Is A Modest Proposal AI-generated? 2 of 7 AI detectors said yes
Jonathan Swift wrote A Modest Proposal in 1729, so no. Fakespot still rated it 98.7% likely AI and TMR 97.6%, which makes 2 of 7 detectors in our test that flagged it. The other five called it human, namely Desklib (0.1%), the OpenAI RoBERTa detector (0.2%), E5-small LoRA (20.5%), MAGE and Fast-DetectGPT.
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 | 0.1% |
| Fakespot AI text detector | AI-generated | 98.7% |
| MAGE | Human | raw score only |
| E5-small LoRA detector | Human | 20.5% |
| OpenAI RoBERTa detector | Human | 0.2% |
| TMR (Target Mining RoBERTa) | AI-generated | 97.6% |
The passage we tested
It is a melancholy object to those, who walk through this great town, or travel in the country, when they see the streets, the roads, and cabbin-doors crowded with beggars of the female sex, followed by three, four, or six children, all in rags, and importuning every passenger for an alms. These mothers, instead of being able to work for...
Two detectors flagged it, five passed it
Fakespot’s 98.7% is about as sure as a detector gets, and it is attached to a text from 1729. Fakespot flagged 34.9% of the 126 human texts in our main benchmark, and in our famous texts test it flagged 13 of 30 passages, so a high Fakespot score on human writing is routine. TMR’s 97.6% is easier to explain. It flagged 83.3% of human texts in the benchmark, including every one that had more than one paragraph, and this passage has four.
The five detectors that passed it include the two with the fewest false positives. Desklib, which flagged 6.3% of human texts, gave it 0.1%, and the OpenAI RoBERTa detector, which flagged 4.8% but also misses most AI text, gave it 0.2%. MAGE and Fast-DetectGPT passed it too. Fast-DetectGPT flagged 15.9% of human texts, with a threshold set on our calibration split that gives it a small advantage over the rest.
Old spelling, modern-sounding argument
Two features of the text likely pull in opposite directions. The spelling is from the early eighteenth century: cabbin-doors, stroling, publick, dropt, the Barbadoes. A language model trained mostly on modern English will find those forms hard to predict, which is the likeliest reason Fast-DetectGPT, a method built on predictability, let it pass. Desklib’s model may react to the same oddities.
The argument itself sounds modern. Swift writes as a sober projector who has “maturely weighed the several schemes” and says that whoever finds “a fair, cheap and easy method” for the problem deserves a statue. That tone of a reasonable expert laying out a policy is close to the voice of an AI answer explaining a solution. Fakespot flagged 70.8% of human news reports and 40.0% of encyclopedia articles in our benchmark, so it seems to react to measured, informative prose, the register Swift is imitating.
Reading a near-certain score
A score that close to certain sounds like proof. When you check real writing, look up how often a detector flags human text before reading anything into its score, and do not count two weak detectors agreeing as confirmation. Leviathan was flagged by the same two detectors. Our false positives page and the famous texts results show how common this is.
Questions people ask
Why does an AI detector say A Modest Proposal is AI?
Swift published it in 1729, so any AI verdict is a false positive. Fakespot and TMR gave one, and in our main test they also flagged 34.9% and 83.3% of human texts.
Can AI detectors recognize satire?
We did not test satire as a category, but there is no reason to expect it. Detectors score word patterns, and Swift's satire works by sounding like a calm, reasonable policy paper, so the joke is invisible to them.
Why did Fast-DetectGPT pass it?
Probably the old spelling. Words like cabbin, stroling, publick and dropt are hard for a modern language model to predict, and Fast-DetectGPT reads unpredictable wording as human.