LoRA detector gave it 93.1%, but it flagged half of our human test texts. Fast-DetectGPT flagged it too; its threshold was set on our calibration split, and it flagged 15.9% of human texts. On the other side, Fakespot gave the passage 2.5% even though it flagged 13 of the 30 famous texts, MAGE passed it, and the OpenAI RoBERTa detector scored it 0.0%.
Why Sun Tzu reads like a chatbot answer
The layout is the most visible feature. The chapter is a run of numbered verses. Verse four lists five factors with numbers in parentheses, and the verses after it define each factor in turn (“Heaven signifies night and day, cold and heat, times and seasons”). Numbered points followed by one-line definitions is a common shape for AI assistant answers, and detectors trained on their output have likely learned it.
The translation probably matters as well. Giles turned compressed classical Chinese into short, balanced English sentences such as “It is a matter of life and death, a road either to safety or to ruin.” The opening line, “The art of war is of vital importance to the State,” could pass for the first sentence of a generated essay. His bracketed notes quoting Chinese commentators sound more personal, and they may be why some detectors passed the passage. The Giles version is public domain and widely copied across the web, which would make it familiar to the language model behind Fast-DetectGPT.
Notes, manuals and study guides
Human writing that uses numbered points and definitions, such as revision notes, training material or a list of principles, shares the features that likely triggered these flags. A detector reading that kind of text is judging the format as much as the author. The Prince, another translation, drew four flags from a partly different set of detectors, and the famous texts overview lists all 30 results. Desklib’s own error rates are on the Desklib page.
Questions people ask
Why would an AI detector flag The Art of War?
The likeliest reasons are its numbered, list-like layout and the plain, even English of the translation, both of which resemble chatbot answers. The Giles translation is also free to reprint and widely copied online, so language models have probably seen it often.
Which detector was most confident The Art of War is AI?
TMR, at 98.3%, followed by the E5-small LoRA detector at 93.1%, its highest score for any of the 30 famous texts we tested. Both flag a large share of human writing, 83.3% and 50.0% of our human test texts.
Does a translated text get flagged more often?
It can, though not always. This translation and The Prince were each flagged by 4 of the 7 detectors, while Meditations, in a seventeenth-century translation, was flagged only by TMR.