Prove what was human-reported,
and what AI touched
A story passes through research assistants, summarizers, translation models and headline tools before it reaches a reader. AI-touched is not AI-generated, but that distinction only counts if you can show it. Swiss Trust Layer seals each published version with an AI-content declaration and a qualified timestamp, producing a certificate anyone can check without contacting your newsroom.
Article 50 of the EU AI Act sets transparency obligations for AI-generated and AI-manipulated content. A qualified timestamp carries a legal presumption of date and integrity under eIDAS Art. 41 . A CMS log gives you neither.
Article sealed and verified
Editorial provenance record
One record: the exact published version, its AI-content declaration, and an independent date, verifiable by anyone.
What a publisher has to be able to show
A reported piece runs through more AI touchpoints than most editorial teams track. A research tool summarizes a source document. A grammar tool rewrites two paragraphs. A headline generator proposes six options. A translation model produces the versions a syndication partner receives. By the time the piece is published, the question of whether a person wrote it has no clean answer, only a chain of partial edits nobody logged.
Across a normal publishing week that means you should be able to show:
- Which exact version of a piece went live, distinct from every draft before and after it
- When that version was fixed, from a source outside your own CMS
- Which parts were reported and written by a person, and where AI generated, modified or assisted the text
- That an archive piece carried your text and data mining reservation on the date it was scraped
- Which version a syndication or licensing partner actually received
Each of these turns into a compliance question, a licensing question or a credibility question the moment someone asks. The record you keep at publication decides how the answer sounds.
Where the pressure on newsrooms comes from
Two EU rules ask the same question from opposite directions
Article 50 of the EU AI Act sets transparency obligations for AI-generated and AI-manipulated content, with the deep fake and public-interest text duties under Article 50(4) sitting with the deployer of the AI system. Article 4 of the DSM Directive lets you reserve your works from text and data mining, but a reservation is only worth what you can prove about what was published and when.
When there is no provenance record, the cost lands in three places:
- A disclosure question arrives, from a regulator or a rival outlet, and the honest answer is a reconstruction from memory and server logs rather than a dated record.
- An AI company negotiating a licensing deal, or disputing a scrape, dismisses your CMS timestamps because the party asserting the date and the party who set it are the same.
- A syndication partner strips context from a piece, and you cannot show which version you delivered or what its declaration said.
A qualified timestamp under eIDAS Art. 41 carries a legal presumption of the accuracy of its date and the integrity of the sealed data. That presumption is what turns an internal belief about a piece into evidence a stranger can check.
A CMS log vs a sealed provenance record
The difference is not archiving discipline. It is whether your account of a piece holds when the person asking has no reason to take your word for it.
| What matters to a newsroom | CMS timestamp, internal log | Swiss Trust Layer sealed record |
|---|---|---|
| Fixes the exact published version | Rows anyone with admin access can edit | Cryptographic fingerprint, tamper-evident |
| Independent date | Your own server clock | Qualified timestamp with a legal presumption |
| AI-content declaration | A policy page, separate from the piece | Generated, Modified, Assisted or Human, bound to the version |
| Checkable by outsiders | Requires trusting your systems | Public certificate, no login, no contact needed |
| Evidence for TDM opt-out and licensing | Open to dispute | Dated record of the reservation on the published text |
Declaration first, signature law underneath
The AI-content declaration is what a reader, regulator or licensing partner actually wants to see. It rests on qualified trust services regulated under ZertES SR 943.03 in Switzerland and eIDAS across the EU. See the compliance hub.
How sealing fits an editorial workflow
Swiss Trust Layer does not run its own cryptographic key infrastructure. Every seal and timestamp is backed by Swisscom Trust Services, an accredited qualified trust service provider in Switzerland.
Lock the final version
The moment a piece clears final edit, seal the exact text that goes live. A cryptographic fingerprint fixes it byte for byte, so a later argument about which version ran is resolvable.
Declare the AI involvement
Record whether the piece was Generated, Modified, Assisted or Human at that same moment, so the declaration is bound to the version instead of reconstructed later.
Keep the verification link
Store the certificate with the piece's internal record. A reader, regulator or licensing partner can check the version, the date and the declaration without contacting you.
The sealing step runs once per published version, not once per AI tool used during drafting. A reporter can run a story through five tools while writing it. What needs sealing is the one version that goes out.
Primary sources: EU AI Act, Regulation (EU) 2024/1689 · DSM Directive (EU) 2019/790 · eIDAS Regulation 910/2014 · ZertES SR 943.03 (fedlex.admin.ch) · Berne Convention (wipo.int)
Where sealed provenance pays off across a newsroom
The same record covers the situations an editorial team actually faces, from daily publishing to archive licensing.
Reported journalism
Seal investigative and reported pieces with a Human declaration, so the claim that a person wrote it is a dated, checkable record rather than a line in a style guide.
AI-assisted production
Summaries, translations and headline variants get an Assisted or Modified declaration bound to the exact version, keeping the AI-touched and AI-generated distinction provable.
Syndication and licensing
Fix the version a partner receives, so a stripped byline or an altered paragraph downstream can be measured against the record you hold.
Text and data mining reservations
Back a mining opt-out with a dated record that a specific text carried your reservation, evidence for AI-training licensing talks and scraping disputes.
Corrections and updates
Seal each revision as its own version, so the correction history of a piece is a chain of dated records instead of an editable log.
Visual journalism
Photographs, graphics and video carry the same declaration and timestamp, covering the transparency questions that manipulated media raises.
Backed by Swisscom Trust Services
Accredited qualified trust service provider
Qualified timestamps
Legal presumption of date and integrity
Verifiable by anyone
Public certificate check, no login, no contact
Editorial provenance, common questions
How do we prove an article was written by a journalist and not AI?
Seal the final version at publication with an AI-content declaration set to Human and a qualified timestamp. Under eIDAS Article 41 the timestamp carries a legal presumption of the date and integrity of the data, so the record shows that exact text carried that declaration on that day. Anyone can verify the certificate without contacting your newsroom.
Does the EU AI Act require publishers to label AI content?
Article 50 of the EU AI Act sets transparency obligations for AI-generated and AI-manipulated content. Under Article 50(4), text generated or substantially manipulated by an AI system and published to inform the public on matters of public interest may need to be disclosed, and the obligation sits with the deployer of the AI system. A sealed declaration gives you the dated record behind whatever label you apply.
How do we evidence a text and data mining opt-out under the DSM Directive?
Article 4 of the DSM Directive lets rightsholders expressly reserve their works from text and data mining, in an appropriate manner such as machine-readable means for online content. Sealing the published version with a qualified timestamp gives you a dated, tamper-evident record that a specific text carried your reservation on a specific day, which is the evidence a licensing negotiation or a scraping dispute turns on.
Can we prove which version of an article we licensed or syndicated?
Yes. Each sealed record is bound to one exact version of the text. If a syndication partner alters the piece or a dispute arises over what was delivered, the certificate shows the version you fixed, its date, and its AI-content declaration. Copyright in the work itself is recognised automatically across the 181 member countries of the Berne Convention, and the sealed record is your evidence of it.
Does sealing an article expose the manuscript or our archive?
No. Sealing fixes a cryptographic fingerprint of the file together with the declaration and the qualified timestamp. Verification confirms that fingerprint, the date, and the declaration, without publishing the text. An unpublished draft, an embargoed piece, or an archive you are reserving from text and data mining stays under your control while the proof remains checkable by anyone.
Put a dated record behind every piece
Book a short walkthrough and we will map the sealing step onto your publishing process, from daily pieces to archive reservations.
Working on AI transparency more broadly? AI provenance for the EU AI Act →
Quick Answers
How can a publisher prove an article was human-written?
Seal the final version the moment it clears editing. Swiss Trust Layer fixes the exact text with a cryptographic fingerprint, binds an AI-content declaration to it, and applies a qualified timestamp, which under eIDAS Article 41 carries a legal presumption of the date and integrity of the data. The resulting certificate shows that this exact version carried a Human declaration on that day, and anyone can verify it without contacting the newsroom.
What does the EU AI Act ask of publishers who use AI?
Article 50 of the EU AI Act sets transparency obligations for AI-generated and AI-manipulated content. Text generated or substantially manipulated by an AI system and published to inform the public on matters of public interest may need to be disclosed under Article 50(4), and the obligation sits with the deployer of the AI system. Disclosure is a recordkeeping problem before it is a labeling problem: a newsroom cannot accurately disclose what it did not track.
How does a publisher evidence a text and data mining opt-out?
Article 4 of the DSM Directive allows rightsholders to expressly reserve their works from text and data mining. That reservation is a claim about a specific text on a specific date, so it needs a record fixed at publication rather than a CMS timestamp the publisher controls. A sealed version with a qualified timestamp gives a dated, tamper-evident record that the reservation was in place, which is what licensing negotiations and scraping disputes turn on.