Earlier this month,
\nAnthropic announced that text produced by Claude would begin carrying an
\ninvisible watermark, a machine-detectable pattern embedded in the choices the
\nmodel makes as it generates language. The company is adopting a system based on
\nGoogle DeepMind’s SynthID technology, part of a broader effort by the
\ntechnology industry to make synthetic content identifiable. Anthropic’s move
\ncomes as new transparency provisions of the European Union’s AI Act take
\neffect, requiring providers to make certain AI-generated content detectable in
\nmachine-readable form.<\/p>\n
The idea is not
\nparticularly radical. Google already uses SynthID to mark AI-generated text,
\nimages, audio, and video. OpenAI uses C2PA Content Credentials and SynthID for
\nsupported generated images, and recently expanded SynthID watermarking to
\nsupported audio. Microsoft adds metadata to certain media generated or altered
\nwith AI and offers visible or audible watermarks in some of its consumer
\nproducts.\u00a0<\/p>\n
There are good
\nreasons for doing this. Generative AI has made it cheap and increasingly easy
\nto manufacture convincing photographs, voices, videos, and documents. And it is
\nimportant for society to understand when something is real and when it is
\nfabricated. A political candidate can appear to say something she never said. A
\nphotograph can depict an event that never occurred. A familiar voice can be
\nsynthesized from a relatively small sample of recorded speech. As the
\ndistinction between synthetic and recorded media becomes harder for humans to
\ndiscern, some mechanism for establishing provenance begins to look not merely
\nsensible but necessary.<\/p>\n
The European
\nregulations reflect this concern. Their stated purpose is to reduce deception
\nand manipulation and protect the integrity of the information ecosystem. But
\nthere is a difference between a label and a watermark, and that difference
\ndeserves more attention than it has received.<\/p>\n
A label is
\nprimarily intended to tell a person something. We put ingredients on food,
\nwarnings on medicine, and disclosures on financial products because we believe
\npeople should have information that affects how they interpret or use those
\nproducts. One could imagine a similar convention for AI: a photograph labeled
\nas synthetic, an article disclosing that AI assisted in its production, a video
\ntelling viewers that portions were generated rather than recorded. <\/p>\n
An invisible
\nwatermark does something subtly different. It creates information about the
\nartifact that can persist independently of what the creator chooses to
\ndisclose. Google’s text watermarking technology, for example, modifies the
\nprobabilities used to select tokens during generation, producing a statistical
\npattern that can later be detected. Other provenance systems can attach
\ncryptographically verifiable information about how an asset was created and
\nmodified.<\/p>\n
If AI-generated
\ncontent was diligently labeled, we may not need deeply technical
\nAI-watermarking. But we live in a world of deepfakes and clickbait and all
\nmanner of profit and political-based deception. This is core to why we\u2019ve
\nneeded to develop watermarking technology and to begin to regulate it. <\/p>\n
Watermarking is
\nmore than labeling. It is evidence. But evidence developed for one purpose has
\na way of acquiring new purposes as institutions, markets, and laws evolve
\naround it. And that\u2019s what I\u2019d like to dig into today.<\/p>\n
\t\t\t\t\t\t\t
For most of the
\nhistory of creative work, the tools used to make something had remarkably
\nlittle to say about who made it.<\/p>\n
Imagine an author
\nfinishing a novel on a typewriter in 1975. When she delivered the manuscript to
\na publisher, she claimed to have written it. In the ordinary course of events,
\nthat claim was accepted. If someone else appeared and claimed authorship, the
\ndispute would have to be resolved through evidence: drafts, notes,
\ncorrespondence, witnesses, perhaps even the distinctive characteristics of the
\ntypewriter.<\/p>\n
Some cases would be
\neasy to settle. Others would not.<\/p>\n
There was an
\nunavoidable element of trust in the system because the tool itself had no
\nmemory of the creative process. The typewriter could leave forensic evidence,
\nbut it did not maintain a ledger of the author’s relationship with it, nor in
\nfact know who was striking its keys. It did not record which sentences it had
\nhelped compose because, of course, the typewriter did not compose them.<\/p>\n
For decades,
\nincreasingly sophisticated digital tools preserved much of that basic
\narrangement. A novelist could write in Microsoft Word without Microsoft
\nbecoming a participant in the novel. A photographer could alter an image in
\nPhotoshop without Adobe acquiring a creative relationship to the photograph.
\nThe software was instrumental to the work, but the conceptual boundary between
\ntool and creator remained relatively clear.<\/p>\n
Generative AI
\ncomplicates that boundary because the tool no longer merely executes
\ninstructions. It can propose. It can write a sentence, redesign an image,
\ngenerate computer code, suggest a melody, reorganize an argument, or offer
\ntwenty alternatives to an idea. Increasingly, it can do these things inside the
\nordinary software people already use to work.<\/p>\n
This creates a
\ncategory that will probably become far more common than either purely human or
\npurely AI-generated work: AI-augmented work.<\/p>\n
The article you are
\nreading belongs in that category.<\/p>\n
I began with the
\nargument. I developed it in conversation with an AI assistant. I supplied
\nexamples and analogies, including the typewriter comparison. The AI helped me
\ntest the argument, suggested ways to organize it, and produced draft language.
\nI made decisions about what belonged, what did not, and what the argument
\nultimately meant. I edited and updated drafts and then loaded them back into AI
\nfor feedback, additional research and verification. The iterative process took
\nmultiple iterations and significant time. I wrote and rewrote much of this
\narticle offline. But I won\u2019t claim to have typed every sentence. I augmented my
\noriginal ideas with modern tooling to drive towards a higher-quality outcome
\nthan if I had not used AI-assistance. The final work emerged from that human to
\nAI back-and-forth.<\/p>\n
So who created it?<\/p>\n
Our instinct is to
\nanswer that question by looking for a percentage. Perhaps a work that is 90
\npercent human and 10 percent AI is human, while one that reverses those
\nproportions is artificial. But creative contribution has never been
\nparticularly amenable to arithmetic. Ten words can contain the central insight
\nof an essay. A single editorial suggestion can transform a book. An art
\ndirector may profoundly influence an image without touching the camera.<\/p>\n
A watermark cannot
\nresolve this problem. It can establish something narrower. And understanding
\njust how much narrower requires separating four ideas that are likely to become
\nincreasingly entangled: provenance, participation, authorship, and ownership.<\/p>\n
\t\t\t\t\t\t\t
Provenance is the
\nsimplest of these concepts. It concerns history: Where did an artifact come
\nfrom? What happened to it along the way? Which tools interacted with it?<\/p>\n
This is precisely
\nthe problem that systems such as C2PA’s Content Credentials are designed to
\naddress. C2PA defines provenance as information about the history of a digital
\nasset and its interactions with actors and other assets. Its credentials can
\ncontain cryptographically verifiable information about an artifact’s origin and
\nsubsequent modifications.<\/p>\n
From provenance, we
\ncan sometimes establish participation.<\/p>\n
If a detectable
\nwatermark associated with a particular model survives in a document, that may
\nprovide evidence that the model participated in generating some of its text.
\nOpenAI makes an important version of this distinction in explaining its
\nimage-verification technology: detecting its provenance signals can indicate
\nthat an image was generated with OpenAI tools, but does not establish that the
\nimage is accurate, unedited, legally owned, or being presented in the proper
\ncontext.<\/p>\n
Participation, in
\nother words, is not authorship.<\/p>\n
This distinction
\nbecomes especially important as AI moves from being a destination, a website or
\nuser interface one visits to ask for something, to being a feature embedded
\nthroughout ordinary software. Consider a photographer who uses AI to remove a
\ndistracting object from an otherwise original photograph. Or a programmer who
\nwrites a large software application but accepts several functions suggested by
\na coding assistant. Or an attorney who writes a brief and asks an AI system to
\nmake two paragraphs clearer. Or an author who submits a chapter for editing,
\naccepts five suggested changes, rejects twenty, and rewrites another three
\nherself.<\/p>\n
In each case, AI
\nparticipated.<\/p>\n
That fact tells us
\nremarkably little about authorship.<\/p>\n
Authorship asks a
\ndifferent and much more difficult question: Who supplied the expressive choices
\nthat make the work what it is? Who conceived the argument, selected the
\ncomposition, determined the structure, chose among alternatives, and exercised
\nthe judgment that produced the final artifact?<\/p>\n
American copyright
\npolicy already recognizes some of this complexity. In its 2025 report on AI and
\ncopyrightability, the U.S. Copyright Office concluded that generative-AI
\noutputs can receive copyright protection when a human determines sufficient
\nexpressive elements, and that using AI as an assistive tool does not itself
\nprevent copyright protection. Prompting alone, by contrast, is generally
\ninsufficient.<\/p>\n
Even authorship,
\nhowever, is not ownership.<\/p>\n
Human beings
\nroutinely create things they do not ultimately own. Employees produce works
\nwhose copyrights may belong to employers. Authors transfer rights to
\npublishers. Multiple contributors can possess different interests in the same
\nwork. Ownership is a legal and economic arrangement layered on top of creation.<\/p>\n
These four concepts
\ntherefore form something like a ladder. Provenance tells us where something has
\nbeen. Participation tells us who or what contributed to the process. Authorship
\nasks who actually created the protected expression. Ownership determines who
\npossesses the rights.<\/p>\n
A watermark begins
\nnear the bottom of that ladder. The question is whether, over time, we will
\nallow it to climb.<\/p>\n
\t\t\t\t\t\t\t
There is no reason
\nto believe that watermarking itself gives an AI company copyright in the things
\nits models produce. Under current American law, it does not. Nor is there
\nevidence that the companies developing these systems are conspiring to use
\ntransparency regulation as a back door to ownership. The more interesting
\nconcern is structural rather than conspiratorial.<\/p>\n
Suppose that 15
\nyears from now, a creator becomes involved in a dispute over a valuable work
\nproduced with substantial AI assistance. The creator says that the idea was
\nhers, that she developed its essential form, and that AI was simply one of
\nseveral tools involved.<\/p>\n
The technology
\nprovider, meanwhile, may possess something the creator does not: an extensive,
\nmachine-verifiable record.<\/p>\n
There could be
\ntimestamps. Model identifiers. Generation records. Cryptographic credentials.
\nStatistical watermarks embedded in surviving portions of the artifact. Perhaps
\nthere will be a history showing dozens or hundreds of interactions between the
\ncreator and the system.<\/p>\n
None of this would
\nnecessarily establish authorship. It certainly would not, on its own, establish
\nownership.<\/p>\n
But it might look
\nremarkably authoritative.<\/p>\n
And this is where
\nthe distinction among the four concepts becomes important. Evidence can be
\nexcellent at answering one question and still be poor evidence for another. A
\nwatermark might provide strong evidence that an AI system participated in
\nproducing a work while providing almost no evidence about the relative creative
\nimportance of that participation.<\/p>\n
My worry is that
\ninstitutions have a natural tendency to privilege what can be measured. The
\ncreator’s evidence might consist of memory, intention, judgment, notebooks,
\nconversations, and testimony about how an idea developed. The corporation’s
\nevidence might consist of cryptographically authenticated records generated
\nautomatically at industrial scale.<\/p>\n
One account is
\nhuman and interpretive. The other looks objective.<\/p>\n
The danger is not
\nnecessarily that the machine record is false. The danger is that better
\nevidence about one question may acquire undue authority over a different
\nquestion.<\/p>\n
Proof of
\nparticipation can begin to feel like proof of authorship. Proof of authorship
\ncan begin to influence assumptions about ownership. The steps are individually
\nsmall. The distance between the first and the last is not.<\/p>\n
\t\t\t\t\t\t\t
Technology often
\ncreates capabilities before society has decided how those capabilities should
\nbe governed. This is especially true of data. Systems built for convenience
\nbecome systems of surveillance. Records collected for security become valuable
\nfor advertising. Data retained for operational purposes becomes discoverable in
\nlitigation. None of these secondary uses needs to have been part of the
\noriginal plan.<\/p>\n
AI provenance may
\nfollow a similar path.<\/p>\n
Today, its
\nrationale is compelling: help people identify synthetic media, combat
\ndeception, and provide greater transparency about what they encounter online.
\nBut once the infrastructure exists, it will exist for other purposes as well.<\/p>\n
Imagine a world in
\nwhich AI systems participate, however modestly, in a substantial share of human
\nintellectual production. They help write books and business plans, edit
\nphotographs and films, produce computer code, design products, draft contracts,
\nanalyze scientific results, write emails and refine inventions. At the same
\ntime, those systems leave behind increasingly durable evidence of their participation.<\/p>\n
The companies
\noperating them will then possess something previous toolmakers generally did
\nnot: a technically sophisticated record connecting their products to the
\ncreative process itself.<\/p>\n
Perhaps nothing
\nconsequential will come of that. Current copyright principles may prove
\nperfectly capable of maintaining the distinction between tool and author.
\nCourts may insist that provenance establishes only provenance and refuse to
\ninfer creative rights from technical participation.<\/p>\n
But it is also
\npossible that the law will evolve. New forms of licensing may emerge. New
\ntheories of machine contribution may be proposed. Contractual terms may change.
\nCourts may confront disputes we have not yet imagined. Economic pressure may
\nencourage companies to seek rights that seem implausible today.<\/p>\n
If any of that
\nhappens, the evidentiary infrastructure will already have been built. And that
\nis what makes the present moment worth examining. The important question is not
\nwhether Anthropic, Google, OpenAI, Microsoft, or anyone else intends to claim
\nownership of AI-assisted work. There is no basis for making that accusation,
\nand intention may ultimately be beside the point. The question is what becomes
\npossible once society has created a persistent technical record of machine
\nparticipation in human creativity.<\/p>\n
For most of
\nhistory, we lived with an imperfect arrangement. People claimed authorship,
\nother people sometimes challenged them, and institutions tried to determine
\nwhat happened from whatever evidence survived. There were ambiguities and
\ninjustices in that system, as there are in any system built around human
\ntestimony and incomplete records.<\/p>\n
We now have the
\nability to replace some of that ambiguity with data. That is usually described
\nas progress. And often it is. But data does not merely resolve uncertainty. It
\nredistributes power toward whoever collects it, controls it, interprets it, and
\npersuades institutions of what it means.<\/p>\n
The great promise
\nof AI watermarking is that, years from now, we may be able to ask whether an
\nartificial intelligence participated in creating something and receive a much
\nbetter answer than we can today.<\/p>\n
We should build
\nthat capability with care. Because over time we may discover that society has
\nbegun asking the watermark a different question. \u201cDid AI participate in the
\ncreation of this\u201d is today\u2019s intended question. But when that question shifts
\nto \u201cWho deserves credit\u201d and \u201cWho owns it\u201d, will AI-platforms have an argument
\nto begin laying claim?\u00a0 The evidence may
\nalready carry more authority than we intended to give it.<\/p>\n","protected":false},"excerpt":{"rendered":"Earlier this month, Anthropic announced that text produced by Claude would begin carrying an invisible watermark, a machine-detectable…\n","protected":false},"author":2,"featured_media":132884,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[53,25,4846,7412,134,10684],"class_list":["post-149228","post","type-post","status-publish","format-standard","has-post-thumbnail","category-anthropic","tag-anthropic","tag-artificial-intelligence","tag-arts","tag-copyright","tag-technology","tag-wral-techwire"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/149228","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=149228"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/149228\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/132884"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=149228"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=149228"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=149228"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}