← Provenote blog
compliance principles

Disclosure isn’t verification

Article 50 makes you say you used AI. That is a different thing from being able to prove how something was made — and the difference decides who ends up protected.

Deepak R Chandran, Ph.D.
Deepak R Chandran, Ph.D.
Founder, Provenote · · 6 min read

It is the question we are asked more than any other, usually with a slightly impatient edge: if I simply declare that I used AI, why would I need you?

It is a fair question and it deserves a straight answer. Declaring is not the same act as proving. The two get used interchangeably in policy conversations, and the gap between them decides who ends up protected when something goes wrong.

A declaration is a claim. Verification is a record.

When you disclose that you used AI, you are asking to be believed. That is not a criticism — most disclosure is sincere. But consider what a declaration cannot do:

A verifiable record is a different kind of object. A provider watermark, a content credential, a signed chain of drafts — these can be checked by someone who does not trust you, which is precisely the situation in which trust is actually needed.

Disclosure is a promise about the past. Verification is evidence from it.

Disclosure only binds the people who were never the problem

Here is the asymmetry a disclosure-only regime cannot escape. Disclosure is an act of voluntary honesty. Someone intent on concealing how their content was made does not file a false disclosure — they simply file none at all, and nothing about the text announces the omission.

So a rule built purely on disclosure regulates the honest and leaves the dishonest untouched. Worse, it hands the honest author nothing when they are the one under suspicion. If a detector flags your essay, your own prior statement that you did not use AI is just another unverifiable claim, offered by the person with the most obvious motive to make it. You are asked to prove a negative about your own writing, and a declaration does not help you do it.

This is the part that gets missed. Verification is not primarily a tool for catching people. It is the only thing that lets an honest person answer an accusation with something other than insistence.

Article 50 actually asks for both

The EU AI Act is clearer on this than most commentary about it. Its transparency duties, in force since 2 August 2026, do not treat marking and disclosure as one obligation. They are two, aimed at two different audiences.

A hidden machine-readable mark does not satisfy the second duty, and a visible label does not satisfy the first. One is for machines to read; the other is for a person to see. Reading the marks correctly is the job Provenote does. Making the disclosure is yours — we are compliance-supporting infrastructure, not legal advice.

Neither duty is universal, and that matters as much as the distinction. Which obligation applies turns on the actor, the type and purpose of the content, and the exceptions the Act itself provides. For public-interest text there is an express carve-out where the AI-generated content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for its publication. So a newsroom that edits and stands behind a piece is not in the same position as an unattended pipeline, and we are not going to tell you otherwise in order to make our product sound more necessary than it is.

What a verification result is allowed to say

If verification is going to be the thing that carries evidentiary weight, it has to be disciplined about its own limits, because four different questions hide inside what people call “proving” something. Was the signed record modified? Which signing identity authenticated it? Which events were actually recorded? And who made the creative decisions? A cryptographic check answers the first two. It does not answer the last two, and no amount of valid signature turns into an answer to them.

So a verdict from us resolves to one of four states, and the fourth has to stay available or the other three are worthless:

Absence of verifiable provenance is absence of evidence. It is not evidence of AI use.

A document can lack a watermark or a content credential for entirely ordinary reasons: the tool never wrote one, an export stripped it, the text was pasted out of its container. A system that reads missing evidence as misconduct has simply swapped one unreliable inference for another, and it would be doing it to the same people the detectors already get wrong. Equally, a valid credential does not establish that a human conceived every sentence — an AI-generated document can carry a perfectly valid one. The signature authenticates the record. What the record covers, and whether what it says is true, are separate questions and we report them separately.

When disclosure genuinely is enough

We are not going to pretend everyone needs this. If you are writing a personal blog post, a newsletter to people who already know you, or a social update, and you say a model helped — that is the obligation discharged. Nobody is going to convene a hearing. Buying verification for that would be theatre, and we would rather say so than sell it to you.

The line is not the length of the document. It is whether a decision follows. Does someone accept or reject the work, award or withhold a grade, publish or spike a piece, admit or exclude evidence, pass or fail a procurement check? Once a consequence attaches to the answer, “trust me” stops being an adequate basis — not because the author is dishonest, but because the person deciding cannot tell the difference from where they are standing.

Who verification is really for

Verification is for the receiver: the journal, the university, the platform, the court, the buyer. And it is for the honest author, who gains the ability to demonstrate rather than assert.

Those two look like opposing interests and are not. The same record that lets an institution act fairly is the record that lets a wrongly-accused writer answer. That is why we build for both sides, and why we cap what any statistical detector is allowed to conclude on its own. One honest limit bears repeating: a record only helps if it exists before the question is asked. We cannot manufacture one afterwards, and we will not suggest otherwise.

Declaring is something you do. Verification is something anyone else can do — without needing to take your word for any of it.
Note

Next up: The case we cannot solve — an author drafts with AI and then edits it into their own voice. Nobody can detect that, and we explain exactly why. See also our Methodology.

Sources

  1. EU AI Act — Article 50 (transparency obligations)
  2. European Commission — Quick facts: transparency rules for AI systems
  3. Provenote — Methodology
  4. Provenote — FAQ
Deepak R Chandran, Ph.D.
Deepak R Chandran, Ph.D.
Founder, Provenote

Deepak R Chandran, Ph.D., is the founder of Provenote. He writes about content provenance, AI watermarking, and building verification that is honest about what it can and cannot prove.

Get provenance you can prove

Provenote is in private beta. Request early access →

More reading: Absence of a watermark is not proof of a human →