Is Civil Disobedience AI-generated? 3 of 7 AI detectors flagged Thoreau
Thoreau's essay appeared in 1849 and is human writing, but MAGE, Fast-DetectGPT and TMR (96.2%) flagged its opening as AI-generated, 3 of the 7 detectors we tested. Desklib (0.1%), Fakespot (4.3%) and the OpenAI RoBERTa detector (0.0%) passed it easily, and E5-small LoRA passed it narrowly at 46.1%.
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 | Human | 0.1% |
| Fakespot AI text detector | Human | 4.3% |
| MAGE | AI-generated | raw score only |
| E5-small LoRA detector | Human | 46.1% |
| OpenAI RoBERTa detector | Human | 0.0% |
| TMR (Target Mining RoBERTa) | AI-generated | 96.2% |
The passage we tested
I heartily accept the motto,—“That government is best which governs least;” and I should like to see it acted up to more rapidly and systematically. Carried out, it finally amounts to this, which also I believe—“That government is best which governs not at all;” and when men are prepared for it, that will be the kind of government which they...
Three flags, four passes
MAGE, Fast-DetectGPT and TMR said AI. TMR’s verdict counts for little, since it flagged 83.3% of the human texts in our main benchmark and every multi-paragraph one among them. MAGE flagged 19.0% of human texts and Fast-DetectGPT 15.9%, the latter with a threshold tuned on our calibration split, a small advantage over detectors that use their authors’ default.
The four that passed it were Desklib at 0.1%, the OpenAI RoBERTa detector at 0.0%, E5-small LoRA at 46.1% and Fakespot at 4.3%. Fakespot’s low score is unusual for a detector that flagged 34.9% of human writing in our test and gave two other famous essays, the Communist Manifesto and A Modest Proposal, 93.8% and 98.7%.
Same author, different result
Thoreau’s Walden makes a useful comparison. Only TMR flagged its opening. Both passages are his, so the difference most likely comes from what each passage is doing.
Walden opens with facts about one man: a house he built himself, a mile from any neighbor, on the shore of Walden Pond. Civil Disobedience opens with general claims about government. “Government is at best but an expedient.” “The standing army is only an arm of the standing government.” Sentences like these state a principle and then turn it over, and that explanatory register is where MAGE and Fast-DetectGPT have trouble with human writing. In our main test, MAGE flagged 40.0% of human encyclopedia articles and Fast-DetectGPT 40.0% of human academic abstracts.
The opening motto may also be too familiar. “That government is best which governs least” is quoted constantly, and Fast-DetectGPT judges a text by how well a language model can predict it. A line the model has seen many times is easy to predict.
Genre more than author
Two passages are a small sample, but they fit our main test, where MAGE and Fast-DetectGPT flagged expository human writing much more often than fiction (MAGE flagged 4.2% of human fiction, Fast-DetectGPT none). If a student’s analytical essay is flagged while their personal essay passes, that may be the genre talking. The famous texts results list every verdict for 30 passages, and the Fast-DetectGPT page breaks down where that detector goes wrong.
Questions people ask
Is the AI verdict on Civil Disobedience reliable?
No. Thoreau published the essay in 1849, and the three detectors that flagged it flagged between 15.9% (Fast-DetectGPT) and 83.3% (TMR) of human texts in our main test.
Did the detectors treat Walden the same way?
No. Walden, which Thoreau published in 1854, was flagged by only 1 of 7 detectors, TMR. The same writer got different verdicts on different texts.
Which detector came closest to changing the result?
E5-small LoRA, at 46.1%, just under its 50% threshold. Had it gone the other way, the passage would have been flagged by a majority of the seven.