There is no shameful level.
There is only an undeclared one.

Preamble

We are not here to slow this technology down.

We are here because a person who used it honestly, and said so, is punished for saying so — and the person who used it and stayed quiet is not. That is a broken incentive, and broken incentives do not resolve themselves. They compound.

Every fight about AI and authorship currently runs through a label with two settings. That label is the problem. Not the models. Not the people using them. The label.

Here is what it costs, and here is what replaces it.

AI is not the problem. Hiding is the problem.

Nothing in this document asks anyone to use less AI. It asks them to say what they did.

The audience already has a dial. The creator still has only a switch.

TikTok lets you choose how much AI-generated content appears in your feed.¹ Pinterest lets you ask for less of it.² The people consuming the work are being handed a gradient. The person who made it gets one checkbox: guilty, or not guilty.

”Made with AI” is not a fact. It is a verdict.

It collapses the surgeon who dictated thirty years of practice into a model and corrected every line, and the script that emitted ten thousand pages last night while its owner slept, into the same three words. A label that cannot tell those two apart is not information. It is an accusation with a spellchecker attached.

We built a binary where reality is a spectrum, and we attached shame to one side of it.

Every failure that follows comes from that single design error.

When honesty is punished and silence is free, silence wins.

This is not a moral failure of creators. It is arithmetic.

It has been measured, and it has a name: the disclosure paradox. In a pre-registered study, people said disclosure of AI use was important — and then rated the work lower when it was disclosed. The authors’ own conclusion: this “risks creating perverse incentives for non-disclosure.”³

We are running an experiment in which we punish the truthful and reward the silent, and then we express surprise at the results.

The penalty is not about quality. It is about effort.

Told a short story was written by a human, readers estimated it took 148 minutes. Told the identical story was written by AI, they estimated six. The label did not change how good they thought it was — not its creativity, not its originality, not their enjoyment. It changed only how much they believed it had cost. And that estimate of cost was what predicted everything else.⁴

This is the whole finding, and it is the reason this scale exists. A switch cannot communicate effort. A scale can. It may be the only form of disclosure that does not punish the person disclosing.

Mark only the machine, and everything unmarked starts to look human.

Flag some of the false headlines and the unflagged ones become more believable — an effect established in Management Science and named the implied truth effect. The fix the same researchers found: also verify the true ones.⁵

So a system that labels only AI makes every unlabelled thing — including all the AI it missed — read as human by default.

This is why the scale starts at zero. The people who use no AI at all need a number too. Not as a courtesy. As load-bearing structure.

There is no shameful level. There is only an undeclared one.

Level 5 is the honest declaration for an automated market report. Level 0 is the honest declaration for a memoir. Neither outranks the other.

A scale that ranks its own levels is a shame ladder wearing a lab coat, and every user will lie their way down it. The moment Level 4 becomes an insult, everyone becomes a 2, and we have rebuilt the binary with extra steps.

Provenance can be proven. Contribution can only be declared.

The cryptography is real, and it is not enough.

C2PA can attach a tamper-evident, cryptographically signed history to an asset. Its own FAQ states that the core specification “does not support attribution of content to individuals or organizations.”⁶ Anyone can implement the open specification, but entry into C2PA’s official trust model requires a conforming product and a signing certificate rooted in its trust list.⁷ Its main deployments concern media assets and documents, not ordinary web prose.

It answers what touched this. It cannot answer whose thinking is inside this. Nothing can, except the person who knows.

A declaration is not a weak form of proof. It is a different thing entirely.

A byline is a declaration. A nutrition label is a declaration. A conflict-of-interest statement at the end of a paper is a declaration. None of them are proofs, and civilisation runs on them anyway.

They work because they are cheap to make and expensive to break.

Detection is not the backstop, and it never was.

Seven commercial AI detectors flagged 61% of genuine university-entrance essays written by non-native English speakers as machine-generated. Ninety-eight per cent were flagged by at least one.⁸

A standard enforced by detection is a machine for accusing the innocent: the immigrant, the dyslexic, the person who simply writes plainly. Any system that needs a detector to work does not work.

Disclosure is not a confession. It is a credit line.

Printers have signed colophons for five hundred years — the typeface, the paper, the press, the print run. Films run their credits to the last runner. A cabinetmaker signs the underside of the drawer.

Nobody has ever been ashamed of the credits. The tools were never the secret.

The cost of hiding compounds, and it is not paid by the person hiding.

It is paid by the honest creator nobody believes any more. By the company accused of something it did not do. By the reader who has started to assume everything is fake and is, increasingly, correct.

Merriam-Webster made “slop” its word of the year for 2025: “digital content of low quality that is produced usually in quantity by means of artificial intelligence.”⁹ That is the reputation now attaching to all of it, indiscriminately — to the careless and the careful alike.

The organisations making the content already recognise the stakes. In a 2026 survey of 27 multinational brands, 82% said transparency about AI was essential to brand reputation and 79% to consumer trust. Yet the same research found fragmented rules and uncertainty about expectations.¹⁰ That uncertainty is not an excuse for silence. It is the reason a shared vocabulary is useful.

There is another debt accumulating underneath the visible one. Models generate content; that content is scraped into later training sets; later models reproduce a narrower version of it; and the cycle repeats. Research in Nature calls the failure mode model collapse: indiscriminate recursive training on generated data can erase the tails of the original distribution and compound errors across generations.¹¹ Synthetic data is not inherently bad, and careful mixtures can remain useful. The danger is losing the ability to tell what kind of material entered the corpus.

A declaration cannot decide whether a crawler is allowed to train on a work — licences, terms and access controls do that. It can give model builders a missing signal: whether the material was human-made, AI-assisted, directed, prompted or published without review. Preserving that distinction is not only courtesy to readers. It helps preserve the diversity of the data future models learn from.

The law’s answer to nuance is an exemption. Ours is a scale.

From 2 August 2026, Article 50 of the EU AI Act requires AI-generated text published to inform the public on matters of public interest to be disclosed — unless it was reviewed by a human who holds editorial responsibility, in which case no disclosure is required at all.¹²

Read that again. The law can see the difference between reviewed and unreviewed work. It just has no vocabulary to express it, so it resolves the nuance by switching the obligation off.

The distinction the law reaches for and cannot name is the distinction between Level 4 and Level 5. We are naming it.

Transparency will feel strange for about a year, and then it will feel like nothing.

Nobody today can imagine a food package without a panel on the back, and nobody is ashamed of the calories printed on it. The label did not kill the food.

It ended the guessing.

What we are asking

Declare your level. Put it on the work. Link it to the definition.

That is all. It is free, it takes thirty seconds, and there is no committee to ask.

The scale is six levels wide and starts at zero. It measures the role AI played in the making — whose substance the work carries, and who stands behind it — not how many characters a model emitted. It is CC0. It is not owned. Fork it if we got it wrong.

If enough of us do this before the habit of hiding sets hard, disclosure stops being a confession and becomes what it always should have been: a line in the credits.

This document declares its own level

This manifesto is Level 3 — Directed.

The diagnosis, the argument, the decision to build this, and every design choice in the scale are the author’s. The research and the prose were produced with a large language model, then read, corrected, and signed line by line. Without the author, this document does not exist. Without the model, it exists — slower, and worse written.

That is exactly the case this manifesto defends. It would be absurd to make it and hide it.

The translations are machine-produced from this English text and are marked as such, under the rule in § Translation.

Sources