AI transparency should identify the machine, not give its provider a claim over human-led work.

Imagine spending a substantial amount of time building a successful software company.

You identify the market problem. You define the product vision, customer workflows, commercial model and strategic direction.

Your engineering team determines the architecture, data model, security requirements, integrations, business rules and performance expectations.

Throughout development, the team uses an AI coding assistant.

The AI helps compare architecture alternatives, generate boilerplate code, create tests, identify defects, propose refactoring, explain unfamiliar libraries and suggest possible implementations.

Some suggestions are accepted.

Some are extensively modified.

Others are rejected because they are insecure, inefficient or inconsistent with the product.

The human team integrates the components, reviews the pull requests, tests the application, resolves the trade-offs and accepts responsibility for the software delivered to customers.

The result is human-led and AI-assisted.

Now imagine that a competitor copies substantial portions of the application’s source code.

Your company takes legal action.

The competitor produces evidence showing that an AI system participated in generating or processing parts of the software. That evidence could come from coding-assistant records, repository integrations, provider logs, commit metadata, content credentials or another form of AI provenance.

The competitor argues:

“The company cannot prove that its employees personally wrote these sections.”

“An AI system generated substantial parts of the implementation.”

“Those parts should receive less – or no – copyright protection.”

“The company’s enforceable rights should therefore be narrower.”

Could that argument succeed under current law?

Not automatically.

Source code is presently outside the EU AI Act’s Article 50 marking obligation, and Anthropic says its statistical text watermark has only limited application to executable code because code often requires exact token choices. It may appear in discretionary elements such as comments, but Anthropic says its effect on actual code should generally be negligible. (Anthropic)

But today’s technical limitations are not the central issue.

The larger question is:

Could evidence that AI participated in creating a commercial asset one day be used to weaken the human owner’s rights – or support an economic claim by someone else?

That possibility does not represent current law.

It is not an accusation that Anthropic has a hidden plan.

It is a precautionary question about what future providers, successors, courts, regulators or legislatures could decide after AI provenance becomes embedded across our documents, research, software and business systems.

The right moment to establish the legal boundary is not after the first ownership claim.

It is now.

This is a precautionary argument, not an accusation

Anthropic currently states clearly that its watermark does not establish ownership or authorship and does not change users’ rights under its terms.

That statement should be taken at face value. (Anthropic)

But no current company statement can permanently bind:

• every future owner of that company;
• every successor or competitor;
• every future business model;
• every future version of commercial terms;
• every court;
• every regulator;
• or every legislature.

The concern is therefore not:

“Anthropic intends to take part of our copyright.”

There is no public evidence supporting that allegation.

The concern is:

Without a clear legal restriction, a future AI provider, industry group, court or legislature could attempt to attach ownership, licensing, royalty or revenue rights to evidence that an AI system contributed to a commercially valuable work.

That outcome is not inevitable.

But no one can guarantee that it will remain legally impossible forever.

Preventive legislation does not require proof that someone is already planning to cause harm. Its purpose is to make an undesirable outcome unavailable before the infrastructure capable of supporting it becomes widespread and difficult to reverse.

What the European Union is trying to achieve

The EU is not introducing AI transparency rules to give technology companies copyright.

Article 50 of the EU AI Act is designed to help people recognise when they are interacting with AI or viewing AI-generated or manipulated content.

The European Commission identifies legitimate risks: misinformation, manipulation at scale, fraud, impersonation and consumer deception. The objective is to protect trust and integrity in the information environment and help people make informed decisions about the content they encounter. The relevant Article 50 obligations became applicable on 2 August 2026. (Digital Strategy)

Providers of qualifying generative AI systems must make generated or manipulated audio, image, video and text content machine-readable and detectable as artificially generated or manipulated.

Deployers must also disclose particular uses, including deepfakes and certain AI-generated text published to inform the public about matters of public interest.

The regulation contains important limits.

Current Commission guidance places source code, short sequences of symbols, machine-to-machine outputs and certain closed-loop industrial or product-development outputs outside the provider-level marking obligation. It also recognises standard-editing exceptions and a narrow exemption for qualifying business-to-business and industrial contexts. (Digital Strategy)

For public-interest text, meaningful human review or editorial control can remove the deployer-facing labelling requirement. The review must be substantive – not merely a spelling or grammar check – and a person or organisation must accept editorial responsibility. (Digital Strategy)

These distinctions show that the EU is attempting to regulate proportionately.

The objective deserves support.

People should be protected from synthetic impersonation, fraudulent media, undisclosed deepfakes and mass-produced manipulation.

But supporting transparency does not mean accepting every future use of the infrastructure built to deliver it.

The relevant question is not only:

Why is AI provenance being introduced today?

It is also:

What might governments, courts and companies decide that it means tomorrow?

What Anthropic’s watermark actually proves

Anthropic announced on 14 August 2026 that future Claude models would generate text containing a statistical watermark to comply with the EU AI Act.

The watermark is not a hidden sentence, an invisible character or an account identifier.

It works by influencing some of the choices Claude makes between otherwise reasonable words. Across a sufficiently long passage, those choices produce a statistical pattern that can be checked by someone holding the corresponding key.

Nothing is visibly added to the text, and Anthropic says the watermark contains no information identifying a user, organisation or conversation. (Anthropic)

The technology has significant limitations.

Anthropic says a detector can estimate the likelihood that Claude participated in producing or processing text, but it cannot distinguish:

• Claude writing the original text;
• Claude heavily rewriting human material;
• Claude substantially editing an existing document;
• or a longer collaborative process between a person and Claude.

Detection is weaker for short passages, factual material, proofreading and content where the model has few reasonable alternatives. Light editing may leave the signal detectable, while a complete rewrite can remove it. (Anthropic)

That means a watermark can support one limited conclusion:

Claude was probably involved at some point.

It cannot determine:

• who originated the underlying idea;
• who supplied the experience or domain knowledge;
• who selected the objective;
• who made the critical decisions;
• who rejected the weak alternatives;
• who verified the claims;
• who integrated the result;
• who assumed the commercial risk;
• or who accepted responsibility for the final work.

A watermark records technological participation.

It does not reconstruct intellectual history.

Today’s technical limitations are not tomorrow’s legal safeguards

The fact that current text watermarking is imperfect does not eliminate the governance question.

It does, however, create an important counterargument:

If the signal is probabilistic, removable and weak in some forms of content, why would a court or regulator rely on it?

They should not rely on it alone.

That is one of the central protections this article proposes.

Current technical weaknesses make a watermark a poor basis for deciding authorship or ownership. But institutions routinely make decisions using incomplete or probabilistic evidence – especially when that evidence is combined with logs, metadata, repository history, provider records or other signals.

Future provenance technology may also develop beyond today’s statistical text watermark.

It may include:

• cryptographically signed content credentials;
• development-platform records;
• coding-assistant logs;
• repository integrations;
• generation histories;
• document revision chains;
• cross-provider provenance;
• or new marking methods that work across a wider range of content.

Whether future systems become substantially more robust remains an open engineering question. We should not assume that every technical weakness will disappear.

But we should not assume that nothing will improve either.

The correct conclusion is:

Technical limitations reduce the reliability of today’s signal. They do not permanently limit how tomorrow’s provenance infrastructure may work – or what institutions may try to infer from it.

The wider risk is purpose expansion

Infrastructure introduced for one purpose can later be used for another.

European institutions already recognise this governance problem through the concept of function creep: a system’s use gradually expands beyond the purpose for which it was originally created.

Eurodac provides a useful precedent.

The European fingerprint system was established to help determine which Member State was responsible for examining an asylum application. Its scope was later expanded to provide law-enforcement authorities with access under specified conditions. (Eur-Lex)

This does not prove that the expansion was illegitimate.

Nor does it prove that AI provenance will follow the same path.

It demonstrates a narrower and more important principle:

Once infrastructure and data exist, new actors often identify additional uses for them.

The original purpose of AI watermarking is transparency.

Possible future uses could include:

• evaluating whether a student worked independently;
• determining whether a researcher qualifies as an author;
• assessing whether a startup owns its software;
• identifying which parts of a commercial work are copyrightable;
• allocating licensing obligations;
• calculating a machine-generated percentage;
• or supporting a claim for royalties or revenue participation.

These possibilities are currently hypothetical.

But purpose limitation exists precisely because a legitimate original purpose does not automatically justify every later use.

Provenance can also protect creators

A balanced policy should not treat all provenance as dangerous.

Properly designed provenance can help creators, entrepreneurs and businesses.

The C2PA content-provenance standard, for example, is designed to help creators, publishers and consumers establish the origin and editing history of digital content. Its principles explicitly state that provenance systems should verify assertions rather than judge whether the content is “good” or “bad,” and that creators should retain control over what information is included. (c2pa.org)

That principle could be extended to AI-assisted work.

Imagine a founder’s development record showing:

• the original product requirements;
• the founder’s initial architecture;
• questions submitted to the AI;
• alternatives proposed by the AI;
• suggestions rejected by the team;
• security corrections made by engineers;
• human-authored revisions;
• pull-request reviews;
• tests;
• and final approval decisions.

Such provenance could demonstrate that AI was used responsibly while the human team remained in control.

It could protect the company against a competitor claiming:

“The AI created everything.”

It could help a researcher demonstrate the evolution of a hypothesis.

It could help a student show that an argument was independently developed even though AI assisted with language or criticism.

The goal should therefore not be to eliminate provenance.

It should be to ensure that provenance supports transparency without becoming an automatic verdict against the human.

Human Intellectual Leadership

The most important question is not:

Who generated the largest number of words or lines of code?

It is:

Who intellectually led the creation?

I propose evaluating this through seven factors.

1. Initiation

Who initiated the project and identified the problem worth solving?

2. Purpose

Who determined what the product, research, strategy or creative work was intended to achieve?

3. Context

Who supplied the domain knowledge, personal experience, private information, values, commercial constraints and real-world understanding?

4. Direction

Who guided the process, asked the important questions and decided which paths should be explored?

5. Judgment

Who assessed the AI’s suggestions, rejected unsuitable alternatives and determined what was accurate, secure, useful or appropriate?

6. Integration

Who combined the ideas, code, evidence and recommendations into a coherent final result?

7. Responsibility

Who approved the work, released or published it and accepted responsibility for its consequences?

AI may make a substantial contribution within this process.

It may identify an important risk.

It may propose an architecture that becomes central to the product.

It may challenge a belief and cause the user to change direction.

It may generate most of the literal prose or a significant amount of the implementation code.

Acknowledging that contribution is intellectually honest.

But AI contribution and AI-provider entitlement are different questions.

Where a person or human-led organisation exercises the dominant leadership across these seven factors, that person or organisation should remain the primary rights holder.

This approach is consistent with the broader direction of current human-authorship analysis. The U.S. Copyright Office says AI-assisted creation does not automatically prevent copyright protection and focuses on whether humans determined sufficient expressive elements, made creative arrangements or meaningfully modified the result. (U.S. Copyright Office)

Human Intellectual Leadership is broader than that existing copyright test.

It is intended as a practical framework for future AI-assisted business, research and creative work.

Imagine an entrepreneur developing a strategy

Consider an entrepreneur developing a strategy for rebuilding European manufacturing.

The entrepreneur identifies the problem, establishes the political and commercial objectives and supplies the initial beliefs, values and practical experience.

AI then helps compare industrial policies, analyse supply-chain dependencies, challenge assumptions, investigate energy constraints, propose financing models and connect manufacturing with defence, education, robotics and AI.

Some of the entrepreneur’s beliefs change because of the interaction.

The final strategy is better because AI participated.

But the entrepreneur remains the main driver if he decides:

• what problem matters;
• which evidence is credible;
• which recommendations are realistic;
• which trade-offs are acceptable;
• and what final position he is prepared to defend.

The work may contain a strong AI watermark because the AI generated much of the final language.

That does not establish that the AI provider originated or owns the strategy.

It establishes only that the system participated in developing or expressing it.

Imagine a researcher testing a hypothesis

A researcher observes an unexpected pattern and develops a hypothesis.

AI helps identify alternative explanations, compare statistical methods, challenge the experimental design and examine relevant literature.

An AI-generated question may expose a serious weakness in the methodology.

The researcher corrects the problem, and the final paper becomes stronger.

AI has contributed materially.

But the researcher still designed the experiment, supplied the data, evaluated the suggestions, interpreted the results and accepted professional responsibility for the conclusion.

A watermark cannot determine that division of labour.

Educational and scientific institutions do need to address genuine misconduct, including students submitting fully AI-generated assignments as their own.

But a binary AI signal is not enough to prove misconduct.

Institutions should evaluate drafts, version history, oral understanding, permitted-assistance rules and evidence of the actual development process.

Different sectors have different stakes

AI transparency does not have equal value in every context.

Education, academia and science

These areas may face the highest personal consequences because the individual’s intellectual contribution is itself being evaluated.

A disputed provenance signal could affect:

• grades;
• qualifications;
• publication;
• research funding;
• professional reputation;
• disciplinary proceedings;
• and future employment.

The key questions are not simply whether AI was used, but whether its use was permitted, whether material assistance was disclosed, who developed the hypothesis or argument and who accepts responsibility for the work.

Private and internal business work

For much routine business activity, independent sentence-level authorship is not the primary objective.

A product manager, engineer, executive or salesperson is normally evaluated on whether the work is accurate, useful, commercially sound and responsibly reviewed.

For internal reports, strategy notes, product requirements, technical documentation, client drafts, meeting summaries and operational procedures, the important questions are usually:

• Is the work correct?
• Is it confidential?
• Has a responsible person reviewed it?
• Does it comply with law and company policy?
• Does it produce the intended outcome?

Whether every sentence was typed independently by a human often has little relevance.

For routine internal business content, an invisible watermark may therefore provide limited practical benefit while creating uncertainty over authorship, ownership and confidentiality.

Public-facing and high-stakes business communication

Business communication is different when it affects the public or carries regulated consequences.

Transparency can have real value for investor reports, health and safety claims, sustainability claims, public product warnings, executive impersonation, synthetic customer testimonials, fraudulent audio or video and AI-generated public statements on matters of economic or social importance.

The distinction should therefore not be:

Education needs transparency; business does not.

The better distinction is:

Transparency has high value where content is used to assess human capability or where synthetic content could deceive the public.

It has much less value for ordinary private and internal business work where a responsible person remains accountable for the result.

The legal rule should be simple

A watermark should be evidence of technological participation, not a verdict on legal authorship.

That principle matters because the evidence chain may become more powerful over time. A future system may combine watermarks, content credentials, provider logs, document histories and repository records. Those signals may be useful for transparency. They may even help show that a person used AI responsibly.

But they should not automatically decide ownership.

The law should make five points clear.

1. Purpose limitation

AI provenance introduced for transparency should not automatically determine authorship, copyright ownership, licensing rights, royalties or revenue participation.

2. No automatic provider rights

Using an AI system should not, by itself, create copyright, co-ownership, royalty, licensing or revenue-sharing rights for the provider.

3. Protect Human Intellectual Leadership

Human contribution should be evaluated through initiation, purpose, context, direction, judgment, integration and responsibility – not by word count, line count or a binary AI signal.

4. Provenance must be contestable and user-accessible

Users should be able to access relevant evidence of their own contribution and challenge automated or institutional conclusions drawn from provenance signals.

5. No retrospective or hidden rights

No future law or provider term should retrospectively attach ownership consequences to previous AI-assisted work. Any co-ownership, royalty or revenue-sharing arrangement should require a clear, specific and voluntary agreement.

What this could look like in law

A suitable rule could state:

Evidence that an artificial intelligence system generated, transformed, edited, suggested or otherwise contributed to a work shall not, by itself, establish authorship, ownership, co-ownership, licensing entitlement, royalty entitlement or revenue participation in favour of the provider, operator or developer of that system.

Such evidence may be considered only for the limited purpose for which it was collected or required, including transparency, disclosure, security, fraud prevention or compliance, and shall not create a presumption against human authorship or ownership where a person or human-led organisation exercised substantial intellectual leadership over the creation, selection, arrangement, verification, integration or approval of the work.

The exact drafting can vary.

The principle should not.

The boundary should be set before the dispute arrives

AI watermarking is not inherently hostile to creators.

AI provenance can support authenticity, protect the public from deception and help responsible users demonstrate how work was developed.

But the meaning of provenance must be legally limited.

AI participation is a technological fact.

Provider ownership would be a legal and political choice.

One does not automatically follow from the other.

The right rule is not anti-AI and not anti-transparency.

It is pro-human leadership.

When people and human-led organisations initiate the work, define the purpose, supply the context, guide the process, exercise judgment, integrate the result and accept responsibility, they should not lose control merely because an AI system helped them think, write, code or decide.

Transparency should identify the machine.

It should not transfer the work.