The United States Declaration of Independence, signed in 1776, has suddenly been “accused” of being machine-written, not by a conspiracy theorist, but by an AI detection tool. The verdict is obviously wrong, yet the incident exposes a deeper problem: how can we trust any claim about who – or what – wrote a text in an age of generative AI?
A founding text flagged as 98.51% AI-generated
The odd story begins with a simple test. SEO specialist Dianna Mason ran the full text of the US Declaration of Independence through an AI detection system. The result: the tool said the document was generated by artificial intelligence with a confidence of 98.51%.
The detector treated one of the most iconic human political texts as almost entirely machine-made.
From a historical standpoint, the claim is absurd. No large language model existed in 1776. The technology behind tools like ChatGPT reached the public only in 2022, nearly two and a half centuries later. Yet the software produced a numerical score, wrapped in the authority of percentages, that could easily mislead anyone unfamiliar with the limitations of these systems.
For Mason and many others watching this field, the result is less a joke and more a warning. If a foundational historical document can be mislabelled as AI-generated, what does that say about how schools, universities or employers are using these same tools right now to decide whether people are cheating?
AI detectors under fire for false accusations
The Declaration of Independence is not the only victim of this technological confusion. Other experiments have pushed long-established texts through detection systems with similarly strange results.
Legal documents from the 1990s have been labelled as AI-written. Passages from the Bible have triggered “high probability of AI” alerts. These works pre-date modern generative AI by decades or centuries.
When sacred texts and court records are flagged as machine-written, the trustworthiness of AI detectors starts to crumble.
Specialists in academic integrity and online publishing have raised concerns for months. Many universities quietly rely on AI detectors to support plagiarism cases. Some lecturers have admitted using these tools to decide whether to open misconduct investigations on student essays.
➡️ Semaglutide May Reverse Damage Caused by Osteoarthritis, Study Suggests : ScienceAlert
➡️ Goodbye hair dyes : the new trend that covers grey hair and helps you look younger emerging
Yet the tools often behave more like pattern-guessing systems than forensic instruments. They hunt for features such as:
- Highly regular sentence structures
- Low variation in vocabulary and rhythm
- Predictable word choices and phrasing
- Unusual consistency across long texts
The problem is that careful, formal writing by humans – especially in legal or religious contexts – can look very similar to AI output on the surface. That is precisely why old documents, drafted in rigid, structured language, tend to trigger false positives.
Why the question “who wrote this?” is getting murkier
In 1776, identifying a text’s origin was far more straightforward. Most documents were handwritten, and the idea of a machine generating natural language was science fiction. Authorship was tied to visible ink, signatures and eyewitnesses.
Today, a polished PDF or a blog post gives no visual hint about whether a human typed every word, an AI helped, or the text was fully generated by software. Style, once our intuitive clue, is now a moving target.
As Mason notes, the real debate may not be whether a text is “purely human” or not, but whether readers actually care, and in which contexts that origin changes how the text should be judged. She told Forbes that when people learn something was made by AI, many immediately distrust it – at least for now.
The social meaning of “AI-generated” may matter more than the technical definition.
Entrepreneurs like Benjamin Morrison, also quoted on the topic, frame it differently: times shift, technology advances, and public attitudes often adjust. What feels suspicious today may feel normal in a decade, just as typing on a computer once felt less “real” than handwriting.
Where AI detection really matters
While mislabelling the Declaration of Independence is amusing, the stakes in other arenas are far higher. Misfiring detectors can affect grades, careers and reputations.
In classrooms and universities
Since 2022, educators have watched AI tools rapidly enter student life. Some institutions reacted by banning ChatGPT; others tried to integrate it into teaching. Many, though, turned to AI detection tools to police essays and exam answers.
That raises several risks:
- Students wrongly accused of cheating on the basis of a detector’s score
- Pressure on teachers to treat the software as a final verdict
- Inconsistent rules between courses and institutions
- Greater inequality, as confident students challenge accusations while others stay silent
Legally, a percentage score from an opaque algorithm is a fragile foundation for sanctions. Ethically, it places heavy weight on a technology that still misidentifies centuries-old texts.
In journalism, politics and public trust
Newsrooms are experimenting with AI tools for drafting headlines, summaries and even entire articles. Political campaigns are already testing AI-generated messages tailored to specific voter groups.
In these spheres, the question of authorship links directly to transparency and trust. A campaign speech drafted largely by AI may still represent the candidate’s views, but voters could reasonably want that fact disclosed. An AI-written news piece framed as human-authored reporting risks misleading readers about how editorial judgment was applied.
Misuse of AI detection cuts both ways: it can hide machine-written content, or wrongly accuse human writers of relying on bots.
Can we ever reliably spot AI writing?
Technically, researchers are trying several strategies to help separate human and machine prose. None offers a perfect fix.
| Approach | What it tries to do | Main limitation |
|---|---|---|
| Stylometric analysis | Study word patterns and sentence shapes | Humans and AIs can both mimic each other’s styles |
| Watermarking AI outputs | Bake hidden patterns into text produced by models | Easy to break by rephrasing or partial editing |
| Metadata tracking | Log which tools were used during writing | Depends on co-operation and honest reporting |
| Hybrid human review | Combine software flags with expert judgment | Time-consuming and still imperfect |
At the same time, AI models are trained to sound increasingly human, with more variation, subtlety and even simulated mistakes. Any static detection method risks turning into an arms race: as detectors improve, text generators learn to evade them.
New norms instead of perfect policing
Given the shaky reliability of detectors, many specialists argue for a shift in focus: from “catching” AI to managing how it is used. That means clearer rules around transparency, consent and accountability.
Some practical ideas being tested include:
- Mandatory disclosure when AI tools are used in scientific papers or news articles
- Assessment formats in education that include oral defences or in-class writing
- Official guidelines on acceptable AI assistance in workplaces
- Industry standards for watermarking content generated by major models
These approaches do not rely purely on detectors. Instead, they try to tie responsibility to people and processes, not just to algorithms scanning finished text.
What terms like “AI-generated” actually mean
The debate is also muddied by language. People use phrases such as “AI-written”, “AI-assisted” and “human-authored” as if they were neat categories. In practice, the boundaries are messy.
A typical scenario might look like this: a student asks a chatbot for a rough outline, then rewrites every section in their own words. Is that AI-generated or human-authored? A journalist feeds bullet points into a model, then heavily edits the result. A novelist occasionally asks an AI to suggest alternative phrasings for one tricky paragraph.
Most real texts will sit somewhere on a spectrum between fully human and fully machine-made.
That nuance rarely appears in simple detector results. A single percentage score can hide the reality that many modern documents are collaborative efforts between humans and software. Designing rules that reflect this spectrum, rather than a strict binary, may be one of the harder policy tasks of the next decade.
Ethical risks and future scenarios
If institutions keep treating AI detectors as oracles, several long-term risks grow. False accusations may erode trust between students and teachers. Companies could use flawed tools to screen job applications, punishing candidates whose writing style just happens to trigger a “likely AI” signal. Governments might rely on detection scores to regulate online speech, with messy consequences for free expression.
At the same time, ignoring origin entirely carries its own dangers: deepfake political speeches, fabricated expert statements and large-scale misinformation campaigns all become easier when no one asks who actually wrote the words.
One possible future involves people treating AI authorship a bit like photo editing. Most pictures today are adjusted somehow, but heavy manipulation must be clearly flagged in journalism or scientific evidence. Text could follow a similar path: light AI help becomes accepted, while undisclosed full automation, especially in high-stakes contexts, faces stronger scrutiny.
For now, the strangest part of the story remains the image of a 21st-century algorithm confidently labelling a parchment from 1776 as machine-made. The glitch is absurd, but the questions it raises about trust, evidence and authorship are very real, and they are landing on teachers’ desks, newsrooms, courts and parliaments far beyond that famous document in Philadelphia.