Is Leviathan AI-generated? 2 of 7 AI detectors flagged Hobbes
Leviathan is Thomas Hobbes's work from 1651 and is not AI-generated. In our test, 2 of 7 AI detectors flagged its introduction anyway, TMR at 96.5% and Fakespot at 55.3%, just over its threshold. Desklib (0.1%), the OpenAI RoBERTa detector (0.0%), E5-small LoRA (18.4%), MAGE and Fast-DetectGPT all read it as human.
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 | AI-generated | 55.3% |
| MAGE | Human | raw score only |
| E5-small LoRA detector | Human | 18.4% |
| OpenAI RoBERTa detector | Human | 0.0% |
| TMR (Target Mining RoBERTa) | AI-generated | 96.5% |
The passage we tested
Nature (the art whereby God hath made and governes the world) is by the art of man, as in many other things, so in this also imitated, that it can make an Artificial Animal. For seeing life is but a motion of Limbs, the begining whereof is in some principall part within; why may we not say, that all Automata...
TMR and a marginal Fakespot flag
Only TMR and Fakespot said AI. TMR flagged 28 of the 30 famous texts we tested and 83.3% of the human texts in our main benchmark, so its verdict here adds almost nothing. Fakespot’s 55.3% is a marginal call from a detector that flagged 34.9% of human writing.
Desklib, with the fewest false positives (6.3%) among detectors that catch a real share of AI text, gave it 0.1%. More telling is E5-small LoRA at 18.4%, because it flagged 50.0% of human texts in our benchmark and still let this one through. MAGE, Fast-DetectGPT and the OpenAI RoBERTa detector passed it as well.
Why most detectors read Hobbes as human
The likely reason is distance from modern text. Hobbes writes “governes”, “begining”, “principall”, “Joynts”, “Wheeles” and “Soveraignty”, capitalizes nouns in the middle of sentences (Heart, Spring, Nerves, Strings) and drops in Latin: CIVITAS, Salus Populi. The classifiers here were trained on recent text, human and machine, and no current model spells like this. For Fast-DetectGPT, which flags text that a language model finds predictable, odd spellings make each word harder to guess, which pushes toward a human verdict.
Unlike Hamlet or the opening of Moby-Dick, this introduction is not a passage most people could quote. “Nature (the art whereby God hath made and governes the world)” is famous among students of political philosophy, but it is likely to appear in training data far less often, so memorization is less likely to push the scores toward AI.
The two flags, and what they are worth
Fakespot’s near-threshold flag may come from the passage’s structure. Hobbes builds a long list of matches between the state and a body (the Soveraignty is the Soul, the Magistrates the Joynts, Reward and Punishment the Nerves), then starts an outline with “First” and “Secondly”. Lists and enumerations are common in AI answers. TMR has no such subtlety: the passage has four paragraphs, and TMR flagged every human text with more than one paragraph in our main benchmark, as the TMR page explains.
Nobody writing today is protected by Hobbes’s spelling, so the useful lesson for real writing is the Fakespot score. A verdict a few points above a threshold could easily have gone the other way, and deserves no more weight than a pass. A Modest Proposal was flagged by the same two detectors, and the famous texts page lists every result.
Questions people ask
Should I trust an AI detector that flags Leviathan?
No. Hobbes published it in 1651, and the two detectors that flagged it, TMR and Fakespot, flagged 83.3% and 34.9% of human texts in our main test.
Why did most AI detectors pass Leviathan?
Most likely because it does not look like anything a modern model writes. The spelling (governes, Joynts, artificiall, Soveraignty), the capitalized nouns and the Latin phrases are far from the modern English these detectors were trained on.
How close was Fakespot to passing it?
Close. Fakespot gave it 55.3% against a threshold of 50%, so a small change in the passage could have flipped its verdict.