Is Frankenstein AI-generated? Only 1 of 7 AI detectors said yes
No. When we ran the opening letter of Mary Shelley's Frankenstein (1818) through 7 AI detectors, only TMR called it AI-generated, with a 97.2% probability, and TMR flagged 83.3% of human texts in our main test. Desklib gave the letter a 2.0% AI probability and Fakespot 13.6%, and MAGE and Fast-DetectGPT passed it.
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 | 2.0% |
| Fakespot AI text detector | Human | 13.6% |
| MAGE | Human | raw score only |
| E5-small LoRA detector | Human | 27.8% |
| OpenAI RoBERTa detector | Human | 0.0% |
| TMR (Target Mining RoBERTa) | AI-generated | 97.2% |
The passage we tested
You will rejoice to hear that no disaster has accompanied the commencement of an enterprise which you have regarded with such evil forebodings. I arrived here yesterday, and my first task is to assure my dear sister of my welfare and increasing confidence in the success of my undertaking. I am already far north of London, and as I walk...
How the seven detectors read the letter
The excerpt is the start of the novel’s first letter, from the explorer Robert Walton in St. Petersburg to his sister Margaret in England. TMR scored it 97.2% and was the only detector to call it AI. Desklib gave it 2.0%, Fakespot 13.6%, the E5-small LoRA detector 27.8% and the OpenAI RoBERTa detector 0.0%, and MAGE and Fast-DetectGPT both passed it.
TMR’s flag is the weakest kind of evidence in our test. It flagged every multi-paragraph human text we gave it, and this letter has two paragraphs. Its overall false positive rate on human writing was 83.3%.
Why Shelley’s formal style passed when the Constitution’s did not
Shelley’s prose is as formal in its way as the US Constitution, which six detectors flagged, yet it drew a single flag. The difference is probably in the kind of formality. The Constitution repeats one legal construction clause after clause. Walton’s letter keeps changing shape: a long sentence about the “cold northern breeze” is followed by a direct question to his sister (“Do you understand this feeling?”), then by a run of Romantic vocabulary such as “inspirited”, “wafted”, “satiate my ardent curiosity” and “diffusing a perpetual splendour”. Rare words and varied sentences are likely harder for a language model to predict, and Fast-DetectGPT flags text that is easy to predict.
The likely reasons a pre-AI text gets flagged mostly miss this passage. It is a famous novel, but the opening letter is probably much less quoted than the scenes with Victor and his creature, so a language model has likely seen it less often. It is neither a translation nor a list.
Ornate prose and plain prose
In our small set of famous texts, ornate or old-fashioned vocabulary seems to make these detectors less likely to call a passage AI. The flip side is that plain, even prose may be at more risk, and that is the style many teachers ask students to write in. If a detector flags such writing, the false positives page shows how often the same detectors flag ordinary human text. A Scandal in Bohemia was also flagged by TMR alone, and you can compare all 30 texts on the famous texts page.
Questions people ask
Does an AI detector think Frankenstein was written by AI?
Almost none of the ones we tested. TMR flagged the opening letter at 97.2%, but it flags nearly every text with paragraph breaks. Desklib, Fakespot, E5-small LoRA, the OpenAI RoBERTa detector, MAGE and Fast-DetectGPT all passed it.
Why is Frankenstein flagged less than the US Constitution?
Both are formal, but in different ways. The Constitution repeats one legal sentence pattern, while Shelley's letter varies its sentences and uses rare words, which likely makes it harder for a language model to predict and less like the uniform style of AI text.