Earlier this month, Anthropic announced that text produced by Claude would begin carrying an invisible watermark, a machine-detectable pattern embedded in the choices the model makes as it generates language. The company is adopting a system based on Google DeepMind's SynthID technology, part of a broader effort by the technology industry to make synthetic content identifiable. Anthropic's move comes as new transparency provisions of the European Union's AI Act take effect, requiring providers to make certain AI-generated content detectable in machine-readable form.
The idea is not particularly radical. Google already uses SynthID to mark AI-generated text, images, audio, and video. OpenAI uses C2PA Content Credentials and SynthID for supported generated images, and recently expanded SynthID watermarking to supported audio. Microsoft adds metadata to certain media generated or altered with AI and offers visible or audible watermarks in some of its consumer products.
Other WRAL Top Stories
There are good reasons for doing this. Generative AI has made it cheap and increasingly easy to manufacture convincing photographs, voices, videos, and documents. And it is important for society to understand when something is real and when it is fabricated. A political candidate can appear to say something she never said. A photograph can depict an event that never occurred. A familiar voice can be synthesized from a relatively small sample of recorded speech. As the distinction between synthetic and recorded media becomes harder for humans to discern, some mechanism for establishing provenance begins to look not merely sensible but necessary.
The European regulations reflect this concern. Their stated purpose is to reduce deception and manipulation and protect the integrity of the information ecosystem. But there is a difference between a label and a watermark, and that difference deserves more attention than it has received.
A label is primarily intended to tell a person something. We put ingredients on food, warnings on medicine, and disclosures on financial products because we believe people should have information that affects how they interpret or use those products. One could imagine a similar convention for AI: a photograph labeled as synthetic, an article disclosing that AI assisted in its production, a video telling viewers that portions were generated rather than recorded.
An invisible watermark does something subtly different. It creates information about the artifact that can persist independently of what the creator chooses to disclose. Google's text watermarking technology, for example, modifies the probabilities used to select tokens during generation, producing a statistical pattern that can later be detected. Other provenance systems can attach cryptographically verifiable information about how an asset was created and modified.
If AI-generated content was diligently labeled, we may not need deeply technical AI-watermarking. But we live in a world of deepfakes and clickbait and all manner of profit and political-based deception. This is core to why we’ve needed to develop watermarking technology and to begin to regulate it.
Watermarking is more than labeling. It is evidence. But evidence developed for one purpose has a way of acquiring new purposes as institutions, markets, and laws evolve around it. And that’s what I’d like to dig into today.
The typewriter that kept no records
For most of the history of creative work, the tools used to make something had remarkably little to say about who made it.
Imagine an author finishing a novel on a typewriter in 1975. When she delivered the manuscript to a publisher, she claimed to have written it. In the ordinary course of events, that claim was accepted. If someone else appeared and claimed authorship, the dispute would have to be resolved through evidence: drafts, notes, correspondence, witnesses, perhaps even the distinctive characteristics of the typewriter.
Some cases would be easy to settle. Others would not.
There was an unavoidable element of trust in the system because the tool itself had no memory of the creative process. The typewriter could leave forensic evidence, but it did not maintain a ledger of the author's relationship with it, nor in fact know who was striking its keys. It did not record which sentences it had helped compose because, of course, the typewriter did not compose them.
For decades, increasingly sophisticated digital tools preserved much of that basic arrangement. A novelist could write in Microsoft Word without Microsoft becoming a participant in the novel. A photographer could alter an image in Photoshop without Adobe acquiring a creative relationship to the photograph. The software was instrumental to the work, but the conceptual boundary between tool and creator remained relatively clear.
Generative AI complicates that boundary because the tool no longer merely executes instructions. It can propose. It can write a sentence, redesign an image, generate computer code, suggest a melody, reorganize an argument, or offer twenty alternatives to an idea. Increasingly, it can do these things inside the ordinary software people already use to work.
This creates a category that will probably become far more common than either purely human or purely AI-generated work: AI-augmented work.
The article you are reading belongs in that category.
I began with the argument. I developed it in conversation with an AI assistant. I supplied examples and analogies, including the typewriter comparison. The AI helped me test the argument, suggested ways to organize it, and produced draft language. I made decisions about what belonged, what did not, and what the argument ultimately meant. I edited and updated drafts and then loaded them back into AI for feedback, additional research and verification. The iterative process took multiple iterations and significant time. I wrote and rewrote much of this article offline. But I won’t claim to have typed every sentence. I augmented my original ideas with modern tooling to drive towards a higher-quality outcome than if I had not used AI-assistance. The final work emerged from that human to AI back-and-forth.
So who created it?
Our instinct is to answer that question by looking for a percentage. Perhaps a work that is 90 percent human and 10 percent AI is human, while one that reverses those proportions is artificial. But creative contribution has never been particularly amenable to arithmetic. Ten words can contain the central insight of an essay. A single editorial suggestion can transform a book. An art director may profoundly influence an image without touching the camera.
A watermark cannot resolve this problem. It can establish something narrower. And understanding just how much narrower requires separating four ideas that are likely to become increasingly entangled: provenance, participation, authorship, and ownership.
From provenance to ownership
Provenance is the simplest of these concepts. It concerns history: Where did an artifact come from? What happened to it along the way? Which tools interacted with it?
This is precisely the problem that systems such as C2PA's Content Credentials are designed to address. C2PA defines provenance as information about the history of a digital asset and its interactions with actors and other assets. Its credentials can contain cryptographically verifiable information about an artifact's origin and subsequent modifications.
From provenance, we can sometimes establish participation.
If a detectable watermark associated with a particular model survives in a document, that may provide evidence that the model participated in generating some of its text. OpenAI makes an important version of this distinction in explaining its image-verification technology: detecting its provenance signals can indicate that an image was generated with OpenAI tools, but does not establish that the image is accurate, unedited, legally owned, or being presented in the proper context.
Participation, in other words, is not authorship.
This distinction becomes especially important as AI moves from being a destination, a website or user interface one visits to ask for something, to being a feature embedded throughout ordinary software. Consider a photographer who uses AI to remove a distracting object from an otherwise original photograph. Or a programmer who writes a large software application but accepts several functions suggested by a coding assistant. Or an attorney who writes a brief and asks an AI system to make two paragraphs clearer. Or an author who submits a chapter for editing, accepts five suggested changes, rejects twenty, and rewrites another three herself.
In each case, AI participated.
That fact tells us remarkably little about authorship.
Authorship asks a different and much more difficult question: Who supplied the expressive choices that make the work what it is? Who conceived the argument, selected the composition, determined the structure, chose among alternatives, and exercised the judgment that produced the final artifact?
American copyright policy already recognizes some of this complexity. In its 2025 report on AI and copyrightability, the U.S. Copyright Office concluded that generative-AI outputs can receive copyright protection when a human determines sufficient expressive elements, and that using AI as an assistive tool does not itself prevent copyright protection. Prompting alone, by contrast, is generally insufficient.
Even authorship, however, is not ownership.
Human beings routinely create things they do not ultimately own. Employees produce works whose copyrights may belong to employers. Authors transfer rights to publishers. Multiple contributors can possess different interests in the same work. Ownership is a legal and economic arrangement layered on top of creation.
These four concepts therefore form something like a ladder. Provenance tells us where something has been. Participation tells us who or what contributed to the process. Authorship asks who actually created the protected expression. Ownership determines who possesses the rights.
A watermark begins near the bottom of that ladder. The question is whether, over time, we will allow it to climb.
When evidence answers the wrong question
There is no reason to believe that watermarking itself gives an AI company copyright in the things its models produce. Under current American law, it does not. Nor is there evidence that the companies developing these systems are conspiring to use transparency regulation as a back door to ownership. The more interesting concern is structural rather than conspiratorial.
Suppose that 15 years from now, a creator becomes involved in a dispute over a valuable work produced with substantial AI assistance. The creator says that the idea was hers, that she developed its essential form, and that AI was simply one of several tools involved.
The technology provider, meanwhile, may possess something the creator does not: an extensive, machine-verifiable record.
There could be timestamps. Model identifiers. Generation records. Cryptographic credentials. Statistical watermarks embedded in surviving portions of the artifact. Perhaps there will be a history showing dozens or hundreds of interactions between the creator and the system.
None of this would necessarily establish authorship. It certainly would not, on its own, establish ownership.
But it might look remarkably authoritative.
And this is where the distinction among the four concepts becomes important. Evidence can be excellent at answering one question and still be poor evidence for another. A watermark might provide strong evidence that an AI system participated in producing a work while providing almost no evidence about the relative creative importance of that participation.
My worry is that institutions have a natural tendency to privilege what can be measured. The creator's evidence might consist of memory, intention, judgment, notebooks, conversations, and testimony about how an idea developed. The corporation's evidence might consist of cryptographically authenticated records generated automatically at industrial scale.
One account is human and interpretive. The other looks objective.
The danger is not necessarily that the machine record is false. The danger is that better evidence about one question may acquire undue authority over a different question.
Proof of participation can begin to feel like proof of authorship. Proof of authorship can begin to influence assumptions about ownership. The steps are individually small. The distance between the first and the last is not.
The infrastructure comes before the law
Technology often creates capabilities before society has decided how those capabilities should be governed. This is especially true of data. Systems built for convenience become systems of surveillance. Records collected for security become valuable for advertising. Data retained for operational purposes becomes discoverable in litigation. None of these secondary uses needs to have been part of the original plan.
AI provenance may follow a similar path.
Today, its rationale is compelling: help people identify synthetic media, combat deception, and provide greater transparency about what they encounter online. But once the infrastructure exists, it will exist for other purposes as well.
Imagine a world in which AI systems participate, however modestly, in a substantial share of human intellectual production. They help write books and business plans, edit photographs and films, produce computer code, design products, draft contracts, analyze scientific results, write emails and refine inventions. At the same time, those systems leave behind increasingly durable evidence of their participation.
The companies operating them will then possess something previous toolmakers generally did not: a technically sophisticated record connecting their products to the creative process itself.
Perhaps nothing consequential will come of that. Current copyright principles may prove perfectly capable of maintaining the distinction between tool and author. Courts may insist that provenance establishes only provenance and refuse to infer creative rights from technical participation.
But it is also possible that the law will evolve. New forms of licensing may emerge. New theories of machine contribution may be proposed. Contractual terms may change. Courts may confront disputes we have not yet imagined. Economic pressure may encourage companies to seek rights that seem implausible today.
If any of that happens, the evidentiary infrastructure will already have been built. And that is what makes the present moment worth examining. The important question is not whether Anthropic, Google, OpenAI, Microsoft, or anyone else intends to claim ownership of AI-assisted work. There is no basis for making that accusation, and intention may ultimately be beside the point. The question is what becomes possible once society has created a persistent technical record of machine participation in human creativity.
For most of history, we lived with an imperfect arrangement. People claimed authorship, other people sometimes challenged them, and institutions tried to determine what happened from whatever evidence survived. There were ambiguities and injustices in that system, as there are in any system built around human testimony and incomplete records.
We now have the ability to replace some of that ambiguity with data. That is usually described as progress. And often it is. But data does not merely resolve uncertainty. It redistributes power toward whoever collects it, controls it, interprets it, and persuades institutions of what it means.
The great promise of AI watermarking is that, years from now, we may be able to ask whether an artificial intelligence participated in creating something and receive a much better answer than we can today.
We should build that capability with care. Because over time we may discover that society has begun asking the watermark a different question. “Did AI participate in the creation of this” is today’s intended question. But when that question shifts to “Who deserves credit” and “Who owns it”, will AI-platforms have an argument to begin laying claim? The evidence may already carry more authority than we intended to give it.