Is A Christmas Carol AI-generated? 2 of 7 AI detectors said yes
No, Charles Dickens published A Christmas Carol in 1843, and only 2 of the 7 AI detectors we tested flagged its opening, TMR and Fast-DetectGPT. Desklib gave it a 0.1% AI probability and Fakespot 1.8%, and MAGE, E5-small LoRA and the OpenAI RoBERTa detector passed it too. TMR flags 83.3% of human writing, so its flag means little.
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 | 1.8% |
| MAGE | Human | raw score only |
| E5-small LoRA detector | Human | 41.0% |
| OpenAI RoBERTa detector | Human | 0.0% |
| TMR (Target Mining RoBERTa) | AI-generated | 98.0% |
The passage we tested
MARLEY was dead: to begin with. There is no doubt whatever about that. The register of his burial was signed by the clergyman, the clerk, the undertaker, and the chief mourner. Scrooge signed it: and Scrooge's name was good upon 'Change, for anything he chose to put his hand to. Old Marley was as dead as a door-nail. Mind! I...
Style detectors and a predictability detector disagree
The split on this passage follows how the detectors work. The classifiers learned what AI text looks like from examples, and every one of them except TMR passed it: Desklib at 0.1%, Fakespot at 1.8%, E5-small LoRA at 41.0%, the OpenAI RoBERTa detector at 0.0%, and MAGE. TMR gave it 98.0%, and it flags almost every text that has paragraph breaks.
Fast-DetectGPT works differently. It has no training on AI examples and instead measures how predictable the wording is to a language model. It flagged the passage, even though it flagged none of the 24 human fiction passages in our main test. Its threshold was chosen on our calibration split, and overall it flagged 15.9% of human texts.
A joking narrator in a famous opening
The style gives the classifiers little to object to, because the voice is idiosyncratic and joking, far from the even tone of most AI prose. Dickens’s narrator interrupts himself (“Mind! I don’t mean to say that I know, of my own knowledge, what there is particularly dead about a door-nail”), asks and answers his own questions, and stacks a comic run of legal titles: sole executor, sole administrator, sole assign, sole residuary legatee. Then he wanders off to Hamlet’s father and Saint Paul’s Churchyard.
What the passage does have is fame. “Marley was dead: to begin with” opens one of the most adapted stories in English, a fixture of every Christmas season. A language model has probably seen these paragraphs many times, so it can predict them well, and Fast-DetectGPT reads high predictability as a sign of a machine. The same mechanism would explain why the method flagged famous passages elsewhere in our set while passing ordinary human fiction.
Quoting or retelling a classic
A school essay that quotes a well-known passage at length, or a retelling that keeps its famous lines, could be flagged by Fast-DetectGPT even when every style-based classifier passes it. Little Women is another famous Christmas opening it flagged. The Fast-DetectGPT page has its full results, and the famous texts overview covers all 30 passages.
Questions people ask
Why did Fast-DetectGPT flag A Christmas Carol?
Familiarity, most likely. Fast-DetectGPT asks a language model how predictable each word is, and an opening this widely read and quoted is predictable to a model that has seen it many times. In our main test it flagged no human fiction passages.
Which detector is most reliable on a text like this?
None is proof, but Desklib had the fewest false positives among detectors that catch a real share of AI text, 6.3% of human texts in our main test. It gave this passage a 0.1% AI probability.