How to Prove You Actually Wrote It

A student hands in an essay she wrote herself, start to finish. A detector flags it as AI. She has no way to prove her innocence, and the accusation alone is enough to derail her.

This is not hypothetical, and it is becoming one of the defining problems of writing in the age of generative AI. The question used to be “is this any good?” Increasingly it is “did a human write this?” And the tools we are using to answer that question are, frankly, not up to the job.

Detection is broken, and the people who built it know

Start with the most telling fact. In July 2023 OpenAI, the company behind ChatGPT, quietly shut down its own AI-detection tool. The reason it gave was “low accuracy.” Internal figures suggested the classifier correctly identified AI-written text only about 26 percent of the time, while wrongly flagging around 9 percent of human writing as machine-made. The makers of the most famous AI in the world could not reliably detect their own AI, so they stopped pretending they could.

It gets worse, and less fair. In a peer-reviewed paper in the journal Patterns, Stanford researchers led by James Zou tested several popular detectors on essays by native and non-native English writers. The detectors were near-perfect on writing by native speakers. They misclassified roughly 61 percent of essays by non-native English speakers as AI-generated.

The reason is bleakly mechanical. These tools often work by measuring “perplexity,” a rough proxy for how surprising or varied the word choices are. Non-native writers tend to use more common words and simpler structures, which scores as low perplexity, which the detector reads as “machine.” In other words, the technology systematically punishes people for writing in clear, plain English in a second language. The same study showed the detectors were trivial to fool from the other direction: ask the AI to rewrite with fancier vocabulary and it sails straight past them.

You cannot prove you wrote something by running it through a detector, and you certainly should not be judged by one.

This is not a wrinkle to be ironed out in the next version. A 2024 analysis bluntly titled “Can AI-Generated Text Be Reliably Detected?” argued that as language models keep improving, reliable detection may be fundamentally, not just temporarily, out of reach. Major institutions have started acting on that reality. UCLA and other University of California campuses declined to switch on Turnitin’s AI-detection feature, citing concerns about accuracy and false positives.

The uncomfortable conclusion: you cannot prove you wrote something by running it through a detector, and you certainly should not be judged by one.

The world is quietly switching from detection to provenance

While detection flails, a different and far more promising idea has been gathering pace across the rest of the media world: provenance.

The principle is simple. Instead of trying to sniff out fakes after the fact, you record where a piece of content came from at the moment it is created, and carry that record with it. Do not detect. Document.

The major effort here is the Coalition for Content Provenance and Authenticity, or C2PA, founded in 2021 by Adobe, Microsoft, the BBC, Intel and Arm, and now backed by a coalition of more than 6,000 organisations. Its user-facing form is called Content Credentials, often described as a nutrition label for digital content: a tamper-evident record attached to a file showing how it was made and edited.

This is no longer theoretical. Leica shipped the first camera with Content Credentials built in back in 2023. Google’s Pixel 10 signs photos at the hardware level. Samsung’s Galaxy S25 attaches credentials to AI-edited images. OpenAI’s own image generator embeds them. The United States Cybersecurity and Infrastructure Security Agency endorsed the approach in a January 2025 advisory, and regulation is now accelerating it: the EU AI Act’s transparency rules and California’s SB 942 both push toward machine-readable disclosure of how content was produced.

There is an honest limit worth stating, because it matters for writers. Provenance proves that a claim was made about a file’s history, not that the claim is true, and platforms often strip the data out when content is shared. It is a chain of custody, not a lie detector. But a chain of custody is exactly what writing has never had.

Writing has no chain of custody. That is the gap.

Here is the strange thing. A photograph taken on a recent phone can now travel with a verifiable record of its origin. A 90,000 word manuscript arrives as a bare file with nothing attached. No history, no record of how it came into being. Just text, which is precisely the kind of thing a machine can now generate in seconds.

Yet writing has something photography does not: it is made over time, in passes, with a visible process. A real manuscript has a history. It was written across dozens of sessions. It has false starts, deleted chapters, a word count that climbed unevenly, paragraphs rewritten five times at two in the morning. AI-generated text, by contrast, tends to arrive fully formed, all at once, with no struggle behind it.

That process is evidence. A record of how a piece was actually written, the sessions, the time spent, the way it grew and was revised, is the writer’s natural equivalent of provenance. It does not try to detect anything. It documents the act of writing, so that when you need to show your work, you can.

What an Authorship Passport does

This is the idea behind one of Tagore’s more unusual features, the Authorship Passport.

As you write on the device, it keeps a record of the writing itself: when you wrote, how long your sessions ran, how the word count progressed, the shape of your revisions, and whether large blocks of text were pasted in from elsewhere rather than typed. It is a log of the process, not just the product.

Two design choices matter. First, it stays local and under your control. It is your record of your work, and you decide if and when to share it, for instance with an editor, a supervisor, or a competition that wants assurance the work is genuinely yours. Second, it documents human effort rather than trying to police it. It is not a detector passing judgment on a finished file. It is a receipt for the work you actually did.

That fits the direction the wider world is already moving: away from unreliable, biased, after-the-fact detection, and toward transparent records of how something was made. For photographs that shift is well underway. For writing it is only beginning, and a writer who can show the full history of a manuscript holds something far more convincing than a detector’s coin-flip verdict.

The future of proving you wrote it will not be a tool that scans the finished text and guesses. It will be the visible, documented trail of you having written it. The work, and the proof, will be the same thing.