Do AI detectors flag famous texts? We tested 30 classics
Yes, often. We ran 30 famous texts, from the Declaration of Independence to Moby-Dick, through seven free AI detectors. All 30 were flagged as AI by at least one detector, and 8 of them by a majority. Genesis 1 from the King James Bible was flagged by all seven detectors, and the US Constitution by six.
Tested
Every famous text we tested
| Text | Written | Flagged by | Detectors that said AI |
|---|---|---|---|
| Genesis 1 (King James Version) | 1611 | 7 of 7 | OpenAI RoBERTa detector, Fakespot AI text detector, Desklib AI text detector, MAGE, TMR (Target Mining RoBERTa), E5-small LoRA detector, Fast-DetectGPT (Qwen2.5-1.5B) |
| The United States Constitution | 1787 | 6 of 7 | Fakespot AI text detector, Desklib AI text detector, MAGE, TMR (Target Mining RoBERTa), E5-small LoRA detector, Fast-DetectGPT (Qwen2.5-1.5B) |
| Alice's Adventures in Wonderland (opening) | 1865 | 4 of 7 | Fakespot AI text detector, TMR (Target Mining RoBERTa), E5-small LoRA detector, Fast-DetectGPT (Qwen2.5-1.5B) |
| Pride and Prejudice (opening) | 1813 | 4 of 7 | Fakespot AI text detector, TMR (Target Mining RoBERTa), E5-small LoRA detector, Fast-DetectGPT (Qwen2.5-1.5B) |
| The Art of War (opening chapter) | 1910 | 4 of 7 | Desklib AI text detector, TMR (Target Mining RoBERTa), E5-small LoRA detector, Fast-DetectGPT (Qwen2.5-1.5B) |
| The Communist Manifesto (opening) | 1848 | 4 of 7 | Fakespot AI text detector, MAGE, TMR (Target Mining RoBERTa), Fast-DetectGPT (Qwen2.5-1.5B) |
| The Declaration of Independence | 1776 | 4 of 7 | Desklib AI text detector, MAGE, TMR (Target Mining RoBERTa), Fast-DetectGPT (Qwen2.5-1.5B) |
| The Prince (chapter one) | 1908 | 4 of 7 | Fakespot AI text detector, MAGE, TMR (Target Mining RoBERTa), Fast-DetectGPT (Qwen2.5-1.5B) |
| Civil Disobedience (opening) | 1849 | 3 of 7 | MAGE, TMR (Target Mining RoBERTa), Fast-DetectGPT (Qwen2.5-1.5B) |
| Dracula (opening) | 1897 | 3 of 7 | TMR (Target Mining RoBERTa), E5-small LoRA detector, Fast-DetectGPT (Qwen2.5-1.5B) |
| Little Women (opening) | 1868 | 3 of 7 | Fakespot AI text detector, TMR (Target Mining RoBERTa), Fast-DetectGPT (Qwen2.5-1.5B) |
| Moby-Dick (opening) | 1851 | 3 of 7 | Fakespot AI text detector, TMR (Target Mining RoBERTa), Fast-DetectGPT (Qwen2.5-1.5B) |
| The Wealth of Nations (opening) | 1776 | 3 of 7 | Fakespot AI text detector, TMR (Target Mining RoBERTa), Fast-DetectGPT (Qwen2.5-1.5B) |
| A Christmas Carol (opening) | 1843 | 2 of 7 | TMR (Target Mining RoBERTa), Fast-DetectGPT (Qwen2.5-1.5B) |
| A Modest Proposal | 1729 | 2 of 7 | Fakespot AI text detector, TMR (Target Mining RoBERTa) |
| Adventures of Huckleberry Finn (opening) | 1884 | 2 of 7 | TMR (Target Mining RoBERTa), Fast-DetectGPT (Qwen2.5-1.5B) |
| Jane Eyre (opening) | 1847 | 2 of 7 | Fakespot AI text detector, TMR (Target Mining RoBERTa) |
| Leviathan (introduction) | 1651 | 2 of 7 | Fakespot AI text detector, TMR (Target Mining RoBERTa) |
| On the Origin of Species (introduction) | 1859 | 2 of 7 | Fakespot AI text detector, TMR (Target Mining RoBERTa) |
| A Scandal in Bohemia (opening) | 1891 | 1 of 7 | TMR (Target Mining RoBERTa) |
| A Tale of Two Cities (opening) | 1859 | 1 of 7 | TMR (Target Mining RoBERTa) |
| Common Sense (introduction) | 1776 | 1 of 7 | TMR (Target Mining RoBERTa) |
| Frankenstein (opening letter) | 1818 | 1 of 7 | TMR (Target Mining RoBERTa) |
| Hamlet's "To be, or not to be" soliloquy | 1603 | 1 of 7 | Fast-DetectGPT (Qwen2.5-1.5B) |
| Heart of Darkness (opening) | 1899 | 1 of 7 | TMR (Target Mining RoBERTa) |
| Meditations (book one) | 1634 | 1 of 7 | TMR (Target Mining RoBERTa) |
| The Federalist No. 10 | 1787 | 1 of 7 | Fast-DetectGPT (Qwen2.5-1.5B) |
| The Gettysburg Address | 1863 | 1 of 7 | TMR (Target Mining RoBERTa) |
| The Great Gatsby (opening) | 1925 | 1 of 7 | TMR (Target Mining RoBERTa) |
| Walden (opening) | 1854 | 1 of 7 | TMR (Target Mining RoBERTa) |
Numbers behind this chart
| Item | Rate | Texts | 95% interval |
|---|---|---|---|
| Fast-DetectGPT (Qwen2.5-1.5B) | 56.7% | 17 of 30 | 39.2 to 72.6% |
| Desklib AI text detector | 13.3% | 4 of 30 | 5.3 to 29.7% |
| Fakespot AI text detector | 43.3% | 13 of 30 | 27.4 to 60.8% |
| MAGE | 20.0% | 6 of 30 | 9.5 to 37.3% |
| E5-small LoRA detector | 20.0% | 6 of 30 | 9.5 to 37.3% |
| OpenAI RoBERTa detector | 3.3% | 1 of 30 | 0.6 to 16.7% |
| TMR (Target Mining RoBERTa) | 93.3% | 28 of 30 | 78.7 to 98.2% |
What we tested
We took 30 well-known public-domain texts from Project Gutenberg: founding documents, scripture, speeches, essays, philosophy and the openings of famous novels. From each we cut a passage of about the length used in our main test, starting at its best-known line, and ran it through the same seven detectors, with the same thresholds, that we use for the AI detector leaderboard.
The texts detectors were surest about
Genesis 1 was the only text that every detector flagged. Even the OpenAI RoBERTa detector, which flagged nothing else among the 30, called it AI-generated. The US Constitution followed, flagged by six detectors. Both are formal, repetitive texts, and both are among the most copied documents in English, so the language models behind many detectors have seen them many times.
That second point matters most for Fast-DetectGPT, which flags text that a language model finds predictable. It flagged 17 of the 30 texts (56.7%). The famous “To be, or not to be” soliloquy from Hamlet passed every classifier but drew one of Fast-DetectGPT’s strongest scores of the whole set. A model that knows a passage by heart finds it predictable, and the detector reads that as machine writing.
Which detectors flagged the classics
TMR flagged 28 of the 30 texts (93.3%), which mostly reflects its reaction to paragraph breaks. Fakespot flagged 13 (43.3%), MAGE and E5-small LoRA 6 each (20.0%), and Desklib 4 (13.3%). Five of the famous novel openings were flagged by TMR alone, while the Bible and the founding documents drew the most flags.
Memorised text looks machine-written
A detector cannot tell the difference between text a machine wrote and text a machine has memorised. Formal, heavily quoted or rule-bound writing looks the same to it. If a student quotes scripture, a legal clause or a well-known passage, a flag is likely to follow. Our false positives page shows the same problem with everyday human writing.
Questions people ask
Why do AI detectors say the Bible or the Constitution is AI-generated?
The likeliest reason is that language models have read these texts many times, so their wording looks highly predictable, and predictable wording is what many detectors treat as a sign of AI. Formal, repetitive phrasing adds to it.
Which AI detector flagged the fewest famous texts?
The OpenAI RoBERTa detector flagged 1 of the 30, but it also misses almost all modern AI text. Among the detectors that catch a real share of AI writing, Desklib flagged the fewest, 4 of 30 (13.3%).
Does this mean AI detectors are useless?
It means a single flag proves nothing. These texts were written between 1611 and 1925, long before AI, and detectors still called many of them AI-generated.