Anthropic has just introduced something that could become far more consequential for the legal industry than another AI feature: provenance.
The company says it will embed imperceptible, machine-readable watermarks into text generated by Claude. The watermark is designed to survive ordinary copying, pasting, and light editing, with Anthropic developing tools that can detect the mark. The move is tied to the EU AI Act’s transparency requirements. (
At first glance, this looks like an AI transparency story.
For lawyers, it could become a litigation story.
Imagine opposing counsel being able to determine that a significant portion of your brief was generated by an AI system. Imagine them being able to identify the sections of your argument where AI did the substantive work.
That changes the dynamics of a case.
AI provenance could become evidence
Lawyers already have professional obligations surrounding competence, confidentiality, accuracy, and supervision. As generative AI becomes embedded in legal work, another issue is emerging: provenance.
Who created this argument?
Who conducted this research?
Who made the strategic decision reflected in this paragraph?
And perhaps most importantly: how much of the work was actually performed by the lawyer?
Today, those questions can be difficult to answer from the finished document. A polished brief does not reveal whether an associate spent ten hours researching the issue, whether a partner drafted every argument, or whether an AI system generated the first version in seconds.
Watermarking potentially changes that.
If provenance technology becomes sufficiently reliable, the final document could contain information about its origin that is invisible to the reader but discoverable by the appropriate technology.
That creates an entirely new layer of information surrounding legal work.
The opposing side gets a new weapon
Consider a hypothetical litigation.
A lawyer submits a 50-page brief containing sophisticated legal analysis. The opposing counsel discovers that substantial portions of the brief contain AI-generated text.
The opposing lawyer now has questions.
Which sections were generated by AI?
Were the cited authorities independently verified?
Were the legal propositions reviewed by counsel?
Did the attorney understand the reasoning?
Were factual assertions generated or merely edited by the lawyer?
Was confidential information provided to the AI system?
And perhaps most strategically: where did the lawyer exercise actual judgment?
That last question could be particularly important.
AI-generated text does not necessarily mean bad legal work. A competent lawyer can use AI as a research assistant, drafting tool, or analytical partner and then exercise substantial judgment over the result.
But provenance could make the distinction visible.
A court, regulator, client, or opposing counsel could potentially distinguish between work that was substantially authored by a lawyer and work that was substantially generated by a machine.
That could affect credibility, discovery disputes, sanctions arguments, professional responsibility questions, and litigation strategy.
Then there is copyright
The copyright implications may be even more profound.
The U.S. Copyright Office has concluded that copyright protects human-authored expression, including human creative selection, arrangement, and modification of AI-generated material. Purely AI-generated material, however, is outside copyright protection. The Office has also said that prompts alone generally do not provide sufficient human control over the expressive elements.
That creates an interesting collision between copyright doctrine and provenance technology.
For years, one of the difficult practical questions has been determining how much of a work was actually generated by AI.
What happens when the technology itself starts answering that question?
If a company claims copyright over software and the underlying code can be identified as AI-generated, provenance could become evidence in a copyright dispute.
The same principle could apply to legal publications, technical documentation, marketing materials, research, books, and other expressive works.
The critical distinction becomes human authorship.
AI can produce something commercially valuable without producing something copyrightable in the same way a human author can.
That distinction is going to matter enormously as AI becomes the primary production layer for knowledge work.
The legal industry may need a new concept of authorship
Law firms have traditionally organized around human authorship.
A partner signs the brief. An associate performs the research. A client owns the work product. A firm is responsible for the quality of the legal analysis.
AI complicates every part of that model.
The lawyer may increasingly become the person who directs, evaluates, edits, validates, and takes responsibility for machine-generated work.
That creates a fundamental question:
If the machine generates the words but the lawyer supplies the judgment, where exactly does authorship begin?
Copyright law is already wrestling with that question. Legal ethics will increasingly have to wrestle with it as well.
And litigation may become the place where these questions get tested most aggressively.
Provenance could become part of the legal record
The implications extend beyond individual briefs.
Law firms could eventually need systems that maintain an auditable record of how important work product was created.
A legal document could have a provenance layer showing:
The human author.
The AI systems used.
The portions generated by AI.
The human modifications.
The sources consulted.
The validation performed.
The final human approval.
That begins to look less like ordinary document management and more like an AI-native chain of custody.
For highly consequential legal work, that could become essential.
The irony is that the same technology designed to make AI more transparent may make AI dependence much more visible.
Today, using AI can be largely invisible.
Tomorrow, it may be measurable.
The strategic question for law firms
The legal profession has spent the last several years debating whether lawyers should use AI.
That is increasingly the wrong question.
The better question is: What happens when everyone can see how the lawyer used AI?
If opposing counsel can identify AI-generated portions of a brief, firms will have to think differently about AI governance.
If clients can see how much of their work was machine-generated, expectations around value and billing may change.
If courts can identify AI-generated arguments, questions about verification and professional responsibility may become more consequential.
And if copyright disputes can use AI provenance as evidence, the economic value of AI-generated intellectual property could become much harder to assess.
This is why Anthropic’s announcement deserves more attention from the legal industry than it is currently receiving.
The important development is not simply that AI-generated text may become identifiable.
It is that AI provenance could turn the invisible process behind knowledge work into evidence.
Once that happens, AI use stops being merely a productivity question.
It becomes a question of authorship, accountability, strategy, credibility, ownership, and ultimately the economics of legal work.
The lawyers who understand that distinction early will have a significant advantage.
Because the future of LegalTech may not just be about knowing what AI can create.
It may be about being able to prove who created what.