Is Moby-Dick AI-generated? 3 of 7 AI detectors said yes
Moby-Dick came out in 1851 and Herman Melville wrote it, so it is human writing. Even so, TMR (95.8%), Fakespot (90.7%) and Fast-DetectGPT flagged its opening as AI-generated, 3 of the 7 detectors we tested. Desklib (0.1%), the OpenAI RoBERTa detector (0.1%), E5-small LoRA (22.2%) and MAGE called it human.
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 | AI-generated | 90.7% |
| MAGE | Human | raw score only |
| E5-small LoRA detector | Human | 22.2% |
| OpenAI RoBERTa detector | Human | 0.1% |
| TMR (Target Mining RoBERTa) | AI-generated | 95.8% |
The passage we tested
Call me Ishmael. Some years ago—never mind how long precisely—having little or no money in my purse, and nothing particular to interest me on shore, I thought I would sail about a little and see the watery part of the world. It is a way I have of driving off the spleen and regulating the circulation. Whenever I find myself...
Fakespot, TMR and Fast-DetectGPT against the rest
Fakespot, TMR and Fast-DetectGPT flagged the opening. In our main benchmark of 126 human texts, Fakespot flagged 34.9% and TMR 83.3%, so neither is a careful judge. Fast-DetectGPT flagged 15.9%, using a threshold we set on our calibration split, which gives it a slight advantage over the other six. What makes its verdict odd is that it flagged no human fiction at all in that benchmark, and Fakespot flagged only 12.5% of it. Moby-Dick still drew a flag from both.
Desklib and the OpenAI RoBERTa detector each gave it 0.1%. E5-small LoRA, which flagged half of all human texts we tested, gave it 22.2%, and MAGE passed it as well.
Why Melville’s opening trips some detectors
On the surface this passage should look as human as prose gets. It has words no model reaches for by default: hypos, spleen, Manhattoes, Corlears Hook, Coenties Slip, spiles. Its sentences swerve, from a joke about knocking people’s hats off to Cato falling on his sword.
The likely explanation is fame. “Call me Ishmael” is one of the best-known first lines in English, and the paragraphs after it are quoted and reprinted constantly. A language model trained on web text has probably seen them many times. Fast-DetectGPT works by asking such a model how predictable each word is, and a passage the model half remembers looks predictable even when its vocabulary is strange.
TMR’s flag has a plainer cause. The passage has three paragraphs, and TMR flagged every multi-paragraph human text in our main test.
A confident score can still be wrong
Fakespot’s 90.7% reads like a near-certain verdict, and it is wrong. A detector’s score measures how closely a text matches patterns the model learned, so a high number on its own says little about who wrote it. Anyone using a free detector on real writing should treat a high percentage the same way as a low one: as one model’s guess. Alice in Wonderland drew flags from the same three detectors plus E5-small LoRA, while six of the seven passed A Tale of Two Cities. All results are on the famous texts page, and the Fakespot page shows where that detector errs most.
Questions people ask
Could Moby-Dick have been written by AI?
No, it was published in 1851. Fakespot, TMR and Fast-DetectGPT flagged its opening, but those three flagged 34.9%, 83.3% and 15.9% of human texts in our main test.
Why is Fakespot so sure Moby-Dick is AI?
We cannot see inside the model. Its 90.7% is a high score, but a confident score is not evidence of AI, and Fakespot gave scores this high to several other books written long before AI, including Jane Eyre and Little Women.
Which AI detector got Moby-Dick right?
Four of them. Desklib gave it 0.1%, tied with the OpenAI RoBERTa detector for the lowest score, and Desklib is the detector with the fewest false positives among those that catch a real share of AI text.