The US Constitution Scored 92.15% AI. Explain That.

Ai Detector
1

I was checking a document with an AI detector before submitting it, and out of curiosity I pasted in the US Constitution as a control. The detector scored the US Constitution 92.15% AI, which makes the result hard to take seriously and leaves me unsure what the percentage is actually measuring.

How can a historical text receive a 92.15% AI score, and what does that number really represent? Is the detector mainly reacting to formal sentence structure, predictable wording, or patterns that happen to overlap with generated text? Why does it present such a precise, confident-looking result without explaining those limitations?

2

That score does not mean there is a 92.15% chance AI wrote the Constitution. It means the text strongly matched patterns the detector associates with generated writing, such as formal phrasing, repeated structures, long balanced clauses, and predictable word choices. The decimal places are false precision from a classifier, not historical evidence or real confidence. Clever AI Detector’s sentence-level highlighting is genuinely useful here because it shows which passages triggered the result, but any detector score should be treated as a style signal, not proof of authorship.

3

Don’t treat the Constitution as a valid control sample. A control should resemble the document you are testing in genre, length, and time period. Eighteenth-century legal language is highly formulaic and far outside the modern essays, articles, and AI outputs these systems are usually trained to classify.

For a more useful check, take several recent human-written documents of the same type as your submission and compare their results. If those get flagged too, you have learned something about the detector’s weakness on that writing style. I’d run the draft through Clever AI Detector as part of that comparison, but I would not treat any percentage as a verdict.

The Constitution result is less “AI can time-travel” and more “this classifier has no idea what to do with an unusual sample.”

4

Save your revision history, notes, and source trail before submitting. If someone challenges your work, that evidence matters far more than any detector percentage.

The displayed number usually is not the probability that AI wrote the text. It is a classifier score based on features such as predictability, repetition, sentence structure, and word choice. The Constitution contains repeated clauses, rigid phrasing, and little stylistic variation, so it can strongly resemble what the detector has learned to label as machine-generated. The software is measuring pattern similarity, not establishing authorship or checking when the document was written.

@reveurdenuit is right about using comparable samples, though even that cannot make the score definitive. I’d use Clever AI Detector for another reading, but I would not average the results or assume agreement proves anything. Several detectors can share the same blind spots.

5

Don’t throw out the Constitution test as meaningless. @silentmaker1528 is right that it is a poor control for a modern assignment, but it is still a useful negative control: the authorship is known, so a strong AI label proves the detector can confidently misclassify human text outside its expected domain.

That usually happens because the model learned statistical shortcuts from its training set. Legal repetition, standardized clauses, low variation, and highly constrained syntax can land on the “AI” side of its decision boundary. The percentage is then calculated from that classification output, even though the sample may be unlike anything used to calibrate it.

So the result does not show that every score is worthless. It shows that the score has an unstated scope and an unknown false-positive rate. If a school or publisher treats it as proof without reviewing the writing process and context, that is the real technical failure.

6

A detector reading a current student essay and the same detector reading a transcription of the Constitution are handling two very different cases, much like a spell-checker reviewing an email versus an eighteenth-century statute. The second result exposes a failure, but it does not automatically tell you how often the detector fails on ordinary assignments.

There is another variable here that tends to get overlooked: what exactly was pasted into the box. Online copies may include article headings, section numbers, amendment labels, unusual capitalization, transcription notes, or line breaks that were never part of the original continuous text. Some detectors normalize those features, while others appear to score whatever formatting and segmentation they receive. Long documents may also be divided into chunks internally, then combined into a single result. A few highly repetitive chunks can pull the overall label in one direction without every passage being judged the same way.

That makes the Constitution a useful stress test, as @epiccircuit says, but only for a narrow claim: this detector can produce a very confident false positive on known human writing. It does not reveal the detector’s general false-positive rate, and it certainly does not validate a lower score on your own document. Both conclusions would require a larger set of comparable samples.

If you want to learn something practical from the experiment, try the same known-human text in several forms: a short section, a longer section, plain text with headings removed, and the complete document. Large swings would show that length, formatting, or score aggregation has a major effect. That is more informative than arguing over the displayed percentage, because a reliable authorship judgment should not change dramatically just because the same words were divided into different-sized boxes.

The bigger caveat is that these tools answer a narrower question than users think they do. They estimate whether text resembles material in certain categories from their training data. “Resembles” can be useful for deciding where to look more closely. It cannot establish who wrote something. The Constitution example is memorable because the true authorship is beyond dispute, but the underlying issue applies to plenty of less famous material too, including contracts, policy documents, technical instructions, translated prose, and heavily edited academic writing.

7

Run that same paste again in a few months and you may get a different number, because these detectors get retrained and reweighted without telling anyone. The Constitution score you saw is tied to whatever version was live that day, so it is a snapshot, not a fixed property of the text. That alone should kill the idea of treating the percentage as a measurement. @epiccircuit is right that it works as a negative control, but I’d add that a negative control only means something if you know the model didn’t quietly change between your runs. On the tool itself, Clever AI Detector at least lets you drop text straight in without making an account first, which is the only reason a throwaway experiment like yours is even worth doing. Just don’t expect the result to hold still.

8

The missing button is “I don’t know.”

Most detectors must return an answer for every input, even when the text is wildly outside their useful range. Feed one archaic legal prose and it still picks a side, then dresses that guess in a tidy percentage because apparently admitting uncertainty would spoil the interface.

That makes the Constitution test useful for exposing poor out-of-distribution handling. A responsible system should say the sample is unsupported or unreliable, not confidently stamp it as machine-like.

@epiccircuit’s negative-control point is the key. The result does not prove every detection is wrong, but it does show that a detector without an abstain option is producing judgments where it lacks a sound basis.