Is Alice in Wonderland AI-generated? 4 of 7 AI detectors said yes
No, Lewis Carroll wrote it, and it came out in 1865. Four of the seven AI detectors we tested flagged its opening anyway, TMR (98.3%), E5-small LoRA (81.7%), Fakespot (57.2%) and Fast-DetectGPT. Desklib gave it 1.2%, the OpenAI RoBERTa detector 0.0%, and MAGE passed it too.
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 | 1.2% |
| Fakespot AI text detector | AI-generated | 57.2% |
| MAGE | Human | raw score only |
| E5-small LoRA detector | AI-generated | 81.7% |
| OpenAI RoBERTa detector | Human | 0.0% |
| TMR (Target Mining RoBERTa) | AI-generated | 98.3% |
The passage we tested
Alice was beginning to get very tired of sitting by her sister on the bank, and of having nothing to do: once or twice she had peeped into the book her sister was reading, but it had no pictures or conversations in it, “and what is the use of a book,” thought Alice “without pictures or conversations?” So she was...
Who flagged the opening, and how reliable they are
The four detectors that called this passage AI-written are some of the least careful in our test. On the 126 human texts in our main benchmark, TMR flagged 83.3%, E5-small LoRA 50.0% and Fakespot 34.9%. Fakespot’s score for Alice, 57.2%, sits only a little above its 50% threshold.
Fast-DetectGPT is the odd one. It flagged 15.9% of human texts overall, with a threshold we tuned on our calibration split, which gives it a small edge over detectors that use their authors’ setting. It flagged none of the 24 human fiction passages in that benchmark, yet it flagged Alice.
The detectors that passed it include the two most cautious ones. Desklib, which flagged 6.3% of human writing, gave Alice 1.2%. The OpenAI RoBERTa detector, which flagged 4.8%, gave it 0.0%, though that means less, because it was built in 2019 for GPT-2 and caught only 9.5% of the AI texts we gave it.
What may have tripped each detector
Familiarity is the likely reason for Fast-DetectGPT. It asks a language model how predictable each word is, and a model has almost certainly read this opening many times, in reprints and quotations. Words the model can see coming look like its own output.
The prose may play a part for the classifiers. Carroll narrates in long, smooth sentences joined by “and” and “but”, with plain words: tired, sister, book, pictures, rabbit, watch. Fluent, evenly paced narration in plain words is also how many AI-written stories read, and a classifier that learned those surface habits can confuse the two.
TMR has a more mechanical cause. Our passage runs to six paragraphs, and TMR flagged every human text with more than one paragraph in our main test. The paragraph break guide shows how much that one quirk drives its verdicts.
Reading a four-to-three split
If a book everyone knows is human splits the detectors four to three, a split verdict on a student’s story tells you very little. Across the 30 famous texts we tested, every one was flagged by at least one detector. A lone “AI” result from TMR or E5-small LoRA is common for human writing. Desklib’s verdict carries the most weight here, and its results page shows where it still goes wrong. Moby-Dick is another novel opening that Fast-DetectGPT flagged.
Questions people ask
Was Alice in Wonderland written by AI?
No. Lewis Carroll's book came out in 1865. The four detectors that flagged it also flagged plenty of human writing in our main test, from 15.9% of human texts (Fast-DetectGPT) to 83.3% (TMR).
Which AI detectors said Alice in Wonderland was human?
Desklib, MAGE and the OpenAI RoBERTa detector. Desklib gave it 1.2% and had the lowest false positive rate of the detectors that catch a real share of AI text, at 6.3%.
Does a 4 of 7 vote prove anything?
No. Of the 30 famous texts we tested, 8 were flagged by at least four detectors, and all 30 were flagged by at least one. A majority vote among free detectors still misfires on well-known human writing.