Academic publishing runs on a question it has mostly stopped being able to answer. Who wrote this?
For most of its history the answer was implicit in the process. A paper was typed by its authors, revised on paper, and submitted. Authorship was still a claim, but it was a claim with friction behind it.
That friction is gone. A competent language model produces a methods section in seconds, and nothing in the resulting file distinguishes it from one a graduate student spent an afternoon on.
Disclosure statements are not evidence
Journals and funders responded with the disclosure statement. A box you tick, a paragraph describing which AI tools you used and how. Better than nothing, and it establishes a norm worth having.
But look at what a disclosure statement actually is. An unverifiable self report, made after the fact, by the person with the strongest incentive to understate.
There is no mechanism to check it. A reviewer who suspects more AI involvement than declared has no way to find out. An author who was scrupulously honest has no way to prove it, which is the part people usually miss. The current system fails careful authors just as badly as it fails dishonest ones.
The same community would never accept this for data. "Trust me, the experiment worked" is not a result. You deposit the data and someone can check. For authorship we are still at trust me.
Why detection tools are worse
AI text classifiers are unreliable in both directions, and one of those directions does real damage. Non native English speakers are disproportionately flagged, because the statistical markers of "AI-like" text overlap heavily with the markers of careful, slightly formal second language prose.
An integrity regime built on a classifier that systematically suspects international students is not one anyone should want to build.
Record it while it happens
Every approach above tries to reconstruct authorship after the fact, from the finished artifact. By then the information is gone.
Provenance inverts that. Record authorship while the document is being written, when the information still exists and costs nothing to capture. The editor already knows who typed what. It knows which passages came from an accepted suggestion and which the author wrote directly. Nothing has to be inferred. The record just has to be kept, and made hard to forge.
In TeXposit every project accumulates a signed authorship record as you write. You can export it as a provenance report showing the composition of the text, which contributors wrote which parts, and how much was AI assisted, signed so that any modification invalidates the signature. A reader verifies the report against the original without taking your word for anything.
The useful property is that an attestation becomes an artifact. "I declare that AI was used only for language editing" turns into something a third party can check.
Four things provenance is not
Provenance is easy to oversell, so here are its limits stated plainly.
It does not measure quality. A fully human written paper can be bad and an AI assisted one can be excellent. Provenance describes composition, not worth, and anyone treating it as a proxy for rigour is misusing it.
It does not establish intellectual contribution. Typing is not thinking. A supervisor who shaped the entire argument may have typed almost none of it. Provenance records text and its origin, which is genuinely different from authorship in the sense the byline means, and it must never become the definition of who counts as an author.
It only covers text written in the editor. Paste in a paragraph drafted somewhere else and the record shows a paste, not a source. This narrows the gap rather than closing it.
It is not surveillance. The record describes the document, not the author's productivity. It is not a tool for auditing how fast someone writes, and we would consider it a failure if it were used that way.
Why this becomes infrastructure
The trajectory is not hard to read. AI assistance in writing is not going to decrease, and tolerance for unverifiable self reports about it is not going to increase. Somewhere in between, journals and funders start asking for evidence instead of assurances. Not because researchers are dishonest, but because a norm nobody can check erodes under pressure regardless of anyone's intentions.
When that happens, the useful thing is not a better detector. It is documents that arrive already carrying a verifiable account of how they were made.
The reason to build it now is that provenance cannot be applied retroactively. It is recorded during writing or not at all. A paper written today in a tool that keeps no record can never be given a provenance report later, because the information was never captured in the first place.
Every TeXposit project generates a provenance record automatically. You can see an example report, verify a document, or read the provenance guide.