How to tell if ChatGPT wrote something

An honest answer: usually you cannot tell reliably from the text alone. Here is what actually works, what does not, and which signals are myths.

By didaiwritethis.ai

The honest answer

From the text alone, usually you cannot — not to a standard that should support any consequence for the person who wrote it. There is no reliable public method for proving a given passage came from ChatGPT. Anyone claiming otherwise is selling something or has not looked closely at the error rates.

That is an unsatisfying answer, so it is worth being precise about which parts are hopeless and which are merely hard.

Why the text alone is so weak

OpenAI has not released a public watermark detection tool for ChatGPT output, and the classifier it did release in early 2023 was withdrawn six months later for low accuracy. Without a watermark to look for, third-party detectors are reduced to inferring from style — measuring how predictable the writing is and reporting that as a probability.

That approach fails in a specific, documented way: it flags plain, conventional prose. Which means it misfires on non-native English speakers, on formal academic writing, and on anyone who writes clearly rather than distinctively. The evidence on detector accuracy covers this in detail. The summary: a detector score is not evidence.

Signals that do not work

A lot of confident folklore circulates about tell-tale signs. Most of it does not survive contact with base rates.

• The word "delve", and its cousins — "tapestry", "testament", "navigate the complexities". These genuinely are more common in model output. They are also words that real people use, disproportionately people trained in formal or academic registers. Accusing on vocabulary alone means accusing a lot of humans.
• Em dashes and tidy tricolons. Popular as a tell, useless in practice. Plenty of careful writers punctuate this way, and anyone can delete a dash.
• "It sounds like AI." This intuition performs poorly under controlled conditions. People are confident and wrong in both directions, and the confidence does not track the accuracy.
• Asking ChatGPT whether it wrote something. It has no memory of what it generated and no ability to recognise its own output. The answer is invented. This one is worth stating plainly because it comes up constantly.

What actually helps

Ordered roughly by how much weight it can bear:

1. Process evidence. Document version history, draft timestamps, commit logs, notes. Google Docs and Word both retain revision history; a piece written over several sessions looks different from one pasted in whole. This is the strongest ordinary evidence available, and the hardest to fake convincingly.
2. A conversation about the work. Someone who wrote something can explain why they cut a section, where an argument came from, what they nearly said instead. Someone who did not, generally cannot — and this is a fair, low-tech check that gives the person a chance to respond.
3. Verifiable specifics. Models fabricate citations, statistics and quotes with great confidence. Checking whether the sources exist and say what they are claimed to say catches a lot, and it catches something real rather than something stylistic.
4. Watermark detection, where the model supports it. Not applicable to ChatGPT today, but a genuinely different class of evidence where available. A positive result is strong; a negative result means little. How watermarking works.

If you are the one being accused

Detector screenshots get treated as proof far more often than they deserve. Reasonable things to ask for:

• What tool produced the score, and what is its documented false positive rate?
• Has it been evaluated on writing like yours — same register, same first language?
• What evidence exists besides the score?

Your own version history is usually the most persuasive answer available, which is a good reason to draft somewhere that keeps it.

Where this is heading

Watermarking is the most promising direction, and adoption is uneven — Google has deployed it on Gemini traffic, and other labs have signalled intent without shipping public detection tools. Even universal adoption would not make the problem go away: watermarks only cover models that implement them, open-weight models can run without them, and paraphrasing degrades the signal.

The realistic future is not a reliable button that says "AI wrote this." It is a world where positive watermark hits are meaningful, negatives stay ambiguous, and anything consequential still rests on process and conversation rather than on a score.