Anthropic is introducing invisible watermarks and provenance mechanisms for content generated by Claude. The move anticipates an important change: in Europe, AI-generated content will increasingly need to be identifiable and verifiable.
This week, Anthropic announced that text generated by its Claude models will incorporate an invisible watermark, while files produced by the models — such as PNG, JPG or SVG images — will include digitally signed provenance metadata.
This is not an isolated decision or simply a commercial initiative. It is one of the first visible implementations of a European obligation that has recently entered into application and that will directly or indirectly affect those who generate, reuse or publish AI-assisted content.
It is therefore worth understanding exactly what the law requires, what a watermark in text actually means from a technical perspective and, above all, what it does not mean.
The obligation: Article 50 of the AI Act
Article 50(2) of the European Union’s Artificial Intelligence Act imposes a transparency obligation on providers of AI systems that generate synthetic content: they must ensure that outputs are marked in a machine-readable format and detectable as artificially generated or manipulated.
The obligation has applied since 2 August 2026.
To facilitate compliance, the European Commission has promoted a Code of Practice on Transparency of AI-Generated Content, a voluntary framework that provides more detail on how this type of content should be marked and detected.
The regulatory logic is straightforward: as it becomes increasingly difficult to distinguish human-generated content from synthetic content simply by looking at it, the legislator places part of the responsibility at the source.
Generating content using artificial intelligence is not prohibited. Instead, mechanisms must exist that make it possible to identify its provenance.
What is a watermark in text?
In an image, the idea of a watermark is intuitive. In plain text, it is much less obvious.
It does not necessarily mean adding a sentence stating “generated by AI”. Identification can be technically incorporated into the content itself through patterns that are imperceptible to the reader but can subsequently be recognised by systems designed to detect them.
According to the announcement, the watermark does not alter the meaning or apparent quality of the text, can remain with it when copied and pasted, and is designed to withstand a certain degree of subsequent modification.
For files, the solution is different. Digitally signed provenance metadata can be incorporated using open standards such as C2PA, allowing information about the origin and subsequent modification of a file to be verified.
The underlying principle is the same in both cases: content can carry verifiable information about its provenance.
A watermark does not establish who the author is
This point is particularly important from a legal perspective.
The fact that a text contains a watermark identifying the involvement of an artificial intelligence system does not necessarily mean that the AI is the intellectual author of the content.
A person may write a text and subsequently use an AI system to improve its style, translate it, summarise it or change its format. The resulting text may contain an AI provenance signal even though the fundamental intellectual contribution remains human.
Likewise, the absence of a watermark does not necessarily prove that content was created by a human. Some systems may not incorporate these mechanisms, and technical questions remain regarding how well watermarks can survive extensive modification of the content.
It is therefore important to distinguish between two concepts:
Provenance indicates which tools or systems have participated in generating or transforming the content.
Authorship determines who made the legally relevant creative contribution.
They are not necessarily the same thing.
From transparency to digital evidence
The issue may also acquire an important evidentiary dimension.
In disputes involving intellectual property, unfair competition, reputation, fraud or liability for content, the ability to technically determine the origin of a document may become relevant.
Digitally signed metadata may constitute evidence regarding the provenance of a file.
However, it should not automatically be treated as conclusive evidence. The reliability of the system used, the integrity of the metadata, possible subsequent modifications and the digital chain of custody of the content will all need to be considered.
Verifiable provenance may therefore develop into a new category of digital evidence.
Who will be affected?
The change has consequences at several levels.
Companies integrating artificial intelligence models through APIs may find that content generated through their products incorporates provenance identification mechanisms.
Media organisations and publishers will need to establish policies governing AI-generated or AI-assisted content.
Law firms and other professional service providers, whose document production increasingly incorporates artificial intelligence tools, should review their internal policies, confidentiality obligations and criteria for disclosing the use of these technologies.
Regulators and courts will also be affected, as verifiable content provenance begins to become a legal and evidentiary issue.
There is also a broader message: marking content does not make the use of artificial intelligence inherently suspicious.
European regulation is based on transparency, not stigma.
The relevant questions will continue to be who is responsible for the content, who made the intellectual contribution and which legal, contractual, professional or editorial obligations apply.
BACS’s view: verifiable provenance as infrastructure
From the perspective BACS has been advocating regarding the development of digital legal infrastructure, this week’s announcement forms part of a much broader development.
Cryptographically signed provenance metadata follows a logic similar to that of legal oracles applied to digital assets: facts or information issued by an identifiable source, verifiable by third parties and capable of being automatically processed by digital systems.
Article 50 of the AI Act applies this logic to AI-generated content.
Digital legal infrastructure can apply a similar logic to rights, decisions and transactions carried out on blockchain networks.
Content transparency and digital legal infrastructure are therefore not entirely separate issues. They share the same underlying principle: verifiability.
How can BACS help?
The implementation of these obligations creates a new area of compliance and governance for companies that develop, integrate or use artificial intelligence systems.
BACS can provide specialised support in defining the legal architecture required to integrate these technologies securely, particularly in projects combining artificial intelligence, blockchain and digital assets.
Services may include:
- analysis of transparency obligations applicable to AI systems and AI-generated content;
- legal review of internal artificial intelligence policies;
- design of traceability and verifiable provenance systems;
- legal analysis of digital signature, certification and digital evidence mechanisms;
- drafting and review of contractual clauses governing the use, responsibility and provenance of AI-generated content;
- dispute resolution mechanisms for disputes involving AI and digital assets;
- integration of legal oracles and verification systems into blockchain architectures;
- design of governance and digital enforcement mechanisms for projects combining AI, smart contracts and tokenisation.
The issue is no longer simply one of formally complying with a transparency obligation. It is about designing systems capable of demonstrating who generated, modified, verified or authorised specific information.
Transparency becomes part of the architecture
For companies using artificial intelligence, the practical recommendation is clear: identify where AI intervenes in content-production processes, review the terms of the providers being used, establish an internal policy on AI use and transparency, and closely follow the development of verification tools.
Transparency is gradually moving beyond being purely a reputational issue and becoming a technical and legal characteristic of the systems themselves.
During the first years of generative artificial intelligence, the major question was whether we could distinguish content created by a person from content generated by a machine.
The next stage will be different.
It will become increasingly important to verify where content comes from, which systems were involved in its creation and who is legally responsible for it.
Artificial intelligence is beginning to leave a trace.
And that trace is beginning to have legal consequences.
