Is Dracula AI-generated? 3 of 7 AI detectors said yes
Dracula is Bram Stoker's novel of 1897, so no. Its opening journal entry was still flagged as AI-generated by 3 of 7 free AI detectors, TMR, E5-small LoRA and Fast-DetectGPT. Desklib and Fakespot, the two highest-ranked classifiers on our leaderboard, disagreed strongly, at 0.1% and 1.1% AI probability. MAGE and the OpenAI RoBERTa detector also passed it.
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 | 1.1% |
| MAGE | Human | raw score only |
| E5-small LoRA detector | AI-generated | 57.4% |
| OpenAI RoBERTa detector | Human | 17.0% |
| TMR (Target Mining RoBERTa) | AI-generated | 96.9% |
The passage we tested
Left Munich at 8:35 P. M., on 1st May, arriving at Vienna early next morning; should have arrived at 6:46, but train was an hour late. Buda-Pesth seems a wonderful place, from the glimpse which I got of it from the train and the little I could walk through the streets. I feared to go very far from the station,...
Three flags, and what each one is worth
Dracula opens with Jonathan Harker’s travel journal, and three detectors called that journal AI-generated: TMR at 96.9%, E5-small LoRA at 57.4%, and Fast-DetectGPT. The verdicts look different once you know each detector’s record. TMR flagged every one of the 24 human fiction passages in our main test, so a TMR flag on a novel tells you nothing. E5-small LoRA flagged 50.0% of all human texts, and its 57.4% here is barely past its 50% cut-off.
Fast-DetectGPT’s flag is harder to explain. Its threshold was tuned on our calibration set, and it flagged 15.9% of human texts overall but none of the 24 human fiction passages. Dracula and a few other famous novel openings are exceptions to that clean record.
On the other side, Desklib gave the passage a 0.1% AI probability and Fakespot 1.1%. MAGE passed it too, and the OpenAI RoBERTa detector put it at 17.0%.
What in a Victorian diary could look like AI
Harker’s journal reads nothing like chatbot prose. It starts with a train time and no subject (“Left Munich at 8:35 P. M.”), notes that the train was an hour late, and breaks off for a private reminder to get a paprika chicken recipe for Mina. The classifiers trained on modern AI output seem to read those clipped, personal details as human.
Fast-DetectGPT is another matter. It asks how predictable each word is to a language model, and the likeliest explanation for its flag is familiarity. Dracula is out of copyright and copied all over the web, so the model behind the detector may have seen this opening many times. The plain travel narration, with its stations, hotels and dinners in order, may also be easy to predict.
Reading a split verdict
When detectors disagree, the useful question is which ones flagged the text. Here two of the flags came from tools that flag a lot of human writing, TMR and E5-small LoRA, and the third from a detector that may know the passage by heart. Desklib and Fakespot, which rarely flag fiction, were close to certain it is human. The same pattern appears for Heart of Darkness, which only TMR flagged. The E5-small LoRA results and the famous texts overview give the wider picture.
Questions people ask
Why would an AI detector flag Dracula?
The likeliest reasons differ by detector. TMR flags nearly every text with several paragraphs. E5-small LoRA flagged half of the human texts in our main test. Fast-DetectGPT may find the opening predictable because the novel is so widely copied online.
Which detectors said Dracula was written by a human?
Desklib (0.1% AI probability), Fakespot (1.1%), the OpenAI RoBERTa detector (17.0%) and MAGE.
How confident was E5-small LoRA that Dracula is AI?
Not very. It gave the passage a 57.4% AI probability, only a little over its 50% threshold, and it flagged 50.0% of the human texts in our main test.