Is the US Constitution AI-generated? 6 of 7 AI detectors said yes
No. The Constitution was written in 1787, but 6 of the 7 free AI detectors we tested flagged its opening as AI-generated. That includes Desklib, which gave it a 79.7% AI probability even though it flagged only 6.3% of human texts in our main test. The one detector that passed it, the OpenAI RoBERTa detector, also misses most AI text and caught 9.5% of 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 | AI-generated | 79.7% |
| Fakespot AI text detector | AI-generated | 52.6% |
| MAGE | AI-generated | raw score only |
| E5-small LoRA detector | AI-generated | 58.8% |
| OpenAI RoBERTa detector | Human | 1.0% |
| TMR (Target Mining RoBERTa) | AI-generated | 98.0% |
The passage we tested
We the people of the United States, in Order to form a more perfect Union, establish Justice, insure domestic Tranquility, provide for the common defence, promote the general Welfare, and secure the Blessings of Liberty to ourselves and our Posterity, do ordain and establish this Constitution for the United States of America. Article 1 Section 1. All legislative Powers herein...
Which detectors flagged it, and how much each flag is worth
The passage covers the preamble and the start of Article 1, up to the original seat counts for each state. Six detectors called it AI-generated, but their verdicts do not carry equal weight.
TMR gave it 98.0%, which says little, since TMR called 83.3% of the human texts in our main test AI. Fakespot (52.6%) and the E5-small LoRA detector (58.8%) were not far past their threshold of 0.5, and they flag 34.9% and 50.0% of human writing respectively. MAGE and Fast-DetectGPT flagged it as well.
Desklib’s flag is the one to take seriously. It gave the passage a 79.7% AI probability, though it flagged only 6.3% of human texts in our main test and only 4 of the 30 famous texts we ran. The OpenAI RoBERTa detector alone called it human, at 1.0%, and that detector calls nearly everything human.
Constant reprinting and fixed legal clauses
Exposure is the likeliest cause. The Constitution appears in textbooks, on government sites and in court opinions, so the language models that detectors are built on have probably seen it again and again. Fast-DetectGPT flags wording that a language model finds predictable, and a model that has read a text many times will find it predictable.
The drafting style probably adds to it. Clause after clause follows one pattern (“The House of Representatives shall be composed of…”, “No Person shall be a Representative who shall not…”), the same capitalized nouns return, and the text moves through numbered articles and sections. The excerpt ends with a list of thirteen states and their seats. Classifiers trained on AI output may have learned to read that kind of uniform sentence and tidy enumeration as a machine signal.
Contracts, bylaws and other formal drafting
Contracts, bylaws and policy documents follow the same conventions, and a detector has no way to tell a human drafter following them from a model reproducing them. Desklib’s flag here means even the most careful detector we tested can misread formal writing, so treat a flag on such text as a reason to look closer and no more. Besides scripture and the founding documents, the only famous text Desklib flagged was The Art of War. The famous texts overview has all 30 results, and the false positives page shows the same pattern in everyday human writing.
Questions people ask
Why do AI detectors say the Constitution was written by AI?
We cannot see inside the detectors, but two explanations fit the results. The text is reprinted so widely that the language models behind these tools have probably read it many times, and its legal drafting repeats the same shall-clauses in a fixed order, which makes the wording very predictable.
Which AI detector said the Constitution is human?
Only the OpenAI RoBERTa detector, which gave it a 1.0% AI probability. That pass carries little weight, because the same detector caught just 9.5% of the AI-written texts in our main test.
Will a detector flag an essay that quotes the Constitution?
It might, if the quotation is long. Six of seven detectors flagged this passage on its own, and Desklib, the most careful of the detectors that catch a real share of AI text, was among them.