Absence of a watermark is not proof of a human
The single most important thing to understand about AI-content verification — and the reason Provenote leads with provenance, not a detector score.
There is a comforting story about AI detection that goes like this: paste the text into a checker, read the percentage, and you know whether a machine wrote it. It is a clean story, and it is wrong in a way that quietly ruins decisions — student appeals, editorial rejections, hiring calls — every day.
The error is treating a detector’s silence as a verdict. A statistical detector that finds nothing has told you one thing: it did not find its signal. It has not told you a human wrote the text.
A missing watermark is missing evidence. Missing evidence is not counter-evidence.
Why absence is so easy to manufacture
Provider watermarks like Google’s SynthID and Anthropic’s new text watermark are real, verifiable positive signals — when they are present. But they are fragile by design. Paraphrasing, translation, heavy editing, or simply generating with a model that never watermarked in the first place all produce text with no watermark at all. So does a human typing from scratch.
That means "no watermark found" is the single most common result you will ever see, and it is consistent with every possible origin. Building a judgment on it is building on nothing.
Provenance first, statistics second
Provenote is structured around this asymmetry. We check the strong, verifiable signals first — provider watermarks, C2PA content credentials, and any signed provenance the content carries. When one is present, we can say something firm and show you exactly why.
- Tier A — provenance: a present, cryptographically checkable watermark or credential. Firm, positive, auditable.
- Tier B — ledger: a prior record that this exact content was seen and attested before.
- Tier C — statistical detectors: useful as corroboration, capped in weight, and never allowed to "confirm" anything on their own.
When none of the strong signals are present, we say so plainly. The honest output is a calibrated maybe with its uncertainty attached — not a confident number dressed up as proof.
What this changes for you
It changes what you are allowed to conclude. A Provenote report will tell you when it has found real provenance, and it will refuse to manufacture certainty when it hasn’t. That refusal is the product. An automated accusation built on an absent signal is exactly the failure mode we exist to prevent.
This is also why we will never build a watermark remover. The same fragility that makes absence uninformative is what bad actors exploit to erase provenance. Provenote proves how content was made; it never helps hide it. See our Principles.
Sources
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.
Provenote is in private beta. Request early access →