Is It True That AI Detectors Flag Older Pre-ChatGPT Writing as AI?

Yes — AI detectors regularly flag human writing that predates ChatGPT, because they measure statistical style rather than actual origin: formal, polished, conventional prose from 2015, 1995, or 1787 can score “AI-generated,” which is exactly what happened when detectors famously flagged the U.S. Constitution.

This one isn’t a myth to bust so much as a rumor to confirm — and then explain, because the explanation changes how you should read every detection score you’ll ever see. People asking “do ai detectors flag old human writing” usually arrive with a specific, unsettling experience: they fed an essay from college, a 2018 blog post, or a decade-old work sample into a detector, and the meter came back red. Text written before the technology existed, “detected” as its product. Here’s why that happens, why it’s not a fixable bug, and what it should do to your confidence in these tools.

Detectors date nothing. They measure style, not history

An AI detector has no idea when your text was written. There’s no timestamp check, no historical database lookup. The tool receives words and computes statistics about them — chiefly *perplexity*, how predictable each word is given the context, and *burstiness*, how much sentence length and structure vary. Text that is smooth, even, and predictable resembles the default output of language models, so it scores “AI-like.” Text that is jagged, surprising, and irregular scores “human-like.”

Notice what’s absent from that description: any connection to authorship or date. The score is a claim about statistical *conventionality*. And human beings have been producing highly conventional prose for as long as writing has had conventions — which is to say, always.

So when your pre-ChatGPT essay gets flagged, the detector isn’t malfunctioning. It’s doing precisely what it always does: reporting that the text’s surface patterns are regular. The error is in the label the industry puts on that measurement. “AI-generated” is a marketing name for “statistically predictable,” and those were never the same thing.

The Constitution test: the most famous false positive there is

The canonical demonstration arrived in 2023, when writers and researchers started feeding historical documents into popular detectors. Ars Technica documented the results in a widely cited piece: detectors assigned high AI probabilities to the U.S. Constitution and to passages from Genesis. Text written in 1787 with a quill, “likely AI-generated.”

The mechanism is worth understanding because it’s almost elegant. Canonical texts saturate the training data of large language models — the Constitution appears in legal corpora, textbooks, court opinions, and a million web pages. A model that has effectively memorized a text predicts its every next word with near certainty. Near-certain prediction is ultra-low perplexity, and ultra-low perplexity is the loudest “AI” signal a detector can register. The most famous human documents in English are, to a detector, the most machine-like text imaginable.

Your old essays get caught by the milder version of the same effect. You wrote carefully. You followed the conventions of academic English — topic sentences, formal transitions, measured hedging. You revised until the rough edges were gone. Every one of those virtues lowered your perplexity and evened your rhythm. We’ve written about what detectors flagging the Bible and Constitution should tell educators — the short version is that a tool this confidently wrong about the founding documents has told you its confidence isn’t evidence.

Which old writing gets flagged most

Not all vintage text trips the meter equally. The flag-prone profiles are consistent:

• Academic and student writing. The five-paragraph essay is a predictability machine by design. Taught structure, taught transitions, taught hedging — a detector reads the lesson plan, not the student.
• Legal, technical, and business prose. Boilerplate exists because the genre demands sameness. Old contracts, documentation, and reports score machine-like for the same reason new ones do.
• Non-native English writing. The 2023 *Patterns* study by Liang and colleagues found detectors flagged the majority of TOEFL essays by non-native speakers — writing that leans on learned patterns and safer vocabulary. That bias applies to a 2016 essay exactly as it does to one written today.
• Heavily edited text. Revision smooths. The more polished the final draft, the more even its statistical surface — which is why professional writing from any era can score worse than a dashed-off email.

Meanwhile, your rambling 2012 LiveJournal entry would probably sail through as gloriously human. The detector rewards mess, and old mess is as safe as new mess.

What this means if it happens to you — or to your accuser’s logic

The practical stakes: someone runs old writing through a detector and draws conclusions. Two scenarios matter.

Your archived work gets flagged. A hiring manager screens your 2019 portfolio piece; a client “checks” a sample you wrote in grad school. The response is the Constitution argument, politely: the tool cannot date text, it flags formal human prose at documented rates, and its own vendors warn against treating scores as proof. Offering process evidence for recent work — drafts, version history — moves the conversation somewhere a fallible statistic can’t.

You want old writing as an alibi. Tempting, but understand its limits. Clean scores on 2019 essays don’t prove you wrote your 2026 submission. What old work *does* provide is a voice baseline — a consistent style across years of dated documents is genuinely persuasive in a dispute, especially paired with why low perplexity is not proof of AI and process evidence for the disputed piece itself.

There’s also a five-minute experiment that will permanently recalibrate how you read detection scores: run one of your old essays through PaperBleach’s detector. Whatever the result, you’ll have watched the tool judge text whose origin you know with certainty — the fastest education in detector epistemology available.

Frequently Asked Questions

Do AI detectors flag old human writing that predates ChatGPT?

Yes, regularly. Detectors don’t know when a text was written — they only measure how statistically predictable its patterns are. Formal, polished, conventional prose scores machine-like regardless of its date. Famous demonstrations include popular detectors assigning high AI probabilities to the U.S. Constitution and passages from the Bible, and plenty of writers have watched their own 2015 college essays come back “likely AI-generated” from text written years before the technology existed.

Why would a detector flag something written in 2019?

Because the flag was never a claim about time — it’s a claim about style. Detectors measure perplexity (word predictability) and burstiness (sentence variation). Careful academic prose, legal writing, technical documentation, and heavily edited text all trend toward low perplexity and even rhythm, which is the same statistical neighborhood AI output lives in. A meticulous human writer in 2019 and a language model in 2026 can produce measurably similar surfaces.

Does the Constitution really get flagged as AI-written?

Yes — this became a famous demonstration in 2023, covered in depth by Ars Technica, when detectors assigned high AI probabilities to the U.S. Constitution and the Book of Genesis. The reason is instructive rather than silly: canonical texts appear all through AI training data, so language models predict their wording almost perfectly, which reads as ultra-low perplexity — the strongest “AI-like” signal a detector can see.

Can I use my pre-ChatGPT writing to prove I don’t use AI now?

It’s weaker evidence than people expect, in both directions. If your old writing scores human, an accuser can say your new work differs; if your old writing gets flagged, you’ve proven the detector unreliable but not your authorship. The stronger use is comparative voice: a consistent style across years of dated documents supports you in a dispute. Strongest of all is process evidence for the current work — version history, drafts, and your ability to discuss it.

What should educators take from detectors flagging historical texts?

That false positives aren’t a rare malfunction but a structural feature. A tool that assigns high AI probability to text written centuries before computers is announcing what it really measures: statistical conventionality, not authorship. That doesn’t make detectors useless — it defines their proper role as a weak screening signal that opens a conversation, never a verdict that closes one, especially for students whose formal style naturally resembles the flagged profile.

The bottom line

AI detectors flag old writing because they never measured *when* or *who* — only *how predictable*. Formal human prose has always been predictable; that’s most of what “formal” means. The Constitution result isn’t a funny edge case, it’s the tool’s true nature caught in good lighting: a conventionality meter wearing an authorship costume. Read every score with that in mind — including the reassuring green ones — and you’ll use these tools the only way they can be used honestly: as one weak signal, dated nothing, proving nothing alone. More in our myth-busting series, or see pricing if you check drafts regularly.