Dr. phil. Leonie Gantenbein

CLEAR Labelling AI Involvement in Texts

Conception · Language · Evidence · Attribution · Responsibility

When it comes to writing, the shadow use of AI is widespread – in literature and research as much as on social media and in private life. We generate a text with AI and humanise it – often in emails, too. And in reply, we receive an equally humanised AI text from a colleague. Along the way, we learn what humanising means and what to watch out for: few dashes, no lists of three, no “both … and”.

But why do we speak of shadow use at all? The word reveals that we regard the use of AI as something better kept hidden. It is usually not forbidden. It is merely considered a flaw.

In the EU, the AI Act ushered in a culture of AI labelling this summer. Since 2 August 2026, deepfakes – deceptively realistic recordings – must be disclosed as artificially generated. This also affects Swiss actors as soon as their content is used in the EU. For texts, the obligation is narrowly defined: an AI-generated text only has to be labelled if it is intended to inform the public on matters of public interest and no human has reviewed it editorially and taken responsibility for it. In the end, the law too asks whether someone stands behind the text.

Public transparency for texts that goes beyond this obligation would be just as important. What if an AI text could simply be labelled, and some people led by example from the start because they are not too proud to say how and where they use AI? And what if this caught on and made it easier for us to assess a text and form an opinion?

Yes, one can lie with a label. But hasn’t that always been the case?

CLEAR is my proposal for disclosing the involvement of artificial intelligence in texts voluntarily and in a differentiated way. Instead of the blanket label “AI-generated”, five short statements record in which stages of the work an AI took part, how far its contribution extends and who stands behind the result. Each person involved makes these statements for their own share.

CLEAR works like a citation: it does not judge; it provides information about how a text came into being.

Responsibility for a statement always lies with whoever makes it, not with the AI. We take responsibility for much of what an AI has contributed, and rightly so, because we are the ones who have to check it. The statement R makes visible whether this check has taken place.

Quick reference

The five statements at a glance:

AxisSubjectLevels
C – ConceptionConception and line of thoughtC0 exclusively human · C1 AI as a discussion partner · C2 individual content parts worked out by the AI · C3 content comprehensively generated by the AI
L – LanguageLinguistic editingL0 without AI · L1 proofreading · L2 stylistic editing · L3 substantive editing (structure and argument) · L4 drafting; suffix (t) for a translation by the AI
E – EvidenceSources and how they were processedE0 without AI · E1 source access · E2 source processing · E3 source interpretation · E4 interpretation of results; suffix (p) published, (n) unpublished material
A – AttributionDeclaration of the toolsAI models used and period of use
R – ResponsibilityAccountabilityR1 fully checked and taken responsibility for · R2 spot-checked · R3 published unchecked, with the risk of having to take responsibility for the text nonetheless

For each axis, give the highest level reached; it includes the levels below. If several people are involved, such as author, editor or translator, each gets a line of their own. The label appears before or after the text.

Where the label goes. It is a one-liner and fits anywhere. Place it where information about how a text came into being is expected, so that no one has to search for it:

  • Book: in the imprint
  • Article: at the end
  • Post and other short texts: also at the end

An example. This is how the text of my book is labelled:

CLEAR (author): C1, L3, E(p)1, A: Claude Opus 4.8, Fable 5, Opus 5, 04–09/2026, R1

The line of thought is mine; the AI was a discussion partner (C1). In the writing, it also intervened in structure and argument (L3) and searched for published sources, which I read myself (E(p)1). Three models were involved between April and September 2026, and I checked every passage and take responsibility for it (R1).

All levels in detail, guidance on classification and the application in schools and universities can be found in the chapter “Specification”. In the chapter “Examples”, twenty invented cases from study, research, work, public life and private life show what labels look like and how to read them. The factsheet at the end of this page summarises all abbreviations and contains a prompt with which a language model proposes its own involvement.

Status

CLEAR is a proposal. It has not yet been tested or evaluated in practice and is certainly still incomplete. That is precisely why I welcome feedback: experience from use, objections and ideas on how CLEAR could be transferred to images, moving images and code. The best way to reach me is via the issues in the GitHub repository.

Origin

CLEAR emerged in my book Zwischen Prompt und Welt. Über das Leben mit KI [Between Prompt and World: On Living with AI], forthcoming on 14 December 2026 from Schwabe Verlag.

CLEAR builds on concepts from my earlier book Annetzung. Handeln durch Technologie und Imagination [Acting through Technology and Imagination] (Schwabe 2025, open access, funded by the Swiss National Science Foundation), above all the concept of world transparency (Welttransparenz): the claim to know whether one is facing something actual, something computer-generated or something imagined.

Citation

If you use CLEAR or develop it further, please cite:

Gantenbein, Leonie (2026): CLEAR – Kennzeichnung von KI-Beteiligung an Texten [CLEAR – Labelling AI Involvement in Texts]. Version 1.0. https://github.com/LeonieGantenbein/clear-kennzeichnung

Factsheet

The factsheet summarises all abbreviations and rules on one page. You can use it for reference or give it to a language model so that it proposes its involvement in a text according to CLEAR.

Notation

CLEAR (role): C?, L?, E(p|n)?, A: models, period, R?

The highest level reached is given for each axis; it includes the levels below. Each person involved gets a line of their own. The label appears before or after the text.

AbbreviationMeaning
C – Conception: conception and line of thought
C0The line of thought was developed exclusively by humans.
C1The AI served as a discussion partner during conception.
C2Individual content parts were worked out by the AI.
C3The content was comprehensively generated by the AI.
L – Language: linguistic editing
L0The language was formulated without AI.
L1Proofreading: spelling, grammar and punctuation.
L2Stylistic editing: wording improved, content unchanged.
L3Substantive editing: interventions in structure and argument.
L4Drafting: the AI writes out the text from content that the human provided or worked out with it.
(t)Translated by the AI. A pure translation already counts as proofreading: L(t)1.
E – Evidence: sources and how they were processed
E0Source access and processing without AI.
E1Source access: the AI finds, obtains or translates sources; the human reads and interprets them.
E2Source processing: the AI summarises sources; work continues with the summaries.
E3Source interpretation: the AI’s interpretation of the sources was adopted.
E4Interpretation of results: a conclusion drawn by the AI appears in the text as a result.
(p)Published material that readers can verify.
(n)Own, unpublished material.
A – Attribution: declaration of the AI models used and the period of their use
R – Responsibility: accountability
R1Fully checked and taken responsibility for by the human.
R2Spot-checked and therefore not fully taken responsibility for.
R3Published unchecked, with the risk of having to take responsibility for the text nonetheless.

Prompt for a language model’s self-declaration

A language model can propose its own involvement. To do so, copy the following text into the same chat at the end of a collaboration. The model then proposes a level for C, L, E and A and gives reasons; R is always set by you.

Please label, according to CLEAR, how you were involved in this text. Base your answer exclusively on our conversation so far.

For each axis, give the highest level reached and justify it in one sentence:

C – Conception: C0 line of thought exclusively human · C1 AI as a discussion partner · C2 individual content parts worked out by the AI · C3 content comprehensively generated by the AI
L – Language: L0 without AI · L1 proofreading · L2 stylistic editing (wording, content unchanged) · L3 substantive editing (structure and argument) · L4 drafting (text written out from content provided or worked out together); suffix (t) if you translated the text (pure translation: L(t)1)
E – Evidence: E0 without AI · E1 source access (finding, obtaining, translating) · E2 source processing (summaries replace the primary source) · E3 source interpretation adopted · E4 the AI’s conclusion adopted as a result; suffix (p) published, (n) own unpublished material
A – Attribution: your model name and the period of our collaboration
R – Responsibility: Leave R open. Only the human can make this statement.

Format: CLEAR (role): C?, L?, E(p|n)?, A: model, period, R?
For L, the deepest involvement counts, not the most frequent. If you cannot determine a level with certainty from the conversation, say so instead of guessing.

Limits of the self-declaration

A model can only assess what it has seen. Its statement is therefore only meaningful if the entire work took place with a single language model in a single chat.

If you used several chats, tools or models, revised the text outside the chat, or if the conversation is so long that the model no longer takes its beginnings into account, it cannot fully assess its involvement. In that case, make the classification yourself.

Even in the best case, the model’s statement remains a proposal that you check and correct where necessary. Only a human can take on responsibility (R).