Is Huckleberry Finn AI-generated? 2 of 7 AI detectors said yes
No. Mark Twain wrote Adventures of Huckleberry Finn, first published in 1884. Its opening, told in Huck's rough vernacular, was flagged as AI-generated by 2 of 7 free AI detectors, TMR and Fast-DetectGPT. Every other classifier passed it, including Desklib at a 1.3% AI probability and the OpenAI RoBERTa detector at 0.0%.
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 | 1.3% |
| Fakespot AI text detector | Human | 44.4% |
| MAGE | Human | raw score only |
| E5-small LoRA detector | Human | 34.5% |
| OpenAI RoBERTa detector | Human | 0.0% |
| TMR (Target Mining RoBERTa) | AI-generated | 96.7% |
The passage we tested
You don’t know about me without you have read a book by the name of The Adventures of Tom Sawyer; but that ain’t no matter. That book was made by Mr. Mark Twain, and he told the truth, mainly. There was things which he stretched, but mainly he told the truth. That is nothing. I never seen anybody but lied...
How the detectors split on Huck’s voice
Two detectors called the opening of Huckleberry Finn AI-generated: TMR, at a 96.7% AI probability, and Fast-DetectGPT. Everything else passed it. Desklib gave it 1.3%, E5-small LoRA 34.5% and Fakespot 44.4%, while the OpenAI RoBERTa detector’s score rounds to 0.0%. MAGE also called it human.
TMR’s flag carries little weight. It flagged 83.3% of the human texts in our main test and every human fiction passage, mostly in response to paragraph breaks. Fast-DetectGPT is a different case. Its threshold was tuned on our calibration split, and it flagged 15.9% of human texts overall but none of the 24 fiction passages, so a flag on a novel is rare for it.
Why Twain’s vernacular passes the classifiers
Huck narrates in the voice of a barely schooled Missouri boy: “that ain’t no matter”, “I never seen anybody but lied one time or another”, “allowed she would sivilize me”. The classifiers learned what AI text looks like from examples of model output, and that output is mostly grammatical, standard English. Double negatives, invented spellings and run-on storytelling are probably far outside what they learned to call AI, which would explain the low scores.
Fast-DetectGPT instead asks a language model how expected each word is. The opening of Huckleberry Finn is one of the most reproduced passages in American literature, and the model has likely read it many times. A model that already knows what comes next may find even “sivilize” predictable. This is our best explanation for the flag, and it matches the pattern for other well-known openings such as Pride and Prejudice.
What a distinctive voice protects you from
A strong personal style seems to help with the trained classifiers, but it gave no protection against a detector that may simply recognize the text. For original writing that is probably less of a worry, since no model has read your essay before. For writing that quotes or closely follows famous passages, a flag from Fast-DetectGPT may reflect the quotation more than the writer. The Fast-DetectGPT results explain how it works, and the famous texts page lists all 30 results.
Questions people ask
Why did Fast-DetectGPT flag Huckleberry Finn?
We cannot say for certain. Fast-DetectGPT flags text that a language model finds predictable, and the likeliest explanation is that the model has seen this famous opening many times. It flagged none of the 24 human fiction passages in our main test.
Does nonstandard grammar make a text look human to AI detectors?
It may help with the trained classifiers. Huck's double negatives and misspellings are unlike the polished English that AI tools produce, and Desklib, Fakespot, E5-small LoRA, MAGE and the OpenAI RoBERTa detector all passed this passage.