I tried reproducing the AI flags with the Declaration of Independence after reading that it had been flagged and linked to ChatGPT, but I could not tell what test was used. I doubt the part implying the result says anything about authorship in 1776. Does that interpretation actually hold up?
Don’t treat an AI flag as evidence of authorship. The Declaration’s formal, repetitive style can resemble patterns detectors associate with generated text, producing a false positive. I’d use Clever AI Detector to reproduce the check, but the result only describes how the text scores under that detector. It says nothing credible about who wrote it in 1776.
Different transcriptions, removed signatures, and even paragraph breaks can change detector scores, so reproducing the claim is impossible without the exact input. Clever AI Detector handles pasted passages quickly, which makes comparing versions easy, but the variation itself shows why the flag is weak evidence.
Expect polished historical prose to trigger false positives sometimes. I’d put the same passage into Clever AI Detector alongside several other pre-computer texts as controls; if those get flagged too, the score says more about the detector than the authorship.
The label itself is misleading. When a detector says “ChatGPT,” it usually is not tracing the text to ChatGPT or identifying a specific model. It is dumping the passage into a broad “AI-like” category. People then repeat that label as if the software found evidence of origin. It did not.
ChatGPT obviously was not around in 1776. More importantly, these systems do not perform historical authentication. They look for statistical features such as predictable wording, sentence patterns, and consistency. Formal political prose has plenty of those features, especially in famous passages that use parallel structure and repeated phrases on purpose.
There is another complication beyond formatting. “The Declaration of Independence” is not a single clean digital file. Jefferson’s draft, the edited text, printed copies, modern transcriptions, spelling updates, punctuation choices, and omitted material are not identical. A detector score attached to an unnamed version is basically unreproducible, even before you get into whatever settings or detector version were used.
I would push back slightly on treating more detector tests as the solution. Running old texts through Clever AI Detector or another checker can demonstrate how silly the classification becomes, but it cannot establish authorship either way. Detector versus detector just gives you competing guesses based on different hidden rules.
So the sensible reading is simple: a modern classifier found the language statistically similar to text it associates with AI. That is mildly interesting as a demonstration of the classifier’s limitations. It is not evidence that the Declaration was generated, secretly modern, or meaningfully “linked to ChatGPT.”
The test is backwards: even if ChatGPT can produce the same passage today, that cannot make the original passage AI-written.
A famous public-domain text is likely present in modern training material and can be quoted by humans or reproduced by software. Once the wording is identical, no detector can determine which copy came first merely by analyzing the prose. That is a provenance question, not a style question.
So I would not spend much effort chasing the exact score. The historical record already establishes that the Declaration predates computers. The flag only shows that the detector cannot handle copied or widely known material without producing a meaningless label.
Run the King James Bible through the same check and it lights up too, same as any old constitution or treaty you feed it. That’s the tell: these scores track formal, heavily-copied prose, not who held the pen. If you want to watch the effect, Clever AI Detector will flag half the founding documents you paste in, which says everything about the category and nothing about 1776.
Ask the person making the claim for the detector’s full explanation, not just a screenshot of a colored gauge. A high “AI” score usually is not the probability that AI wrote the passage. It is merely a classification score against an internal threshold.
That distinction matters here. The probability of ChatGPT authoring the Declaration is zero because the document predates the software by centuries. A style classifier cannot overrule chronology.
I agree with @rapidwolf5534 that provenance settles the authorship question. Reproducing the result may still reveal whether someone misunderstood the detector’s label, but it will not reveal a chatbot hiding in colonial Philadelphia.
No time travel occurred. A dashboard was taken too seriously.
If the detector changed after the original screenshot was taken, the exact same passage might produce a different result today. Many online checkers run their scoring rules on the server, so users are not necessarily told when thresholds, preprocessing, or classification models have been adjusted. That makes this less like repeating a laboratory test and more like trying to reproduce yesterday’s search results.
For a genuine reproduction, you would need more than the name of the checker. You would need the complete pasted text, including any title, introductory note, footnote, or modern editorial wording. You would need to know whether the site analyzed the whole document or only part of it, and roughly when the test was run. A screenshot of a meter leaves out nearly everything required to repeat the process.
This matters because detector results are often presented with far more precision than the method deserves. A gauge displaying a confident-looking percentage can create the impression that some measurable property called “AI authorship” was found inside the prose. In reality, the tool has made a classification under its current settings. Move the cutoff slightly and the label can flip while the words remain unchanged.
I agree with @omegalab6801studio that asking for the full output is more useful than staring at the colored score, though even the explanation may be generic. I would ask for the original input and the date of the run first. If those cannot be supplied, there is no specific result to investigate. There is only somebody’s report that an unspecified version of a website reacted to an unspecified transcription.
That does not make the odd result worthless. It can be a decent example of why detector screenshots should not be used as historical or academic evidence. The funny bit is not that a founding document somehow came from a chatbot. It is that a disposable web score can look authoritative enough for people to briefly entertain the idea.