Is Meditations by Marcus Aurelius AI-generated? Only 1 of 7 AI detectors said yes
No. Marcus Aurelius wrote the Meditations in Greek in the second century, and the passage we tested comes from Meric Casaubon's English translation, first published in 1634. Only 1 of the 7 AI detectors we tested flagged it, TMR, at a 96.0% AI probability, and TMR flagged 83.3% of human texts in our main test. Desklib gave it 0.1%.
Tested
What each AI detector said
| Detector | Verdict | AI probability |
|---|---|---|
| Fast-DetectGPT (Qwen2.5-1.5B) | Human | raw score only |
| Desklib AI text detector | Human | 0.1% |
| Fakespot AI text detector | Human | 14.4% |
| MAGE | Human | raw score only |
| E5-small LoRA detector | Human | 29.3% |
| OpenAI RoBERTa detector | Human | 0.1% |
| TMR (Target Mining RoBERTa) | AI-generated | 96.0% |
The passage we tested
Of my grandfather Verus I have learned to be gentle and meek, and to refrain from all anger and passion. From the fame and memory of him that begot me I have learned both shamefastness and manlike behaviour. Of my mother I have learned to be religious, and bountiful; and to forbear, not only to do, but to intend any...
One flag, from TMR
Apart from TMR, which gave the passage 96.0%, every detector called it human. Desklib scored it 0.1%, the OpenAI RoBERTa detector 0.1%, Fakespot 14.4% and the E5-small LoRA detector 29.3%, and MAGE and Fast-DetectGPT did not flag it either. TMR’s flag counts for little, since it flags almost any text with paragraph breaks, and this one is split into numbered sections.
A list that did not trip the detectors
Book one looks like an easy text for a detector to mistake for AI. It is a catalogue of debts, and each entry is built on the same frame: “Of my grandfather Verus I have learned…”, “Of my mother I have learned…”, “Of Diognetus, not to busy myself about vain things”. Repeated sentence openings and enumerated points are often listed as signs of chatbot writing, and a translation can flatten an author’s voice into something more uniform.
Two other translated texts we tested, The Art of War and The Prince, were each flagged by four detectors. The likeliest reason this one passed is the age of its English. Casaubon wrote in the seventeenth century, and his vocabulary is far from anything a modern chatbot produces: “him that begot me”, “shamefastness”, “prestidigitators”, “the coursers in the circus”, and Latin names such as the Parmularii and the Secutores. A language model has little reason to predict those words, so Fast-DetectGPT probably found the text surprising, and classifiers trained on modern AI output had little in it to recognize. The Giles and Marriott translations, from the early twentieth century, read much more like plain modern English.
Plain lists written today
The comparison suggests these detectors react less to structure than to wording that sounds like current, plain English. That is bad news for people who write clearly and in lists. Study notes, reflective journals and step-by-step instructions written today may be flagged where a seventeenth-century list of virtues was not. The famous texts overview shows all 30 results, and the TMR page explains why its flags carry so little weight.
Questions people ask
Why did TMR say Meditations is AI-generated?
Probably because of its paragraph breaks. TMR flagged all 101 multi-paragraph human texts in our main test, and only 25.7% of them once we removed the line breaks.
Is a list-like text more likely to be flagged as AI?
Not on its own, judging by this passage. Book one of the Meditations is a list of people and what the author learned from each, yet only TMR flagged it. The archaic wording likely outweighed the list structure.
Would a modern translation of the Meditations be flagged?
We did not test one. A translation in plain modern English would lose the old vocabulary that probably helped this one pass, so the result could differ.