A routine experiment with an AI-detection tool has triggered a strange claim about a cornerstone document of American history.
The US Declaration of Independence has survived wars, fires and political storms for nearly 250 years. Now it faces a challenge from a far more unexpected rival: a commercial AI detector that insists the 1776 text was “almost entirely” written by artificial intelligence.
An AI detector versus the founding fathers
The controversy began when Dianna Mason, a US-based search engine optimisation specialist, plugged the Declaration of Independence into an AI-detection service out of sheer curiosity.
The tool returned a startling verdict: the document was supposedly 98.51% “AI generated”. In other words, the software rated one of the most studied pieces of human writing as almost certainly machine-made.
The same type of tools that some universities now use to spot student cheating just flagged Thomas Jefferson as a probable chatbot.
No serious historian or technologist believes the Declaration was produced by a machine. In 1776, the concept of a large language model did not exist, and electricity itself was still a scientific curiosity. ChatGPT would not arrive until 2022, nearly two and a half centuries later.
Yet the incident exposes a growing problem: AI detectors are increasingly treated as forensic instruments, while their error rates remain high and poorly understood.
Historic texts branded as “AI-generated”
The Declaration is not the only victim. Over the past year, researchers and journalists have fed a series of old documents into detection tools, with oddly similar results.
- Legal case summaries from the 1990s scored as “likely AI”.
- Passages from the Bible were labelled machine-written.
- Classic literature sometimes triggered “high AI probability” warnings.
These false alarms show how brittle current systems can be. Many detectors are trained mostly on recent Internet text and on examples of AI output generated by modern models. When confronted with unfamiliar styles, archaic phrases or highly structured prose, they can misfire dramatically.
When a 30-year-old legal brief or a centuries-old religious text is branded as AI-written, the tool is telling you more about its own limits than about the document.
This matters because the same software is now being used in classrooms, newsrooms and offices. Students have reported being accused of cheating based solely on an automated score. Some editors quietly screen freelancers’ work with detectors, sometimes without explaining the process.
How AI detectors actually work
Most tools on the market do not “recognise” AI in the way a fingerprint scanner recognises a person. Instead, they estimate how predictable a string of words looks.
AI-generated text, especially from older or poorly tuned models, tends to be highly regular. Sentences often follow familiar patterns. Certain phrases appear again and again. That statistical smoothness can make it easier to spot.
Human writing is typically more jagged. People repeat themselves, switch tone mid-paragraph, or wander away from the main point. Paradoxically, those flaws can be signals of authenticity.
Modern AI systems, though, are trained precisely to sound more human. They add variation and noise to mimic our quirks. As they improve, they creep ever closer to the messy middle ground where detectors struggle to tell man from machine.
Why old documents confuse modern detectors
Historical texts like the Declaration of Independence add another layer of complexity.
- They use archaic grammar and spelling that rarely appears in current training data.
- The rhetoric is formal and repetitive, a style that resembles the “over-polished” tone of some AI outputs.
- Sentences are long and heavily structured, which can look statistically regular to an algorithm.
To a 2020s detector, Jefferson’s dense, carefully crafted prose can look oddly similar to a model fine-tuned to sound authoritative and legalistic.
Does the origin of a text really matter?
Mason argues that the more interesting question is not whether a text was produced by AI, but whether the audience cares.
“When people know something is AI-created, many still walk away from it,” she told business magazine Forbes. For now, that instinctive distrust shapes how organisations present their content, even when they quietly rely on tools behind the scenes.
The value of a text may end up judged less by who — or what — wrote it, and more by whether it is accurate, honest and useful.
Some entrepreneurs share a pragmatic view. As one put it, “Times change, technology moves on.” They see AI not as a threat to authorship but as just another step in the evolution from quills to typewriters to laptops.
Yet for teachers, publishers and regulators, the question of origin has concrete consequences. Exams, professional certifications and legal documents often demand clear human accountability. If detectors cannot provide reliable evidence, institutions have to rethink how they police integrity.
Ethics, ownership and the line between help and cheating
The spread of generative AI has scrambled long-standing assumptions about creativity and responsibility. If a student uses a chatbot to rewrite clumsy sentences, is that editing help or plagiarism? If a journalist leans on AI to propose headlines, who owns the final wording?
Different sectors are already drawing the line in different places:
| Context | Typical stance on AI use |
|---|---|
| Universities | Allow support tools in drafts, ban fully generated essays |
| Newsrooms | Permit AI for research and summaries, require human-written final copy |
| Marketing | Often embrace AI drafts, with human polishing for tone and accuracy |
| Courts and law firms | Use AI cautiously for search, demand human verification of every citation |
In all these settings, detectors are sometimes treated as referees. The Jefferson episode shows how shaky that role can be.
Practical risks of trusting detectors too much
Overconfidence in AI detection can create its own harms.
- False accusations: Genuine writers may be penalised based on a single automated score.
- Unequal treatment: Non-native speakers and very polished writers often trigger higher “AI probability” scores.
- Complacency: Institutions may stop investing in human review, assuming software will catch every case of misconduct.
- Arms race: As some tools promise to “evade detection”, both side’s tactics become less transparent.
When a detector wrongly brands the Declaration of Independence as machine-made, it underlines a wider point: these tools are aids, not judges.
For educators and managers, a healthier approach is to combine multiple signals: conversation with the writer, drafts and notes, version history in documents, and, only then, automated analysis as one datapoint among several.
Key terms readers keep hearing
What “AI-generated” really means
In everyday news coverage, “AI-generated text” usually refers to content produced by a large language model such as GPT, Claude or Gemini. The system predicts the next word based on patterns learned from billions of sentences.
There is no understanding in the human sense. The model does not “know” the Declaration of Independence; it recognises patterns that look like similar documents and reproduces them statistically.
What an “AI detector” is actually detecting
When headlines say a detector has “found AI”, what has really happened is closer to this: the software has assigned a probability score that the statistical fingerprint of the text resembles the outputs on which it was trained.
Those scores are influenced by length, structure, repetition, common phrases and average sentence complexity. Short texts are especially hard to judge; a four-line email can often trigger wildly different readings across platforms.
Where this could go next
Looking ahead, experts expect two trends to collide. Language models will continue to grow more fluent and personalised, making their output harder to distinguish from a careful human writer. At the same time, researchers will build more nuanced detectors that look beyond surface statistics and incorporate metadata such as editing history.
Some technologists suggest a different route altogether: watermarking AI output at the model level, so text carries an invisible signal of its origin. Others push for social solutions, such as honour codes, disclosure norms and random oral checks for high-stakes work.
The strange moment when a commercial detector effectively accused the American founders of using a time-travelling chatbot may age as an amusing anecdote. For students, professionals and regulators grappling with AI’s rapid advance, it also serves as a cautionary case study in what happens when automated certainty meets messy human history.
People Who Don’t Feel The Need To Dye Their Hair As They Age Often Share These 8 Remarkable Traits
