“I wrote this myself, but it was flagged as AI-written.” You see this complaint often. Students on assignments, applicants on their submissions, hobbyists writing fiction: no one wants to watch people accused of something they did not do.

The usual response to these errors is to ask how detection can be made more accurate. Core takes a different approach: instead of judging the finished text, it records the process of writing itself. This post explains the thinking behind that choice.

The limit of looking only at the finished text

The trouble is trying to satisfy that wish by judging the finished draft. That approach has a limit built into it.

There is research that shows the limit. In 2023, Stanford researchers published a study in which English essays by non-native speakers, all written by humans, were run through seven commercial AI detectors, and 61% were misclassified as AI-generated. One detector put that figure at about 98%. Essays by native English speakers, by contrast, were misflagged only about 5% of the time.

And what happens once people start learning to write from AI? If a human learns from AI and produces text much like an AI’s, there is no reliable way to tell the two apart from the finished text alone. Statistical detection guesses from quirks of style and vocabulary, and as long as the finished product is the only clue, there will always be cases that cannot be judged even in principle.

What remains when a person actually writes

So the object of inspection has to change. Look not at the product, but at the process.

The same sentence can be pasted in all at once, or rewritten many times, cut along the way, and built up over time. That difference exists only while the writing is happening. You cannot see it in the finished text; it survives only if the process is recorded.

The only thing that can keep this record, this provenance, is the tool used to write. Having the tool that produced the text attest that a human wrote it makes more sense than appraising the finished product after the fact.

Anticipated objections

A provenance-based approach invites a few objections.

First: could the record, or the certificate attesting that a person wrote the text, be forged? Features built into the writing tool, such as signatures and timestamps, can prevent that.

Second: if someone copies an AI’s output by hand, could they fake a human-looking history? But mere transcription and genuine composition leave different traces. Copying runs almost linearly, while writing your own text always leaves revisions, deletions, and a back-and-forth behind. Provenance can read that difference and judge on it. This is the decisive difference from judging the finished product alone.

Make provenance part of writing

The current confusion over “did a person write this, or did an AI?” comes from trying to judge text that has already been written.

If recording provenance becomes an utterly ordinary feature of writing tools, problems like these disappear. When a question comes up, you just check the record.

Stop trying to guess whether an AI wrote something, and let people keep their own evidence that they did. In Core, this idea takes the form of a feature that records your writing’s provenance, built into the app.