Did AI Write the U.S. Declaration of Independence?

Opening

Lauren Jaeger, a chemistry major at Idaho State University, was applying to PhD programs when she ran into a strange warning. Some application portals stated that “if AI-generated traces are detected in your personal statement, your entire application will be voided.” Jaeger had never used AI. But out of anxiety, she ran her own writing through a few online detectors — and every single one flagged it as “nearly 100% AI.”

So Jaeger rewrote her personal statement. Not to make it better, but deliberately worse, to dodge the detectors. She kept lowering the score until it hit 30%, then submitted it thinking “this should do.” (She was, thankfully, admitted to a PhD program at the University of Utah.)

Around the same time, something even more absurd went viral online. Someone ran the U.S. Declaration of Independence — written in 1776 — through a detector, and it came back “95–100% AI-written.” A 250-year-old document.

Here’s the question that raises: can we really trust these tools that universities are using to catch cheating? The bottom line, up front: the real problem isn’t that detectors are sometimes wrong. It’s that a tool capable of being wrong is being treated as if it were irrefutable proof.

🔍 Why Universities Keep Reaching for Detectors

With the arrival of generative AI like ChatGPT, the cost of having someone (or something) else write your text has collapsed to essentially zero. Cath Ellis, who oversees academic integrity at Western Sydney University, calls this “a fundamental shift in scale.” Ghostwriting existed before, she says, but now “the sheer volume of output that’s at least substantially fabricated or manufactured has exploded.”

The trouble is that the old weapons don’t work well anymore. Plagiarism checkers like Turnitin work by matching text against a vast archive of existing publications and sentences — hunting for signs of copying. But sentences generated by AI aren’t lifted wholesale from anywhere; they’re freshly assembled on the spot. With no original to match against, there’s nothing for a plagiarism checker to catch. It’s an irony: AI produces output by training on a vast trove of past writing, yet plagiarism tools that compare against past writing can’t catch it.

That’s the gap AI detectors claim to fill — tools that promise to determine “was this written by another AI?” Copyleaks, GPTZero, ZeroGPT, and tools built by Grammarly, QuillBot, and Turnitin are all on the market. Facing a flood of assignments and applications, universities and graduate programs have started leaning on these tools.

Why “Well-Written” Gets Flagged as AI

Many detectors rely on a metric called perplexity¹ — in plain terms, a measure of how predictable the next word is. AI-written text tends to follow statistically smooth, predictable patterns. So text with low perplexity (highly predictable) gets suspected of being “written by a machine,” while text full of jarring, unpredictable phrasing gets judged “human.”

See the problem? By this logic, the more precisely someone follows grammar rules and writes cleanly, without clutter, the more likely they are to be suspected of using AI. This is exactly what Jaeger meant when she said, “I studied grammar books and I write by strictly following the rules — I think that’s why I got flagged as AI.” She got caught for writing well.

Here’s where it gets more serious: non-native English speakers. In 2023, a Stanford research team ran 91 TOEFL essays — all written before ChatGPT existed, prior to 2020 — through seven detectors.

The results were stunning. On average, 61.3% were wrongly classified as “AI-written.” Meanwhile, essays written by American 8th graders (ages 13–14) were classified as “human” almost flawlessly.

Non-native writers tend to use less varied phrasing, which produces lower perplexity — and the detectors mistook that for “sounding like AI.” A 61% false-positive rate² means 6 out of every 10 students flagged were flagged unjustly.

The Declaration of Independence got caught by the same mechanism. After 250 years of being quoted and refined countless times, its phrasing has become statistically extremely “predictable.” Nature itself ran excerpts of the 1776 original through ZeroGPT multiple times, and every single time it came back 95–100% AI.

Is this some freak accident? No. A 2025 study evaluated GPTZero, one of the most widely used detectors. It caught fully AI-generated text with high confidence — but misidentified human-written text as AI about 16% of the time. The researchers concluded that “reliability in distinguishing human-written text remains limited.”

⚡ Even a More Accurate Detector Leaves a Problem Behind

Not every detector performs this poorly, of course. Pangram Labs, based in New York, claims a false-positive rate of “essentially zero.” Their approach is different: they train on massive amounts of human-written text alongside AI rewrites of that same text, and update the model to learn each new chatbot’s style as it emerges. They don’t rely on perplexity alone. Independent evaluations by the University of Chicago’s Booth School of Business and the University of Maryland found accuracy never dropped below 99.8%.

But here’s the point I want to underline. Even a more accurate detector doesn’t end the problem.

First, there’s hybrid text. Mike Perkins, who studies AI’s impact on academia at British University Vietnam, says, “Detectors are pretty good at catching pure AI text, but the moment you start tweaking that text, detection falls apart.” A human doesn’t even need to touch it — you can simply ask another AI to “rewrite this,” or run it through a “humanizer³” tool designed specifically to lower detection scores. Every time detection companies try to catch these workaround tools, new workaround tools appear. In Perkins’s words, it’s “a massive arms race that helps nobody.”

Second, there’s a deeper problem. This year, a team led by Marzena Karpinska used Pangram to analyze 186,000 articles across 1,500 U.S. newspapers from the summer of 2025. About 9% were detected as partially or fully AI-written. An interesting figure. But Karpinska herself was quick to draw a line: this can be used to observe “large-scale trends,” but it should never be used as grounds to judge an individual author’s guilt.

“You can’t use this to mass-disqualify people.”

That line is the crux of today’s story. The same tool can be valid for trend analysis while being unfit for judging an individual. Statistics and verdicts are different games.

Oswald’s Lens

Honestly, I don’t read this story as a debate about “detector performance.” As someone who’s worked with data, what jumps out to me first is threshold and base rate.

There’s a trap I’ve run into repeatedly while doing data analysis: even a model with 99.9% accuracy will still produce a lot of wrongly accused people, if the population is large enough and true positives are rare enough.

Say you have an excellent detector with a 0.1% false-positive rate. Apply it to 10,000 students who all honestly wrote their own work, and about 10 of them get branded “AI cheaters” despite doing nothing wrong. 99.9% accuracy is zero comfort to those 10 people.

This connects to what Perkins pointed out. People see a score and just believe it. In the plagiarism-checker era, there was visible evidence: “this sentence here is identical to that paper.” AI detection scores offer no such evidence. There’s no way for a student or a professor to verify what an “87% AI” figure is actually based on. In my work building go-to-market strategy, I’ve seen countless times how easily “confidence dressed up as a number” can mislead an organization. A score is supposed to assist judgment — but it often ends up replacing it.

So here’s where I land. Detectors can be used as an investigative lead. But the moment they’re used as proof of guilt, the first people caught in the net are exactly like Jaeger — people who write well, and non-native speakers. I’m not saying scrap the tools. I’m saying give their output only the weight it actually deserves.

Closing

Three takeaways from today:

First, many AI detectors mistake well-written prose and non-native speakers’ writing for AI, with false-positive rates reaching as high as 61%. Second, even more accurate detectors break down against hybrid text and humanizer tools — and more fundamentally, “trend” and “individual judgment” are different problems. Third, the real danger isn’t that a tool is sometimes wrong — it’s treating a score that can be wrong as if it were evidence.

If you’re ever in a position to evaluate or submit written work, remember just this one thing: an AI detection score is a question to start with, not a verdict to end with.

Jaeger deliberately made her own writing worse to get admitted — sabotaging her own words just to dodge a detector. Have you ever had your own writing, or a student’s or colleague’s, wrongly flagged as AI? Tell me what happened in the comments. I’ll consider it for a future issue.

---

💬 Been wrongly flagged as AI? Tell me about it in the comments — I might feature it in a future issue.

📨 If you think this could help a fellow writer, please pass it along.

---

References & Further Reading

Primary sources

• Weixin Liang et al., “GPT detectors are biased against non-native English writers”, Patterns (Cell Press), 2023. : This is the source of the 61.3% false-positive rate — the first paper to document bias against non-native writers with hard data. Worth starting from the Results section.
• “Assessing GPTZero’s Accuracy in Identifying AI vs. Human-Written Essays”, arXiv preprint, 2025. : Source of the 16% GPTZero false-positive figure.
• Jenna Russell, Marzena Karpinska et al., “AI use in American newspapers is widespread, uneven, and rarely disclosed”, arXiv preprint, 2025. : The study finding 9% AI detection in newspaper articles. The authors’ restraint — “good for trends, not for individual judgment” — is striking.
• W. H. Walters, “The Effectiveness of Software Designed to Detect AI-Generated Writing: A Comparison of 16 AI Text Detectors”, 2023. : The study finding only 3 of 16 tools (Copyleaks, Turnitin, Originality.ai) performed strongly.
• M. Suvanto et al., preprint, arXiv, 2026. : Source of the “telltale signs of detection” graphic referenced in the article, visualizing how perplexity works at the word level.

Background

• Pangram Labs’ collection of third-party evaluations, “Pangram evaluations”, “Chicago Booth Review”. : Check here if you’re curious how the “near-zero false-positive rate” claim was verified.

Original article

• “Universities are relying on AI-detection software to catch cheating. How well do the programs work?”, Nature (2026). : The article that sparked today’s newsletter.

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. Perplexity: a measure of how predictable the next word in a text is. Lower perplexity (“more predictable”) gets judged as more AI-like; higher perplexity gets judged as more human. ↩

2. False positive rate: the rate at which text actually written by a human is wrongly judged to be “written by AI.” Think of it as the rate at which innocent people get caught. ↩

3. Humanizer: a tool that reworks AI-generated text to look more human, lowering its detection score. ↩