1. Attribution & Accountability

In November 2025, the music streaming platform Deezer and polling company Ipsos published the results of a listening study — 9,000 participants across eight countries were asked to distinguish AI-generated music from human-made work in a blind test. Ninety-seven percent couldn’t tell the difference.

Okay. So AI is getting better at making things. But here are the practical implications.

By June 2026 Deezer was receiving roughly 90,000 fully AI-generated tracks a day — more than half of all new uploads to the platform, up from 10,000 a day and a tenth of uploads eighteen months earlier. The company estimates that as much as 85 percent of the plays those tracks get are fraudulent, with no human involved in the “listening.” Spotify said in September 2025 that it had removed more than 75 million spammy tracks over the preceding twelve months. And on Amazon, a company that sells AI-detection software scanned 844 books in the “Success” subcategory and judged 77 percent of them were likely AI-generated.

The basic legibility of creative work — who made this, how was it made, is it what it claims to be — is breaking down. Not gradually, but structurally, and in a tidal wave. The mechanisms we’ve relied on for centuries to establish authorship, assign credit, and hold people accountable for what they produce are all failing to keep pace.

This essay attempts to map that breakdown and address its consequences. It’s part of a three-system analytical framework that organizes issues in AI-and-creativity not by issue or jurisdiction — categories that shift with every court ruling, legislative session and corporate licensing deal — but by function: the role each system plays in how creative work gets made, valued, and protected. Attribution and accountability is the first of those systems, built around a simple question:

Who or what gets recognized, trusted, or held responsible for creative acts?

The answer used to be straightforward. It isn’t anymore.

#Traditional attribution

For most of the history of creative work, attribution was simple. A painter signed a canvas. A writer got a byline. A composer’s name appeared on the score. A copyright notice established who owned the work.

These conventions served multiple functions at once — establishing authorship, enabling compensation, creating legal accountability, building reputation, establishing provenance. The system worked because the question “who made this?” had, in most cases, a clear and, importantly, verifiable answer.

That bundling is coming apart. AI doesn’t challenge just one of these functions; it forces all of them to operate independently, often at different layers of the same creative act, and often in tension or opposition with one another.

The court weighing whether OpenAI’s training on New York Times journalism qualifies as fair use is asking a different question from the legislator trying to require AI-generated content to carry a label, which is a different question from the musician trying to prevent an AI clone of her voice from appearing on streaming platforms. These might all seem like different issues, but they’re not. They’re expressions of the same structural problem: the old attribution system assumed a world in which humans made things, and that we could tell.

Both assumptions are now unreliable.

#From works to layers

Here is the core analytical reframing that seems essential: AI has disaggregated creative output into multiple extractable layers, each of which now requires its own attribution mechanism.

A single artist’s creative identity can be decomposed — a specific recording, a recognizable voice, an accumulated stylistic vocabulary, the individual performance choices that distinguish one interpretation from another. Before generative AI, these layers traveled together. You couldn’t extract Billie Holiday’s phrasing from her voice from her songs from the biographical experience that inflected all of it. Now, in principle, you can isolate each one and define and value it independently. And each layer faces a different attribution challenge.

There are seven layers.

Authorship. Who or what created this work? The Copyright Office’s framework holds that copyright requires a human author, and the D.C. Circuit endorsed the narrow version of that principle last year in Thaler v. Perlmutter, holding that a machine cannot be an author — in a case where the applicant had named the AI system itself as sole author. The Supreme Court declined to hear it this past March, leaving the ruling undisturbed without endorsing it.

What the court pointedly did not decide is the harder question: how much human contribution is enough when a person uses AI as a tool? Judge Patricia Millett wrote that line-drawing disagreements over how much AI contributed were “neither here nor there in this case.”

So the spectrum runs from unprotectable pure-AI output through protectable human modifications to fully protectable human-AI collaboration, and its boundaries are undefined. Allen v. Perlmutter, in which Jason Allen argues that hundreds of iterative Midjourney prompts plus post-processing amount to authorship of Théâtre D’opéra Spatial, is the only case squarely on that threshold. It was fully briefed early this year and remains undecided.

Provenance. How was this made, and can we verify the chain of creation? This is the infrastructure layer, the plumbing that makes other forms of attribution possible. The C2PA (Coalition for Content Provenance and Authenticity) content credentials standard — steered by Adobe, Google, Microsoft, OpenAI and others — embeds tamper-evident metadata recording how content was created.

Implementation is uneven in ways worth noticing. Nikon shipped capture-side credentials last August, but on a single camera body. Samsung’s Galaxy S25 phone attaches credentials only to AI-generated or AI-edited images, so an ordinary photograph taken on one carries no provenance data at all.

California has built the most developed statutory version, and it’s instructive that the obligation walks is progressive rather than landing all at once. Training-data disclosure took effect in January. The state’s AI Transparency Act, requiring a free detection tool and embedded latent disclosures, kicks in this August. Provenance duties reach large online platforms and generative AI hosting services in January 2027, and camera manufacturers in 2028. Notably, the statute is standard-agnostic — it requires provenance data consistent with widely accepted industry practice and never names C2PA.

And California wasn’t first. China’s labeling measures, requiring labels embedded in file metadata, took effect last September.

But provenance infrastructure has a known and fatal vulnerability: metadata can be stripped, rendering it useless and open to corruption.

Identity and likeness. Is this person who they appear to be? Synthetic media can now produce convincing replications of real people’s faces and voices.

The NO FAKES Act would establish a federal right to control AI-generated digital replicas, with a notice-and-takedown process modeled on the DMCA (Digital Millennium Copyright Act). It has drawn an unusual coalition — the studios, the music labels, the guilds and the major AI developers — along with testimony from performers including FKA twigs and Martina McBride, though a year apart and before different subcommittees. A revised version advanced unanimously out of Senate Judiciary in June, making it the furthest-advanced federal AI bill of its kind.

At the state level, Tennessee’s ELVIS Act extended that state’s right of publicity to voice and AI-generated replicas. Roughly a dozen states have now enacted digital-replica or AI-likeness provisions, most recently Washington. Though several — California, New York, Illinois — arrived by amending long-standing right-of-publicity statutes on their own tracks rather than following Tennessee’s model.

This is arguably the most enforceable of the attribution functions, because faces and voices are detectable even when other AI uses aren’t.

Artistic essence. This is the one that current frameworks don’t cover. When GPT-4o’s image generator arrived in March 2025 and the internet filled with Studio Ghibli-style renderings, Studio Ghibli said nothing at all. “We are not making any comments,” a spokesperson offered.

Copyright protects specific expressions, not style. Likeness rights protect faces and voices, not aesthetic sensibilities. What was being extracted — the accumulated creative identity of a lifetime of artistic decisions, the thing that makes a Ghibli frame look like a Ghibli frame — doesn’t yet have a legal name.

This matters enormously, because style transfer is arguably the most common form of AI extraction from living artists. The prompt isn’t “reproduce this painting.” It’s “make something that looks like this artist’s work.”

And the question it raises — when does Bach cease to be Bach, as a recent profile of the harpsichordist Jean Rondeau put it to his reworked transcription of Bach’s Goldberg Variations — is one the policy conversation has barely begun to address. Rondeau’s own answer is that the question is less fraught than it sounds: “Bach did not hesitate to reuse and transform his own material.” Which is exactly the difficulty. The line between homage and extraction has never been drawn by law, because until now nothing could cross it at scale.

Transparency and disclosure. Must the origin of content be declared? Multiple federal labeling bills are pending. The EU AI Act requires transparency for general-purpose systems. Platforms have responded with their own approaches — YouTube’s synthetic media labeling, Deezer’s AI tagging, iHeartRadio’s “Guaranteed Human” pledge.

The logic is sensible: if audiences can’t detect AI content on their own — and 97 percent of listeners couldn’t — shouldn’t someone be required to tell them?

But this kind of labeling faces problems the other layers don’t, and I’ll come back to those.

Training data accountability. Whose work built the system that produced the output? This is attribution traced backward. Not “who made this?” but “whose creative labor made this possible?”

In America, the TRAIN Act would create a discovery mechanism modeled on the DMCA subpoena: on a sworn declaration of good-faith belief, a copyright holder could compel an AI developer to produce the training material or records sufficient to identify it, with non-compliance creating a presumption that copying occurred. The bill sits in committee without a hearing. The COPIED Act would direct NIST to set provenance standards, bar the removal of provenance information, and prohibit training on provenance-bearing content without authorization. It’s waiting too.

The EU requires providers of general-purpose models to publish a “sufficiently detailed summary” of training content on a Commission template. But be clear about what that is and isn’t. A public narrative summary with domain lists and crawler information is not itemized disclosure, and it won’t tell an individual author whether her book was used.

And the courts are reaching different conclusions in different jurisdictions — though not in the way the headlines suggest.

Last November the Munich Regional Court held in GEMA v. OpenAI that OpenAI had infringed German copyright. But the holding is narrower and stranger than it sounds. The court accepted that the initial scraping and data-conversion phase fell within the text-and-data-mining exception. It found infringement instead in memorization — lyrics embedded and retained in the model’s parameters, which the court held was a reproduction beyond what mining permits — and in the reproduction of those lyrics in ChatGPT’s outputs. On whether the act of training itself infringes, the court said the question “has not been clearly clarified” and declined to reach it. The judgment is under appeal.

In London, where the first trial anywhere on AI was supposed to settle the question, it was never reached at all: Getty abandoned its training claim before closing submissions for want of evidence that the training had happened in the UK. And this month the Delhi High Court read training as fair dealing under Indian law and refused an injunction against OpenAI.

Two years in, no court anywhere has finally resolved whether training on copyrighted creative work infringes. What the courts have resolved are adjacent questions, each of which happens to be more tractable: what a finished model is, whether a model memorized, whether a party could prove where the copying occurred.

There’s one more route, and I find it the most interesting thing in this whole section precisely because nobody designed it for this. Both the independent artists’ class actions against Suno and Udio and Sony’s suit against Udio this month plead circumvention under the Digital Millennium Copyright Act — the allegation being not that training infringes but that the companies defeated technical protection measures to acquire the files, stream-ripping from YouTube and Spotify. That theory carries no fair-use defense, which means it survives whatever a court eventually decides about training. In May a court sustained those claims against Udio while dismissing the output-infringement claim.

And the evidence underpinning it arrived from outside the policy system entirely. This month, source code obtained in a breach of Suno documented the company’s training corpus by name — YouTube Music, Deezer, Genius, Jamendo, Pond5, IMSLP — with clip counts and hour counts attached.

Every other training-provenance fact this project tracks arrived through litigation discovery, a regulatory disclosure mandate, or company self-reporting. This one arrived through a hack, a route no policy instrument controls and no defendant can game out.

Audience. The first six functions describe the supply side of the attribution chain — who made what, and how. But creative work is a transaction, and the demand side is just as compromised.

Our primary systems for establishing cultural value lean heavily on audience metrics — streams, engagement, follower counts, chart positions. These metrics function as attribution. When Spotify reports that a track has 10 million streams, that number is doing more that fiscal accounting, it’s doing attribution work, measuring cultural significance, audience interest, market demand. It’s saying people chose this.

It’s often the basis for what we determine is a successful work and what is not. Royalty pools distribute accordingly. Recommendation algorithms amplify accordingly. For right or wrong, the economic infrastructure of creative industries treats audience metrics as a reliable proxy for what human beings actually care about.

That proxy is failing. But be certain about how, because two different failures get run together and they have different remedies.

The first is dilution. AI-generated work takes a large share of a catalogue and a small share of consumption, and three independent datasets now measure the same asymmetry. AI titles are roughly 20 percent of Amazon’s self-published fiction catalogue and 12 percent of its sales. They’re 5 percent of Spotify’s catalogue and 1 percent of its recommendation graph. They’re more than half of Deezer’s daily uploads and between 1-3 percent of its streams.

What the flood does, in the first instance, is spread a slowly growing revenue pool across a catalogue growing far faster than demand. That requires nobody to be fooled. It’s arithmetic. And it sets a test any proposed remedy has to pass: does it change the ratio or only the labeling?

The second failure is fraud, and that one is an attribution problem in the sense this essay means. Up to 85 percent of the streams that AI tracks get at Deezer are fraudulent. Those streams don’t represent human listeners making choices. They represent synthetic engagement manufacturing the “evidence” currency that determines who gets paid, who gets discovered, and what the culture appears to value.

Spotify’s removal of 75 million “spammy” tracks and Amazon’s three-book-per-day upload limit are platform responses to the same problem. So is the scale of the monetized slop economy: one study identified 278 channels on YouTube posting exclusively AI-generated content, drawing 63 billion views and an estimated $117 million in annual ad revenue.

Put the two failures together and the shape of the problem changes. It isn’t that audiences are being swamped. It’s that a catalogue nobody can sort is being paid out of the same pool, and that a measurable fraction of the signal determining those payments is manufactured.

This sits, admittedly, on top of an already unreliable system of audience generation. Channels with wider reach make the work seen in them disproportionately more popular than it would be on smaller platforms. Content chosen by curators evaluating quality is one thing. But when algorithms determine what gets seen by usage or attention metrics rather than any kind of intrinsic creative quality, the shift in value to attention is profound — and easily manipulated by whoever controls the dials.

The attention problem goes deeper than bot farms. Even legitimate audience metrics were designed by platforms for their own platform purposes — advertising targeting, engagement optimization — not for accurately representing the artist-audience relationship. Someone who lets a playlist run in the background while cooking is attributed the same cultural choice as someone who sought out a piece of music, listened closely, and was changed by it.

That flattening was surely a problem before AI. But AI makes it structural, because now you can generate both the content and the audience at scale. Synthetic content performing for synthetic audiences, flowing through measurement systems that can’t distinguish either from the real thing, isn’t just a market distortion, it’s a market fiction, and ripe for manipulation and fraud.

If we’re building an attribution framework that accounts only for the supply side — who made the work — while leaving the demand side unverified, we’re only describing half a system. The artist-audience compact is also an attribution relationship, and it’s breaking down by the same mechanisms, and for the same reasons as every other function on this list.

#The multi-level problem

These seven functions don’t operate in isolation. A single AI system can extract from a single artist at multiple layers simultaneously — using a specific recording (training data), a recognizable voice (identity), accumulated stylistic choices (artistic essence), and producing something that presents as a new work (authorship). Then the output is validated by audience metrics that may themselves be synthetic.

That’s five distinct attributions spanning both sides of the creative chain. And the policy responses to each are being developed in separate silos, by different actors, with different legal theories.

Holly Herndon is one of the artists who seems to have grasped this. Her Holly+ project treated her voice as a separable creative asset with its own approval and revenue-sharing framework — governed by a DAO, a decentralized autonomous organization, of “stewards” who vote on whether works made with her voice should be approved and co-sold — distinct from any specific recording or composition. Her stated reasoning was as much about succession as commerce: “In the event of my death, I feel more comfortable with distributed ownership of the rights to my voice model among a DAO of stewards.”

SAG-AFTRA’s contracts do something similar, though they cut along a different axis, and the difference is instructive. Rather than separating likeness from voice — the union’s definition of a digital replica bundles “voice and/or likeness” together — the agreements distinguish replicas made with a performer’s participation on a job from those built without it, and both from Synthetic Performers, wholly fabricated figures resembling no identifiable person.

The union sorted by consent and employment relationship rather than by which layer was extracted. That’s a defensible choice for a bargaining unit and a revealing one for this framework: the layers AI separates are not the layers labor law knows how to grip. The Interactive Media agreement ratified last July, after a 320-day strike, went furthest — including the right to suspend consent for generating new material during a strike, which I’d call the most structurally interesting AI provision in any of them.

What this decomposition points toward, and what the next essay in this series takes up, is a fundamental question about recompilation. If creative identity can be decomposed into separable layers, can it be reassembled into something licensable and compensable at scale?

Perhaps something like the way ASCAP solved a structurally similar problem a century ago, when music was being performed in too many venues and too many configurations for per-use licensing to function. That problem got solved by abstracting up to collective administration with blanket licenses. The AI version operates at enormously greater scale and across more layers of creative identity. But the architectural logic may be the same.

That’s a value question, though, not an attribution question. The attribution system’s job is to identify the layers.

#The excellence problem

There’s a difficulty with this framework.

In the artist world, the AI creativity debate has landed on “excellence” as a central pillar of resistance. The argument runs something like this: AI cannot produce genuinely excellent work because it lacks human experience, training, and embodied perception. Ergo, human creativity deserves protection.

That’s true. But in an open marketplace, where “I like that” is really the only final authority, excellence is in the eye — or ear — of the beholder. And the excellence argument is making three different claims and conflating them into one.

Provenance is the idea that human origin confers value. Credentialism is the idea that proper preparation confers legitimacy. And excellence — the idea that the work achieves what it sets out to achieve — is not the same thing as either.

We’ve been calling credentials excellence for a long time. The musician who clears the audition at Curtis looks excellent, but what we’re often hearing, as Stanford Thompson — a Curtis graduate who founded Play On Philly — wrote this spring, is “access layered over time and expressing itself as readiness.” Not just talent, in his phrase, but “years of accumulated opportunity, quality of instruction a student had access to.” We’ve been calling provenance excellence too.

When Yuval Sharon staged Isolde giving birth moments before the Liebestod in his Met debut this year, the objection was not that Wagner didn’t intend it — Sharon’s defense was that he’d gone back to Wagner’s own word, Verklärung, transfiguration. The objection was that the staging was too earthbound for music that isn’t. That’s an excellence argument, and it’s worth noticing how rare that is. Most objections of this kind reach for provenance instead, because provenance is easier to win.

AI makes the conflation visible because it strips away the biographical credentials and leaves only the output. If the output holds up — if it moves people, challenges assumptions, does the work it was supposed to — then insisting it isn’t excellent because it was made by AI is a provenance claim, not an excellence claim.

That doesn’t mean provenance doesn’t matter. It does. Human experience, cultural situatedness, the biographical stakes of making something — these are real values. But they’re not the same as quality, and conflating them in the policy debate doesn’t clarify anything.

This matters for the attribution framework because the disclosure and transparency apparatus — the function absorbing the most legislative energy — is essentially a provenance argument. It asks: was this made by a human? And the policy assumption is that the answer matters.

But consider. The 97 percent detection gap means audiences already can’t enforce the distinction. The technology trajectory will make detection harder, not easier. And the incentive structure is brutal, because disclosure appears to carry a market penalty — in one survey this spring, two-thirds of respondents said they’d be less interested in a release marketed as fully AI-generated.

So if disclosure carries a market cost, and detection is impossible, and the distinction between human and AI contribution is increasingly blurred in practice, then the disclosure apparatus is asking creators to voluntarily devalue their own work on the honor system, with no enforcement mechanism.

That doesn’t make transparency worthless. But it means the policy architecture may need to put more weight on the attribution functions that can be enforced — provenance infrastructure built into systems at the point of creation, and training data accountability that operates upstream — rather than relying primarily on downstream disclosure that depends on either detection (which is failing) or self-reporting (which incentives undermine).

The uncomfortable evidence for that proposition is that the most enforceable attribution mechanism currently operating is not an attribution mechanism at all. It’s an anti-circumvention provision written in 1998, being used to reach a question copyright can’t. That’s either an indictment of the architecture this essay maps or a clue about where the leverage actually is.

Some of these attribution functions are more viable than others. But the policy energy may not be going to the right places. And a framework that can’t say so isn’t useful to the researchers and policymakers who need it most.

#Why the functional lens matters

I’m making an implicit argument here for organizing all of this by function rather than by issue or jurisdiction. Three reasons:

First, convergence across domains becomes visible. A visual artist fighting an AI image generator, a musician fighting a voice clone, a journalist whose work trained a large language model — all are experiencing versions of the same attribution failure. Conventional tracking treats these as separate stories in separate sectors. The functional lens shows they’re the same structural breakdown at different layers.

Second, the jurisdiction problem comes into focus. A Munich court held that OpenAI infringed German copyright — not by training, but because lyrics were memorized in the model’s parameters and reproduced in its outputs. Eight months later the Delhi High Court read training as fair dealing under Indian law and refused an injunction. In London, the first trial anywhere never reached the question.

Japan permits use for non-enjoyment purposes such as model training, which makes most training lawful — but it carries a proviso withdrawing the exception where use would unreasonably prejudice the copyright owner’s interests, and the guidance already excludes fine-tuning aimed at reproducing a specific creator’s style. South Korea, by contrast, has no text-and-data-mining exception at all. Its AI framework law governs transparency and risk but is silent on training data and copyright, and a government proposal to add an exception with remuneration and opt-out is being contested by rightsholder groups.

Organizing by jurisdiction produces a confusing patchwork. Organizing by function reveals that the questions are consistent and that it’s the answers that diverge — and, more troublingly, that some jurisdictions aren’t answering the question they appear to be answering. Japan’s guidance, alone among these, has an administrative answer to the style question, which is more than the United States has. That divergence creates an effectively impossible compliance environment for any artist operating across borders.

Third, policy gaps become visible. Map the seven attribution functions against existing and proposed interventions and the holes show up clearly. Provenance infrastructure exists but has a metadata-stripping vulnerability. Training data accountability has a discovery mechanism on paper but no enforcement teeth — transparency without remedy, and in the EU’s case a summary rather than a disclosure. Likeness protection is advancing fastest of the seven, and has now cleared a Senate committee. Artistic essence, arguably the most common form of AI extraction from living artists, has no policy mechanism at all outside a paragraph of Japanese administrative guidance. And audience attribution — the demand-side signal that determines who gets paid and what appears to matter — is being gamed at industrial scale with no systemic response beyond platform-by-platform whack-a-mole.

And there’s a structural gap that cuts across all seven: the unaffiliated individual artist. SAG-AFTRA members have digital replica protections. WGA members have AI contract provisions. The Authors Guild has a partnership with an opt-in licensing marketplace where individual authors enroll works and set their own terms. But the vast majority of working artists aren’t members of guilds with bargaining power, and an opt-in marketplace is not the same instrument as a collective agreement. Most of the attribution infrastructure being built serves institutional or collective actors.

That gap can now be measured. Every one of the 379 developments this project tracks as of July 29, 2026, has been coded for who made the operative decision and whether creators were a party to it. Across the 315 in which something was actually decided, creators were a party in 21 percent, formally consulted in 8 percent, and absent from 72 percent. Creators or their unions held the pen in 18 percent — without a union behind them, in 6 percent. Ties in the coding were broken toward “absent,” which makes that a floor.

Six percent is the unaffiliated artist’s share of the decisions that will govern their work.

There’s one route that reaches them, and it emerged from litigation rather than policy: the independent artists’ class actions against Suno and Udio, where the circumvention claims survived dismissal in May. Those are the only live vehicle by which artists excluded from the major-label settlements reach the AI companies directly. And, again, the surviving claim isn’t about training at all. It’s about how the files were acquired.

#The institutional dimension

Attribution systems don’t operate in a vacuum. They require functioning institutions to administer, interpret, and enforce them. Which seems obvious until you watch the institutional ground shift.

In May 2025 the Trump administration fired Carla Hayden, the Librarian of Congress. Two days later it fired Shira Perlmutter, the Registrar of Copyrights — one day after her office released Part 3 of its AI and copyright analysis, which leaned toward limiting fair use claims for commercial AI training and which remains posted today only as a “pre-publication version.”

A Justice Department official was designated Acting Registrar. He never took office: he and a colleague were turned away from the Copyright Office two days after that. The Office paused issuing registration certificates for twelve business days, affecting roughly 20,000 pending applications.

Perlmutter sued. The D.C. Circuit reinstated her that September, finding the removal likely unlawful, and last month the Supreme Court declined to remove her while the case proceeds. She remains Registrar of Copyrights. Meanwhile the framework she built became more settled: her office’s human-authorship position was endorsed by the D.C. Circuit and left standing by the Supreme Court’s refusal to hear Thaler.

The institution held. But it held by litigation, over fourteen months, and the merits are still unresolved. That’s not a reassuring answer to the question of whether the institutions administering attribution can be relied upon. It’s an answer about how much force it now takes to keep one in place.

The broader point is structural. The most sophisticated attribution framework is inert without institutions capable of implementing it. And institutional disruption — whether by political intervention, regulatory capture, resource starvation, or simple inattention — is itself a form of accountability failure.

#The structural tensions

Six tensions run through everything here:

Credit versus traceability. The old system rewarded claiming authorship. The emerging system may reward proving process. These are different incentive structures with different winners and losers. An artist who can demonstrate their creative process — sketches, drafts, iterative decisions — may be better positioned than one who produces finished work with no visible trail, regardless of quality. And provability requirements can be cumbersome and expensive. Movie studios have built provenance tools. Individual artists may not have the resources to document. This inequity could render the whole system unworkable, creating a parallel cheap undocumented market that subverts the documented market.

Transparency versus market penalty. Every disclosure regime asks creators to accept a potential market cost for honesty, and current research suggests that cost is real. Whether it’s durable or whether audiences will eventually stop caring about the human/AI distinction, as they stopped caring whether photographs were “real art,” is uncertain. The say-do gap in the research, stated preferences against AI content even as respondents actively use AI in their own work, suggests the penalty may be partly performative. But we don’t know yet.

Individual rights versus systemic infrastructure. Likeness protection works for recognizable individuals — celebrities, and established performers. Provenance infrastructure works at system scale. But the vast middle class of working artists, not famous enough for likeness claims yet whose work feeds training datasets, falls between both frameworks. This is the unaffiliated artist problem.

Enforceability versus aspiration. Some of what’s being built might work. Provenance infrastructure embedded at the point of creation might be enforceable because it operates upstream. Training data accountability is at least architecturally possible. But downstream disclosure — tell us whether AI was involved — depends on detection that’s failing, or self-reporting that incentives undermine. The policy conversation needs to be honest about which attribution functions can carry enforcement weight and which are merely aspirational. And it should notice that the mechanism currently carrying the most weight — the DMCA — was written in 1998 for an entirely different purpose.

Provenance versus excellence. Both matter, but they’re not the same claim. The case for valuing human origin is real and worth making on its own terms — not as a proxy for quality but as an independent value. Human experience, cultural situatedness and embodied perception shape what art means and why it matters. But building a regulatory architecture on the assumption that human-made and AI-made are distinguishable categories, when the technology and the evidence suggest they increasingly are not, is building on sand that’s shifting under your feet.

Supply-side attribution versus demand-side attribution. Nearly all the policy energy is going toward who made the work. Almost none is going toward whether the audience response is measurable or real. If the metrics that determine compensation, discovery, and cultural significance are themselves unreliable — synthetic audiences validating synthetic content — then even a perfect supply-side attribution system is feeding a broken measurement apparatus. The attribution chain runs in both directions, and the demand side is, if anything, further from a solution than the supply side.

#Where this is heading

The Observatory tracks 379 developments (as of July 29, 2026) across litigation, legislation, regulation, labor agreements, commercial deals, and research. Several trajectories are visible, though none is certain.

• Attribution is shifting from a system of credit to a system of infrastructure. The old model asked who gets the byline. The emerging model asks whether we can trace the chain of creation, verify the process, and identify the contributions at each layer. Content credentials, training data registries, California’s transparency laws — these are infrastructure investments, not individual protections.
• A gap is widening between jurisdictions building attribution infrastructure and those that aren’t. The EU and California are furthest ahead, though both have pushed their operative dates back. The US federal government has spent fourteen months litigating who runs its copyright office. Japan and South Korea have arrived at genuinely different places — Japan with a permissive exception narrowed by administrative guidance, South Korea with no exception and an unresolved fight about whether to create one. For artists working across borders, the result is an effectively impossible compliance environment, the rules changing depending on where the work is made, where the model was trained, and where the audience is.
• The artistic essence question is likely the next frontier. Can an artist’s accumulated creative identity be extracted and reproduced without consent or compensation? It has no legal framework and no clear path to one, outside that paragraph of Japanese guidance. But it’s the form of AI extraction most artists experience most directly, and the pressure to address it will only grow.
• The commercial deals are not waiting for the policy apparatus to catch up. Most of the major labels have shifted from suing AI companies to licensing their catalogs to them — though not all. Sony settled with neither of the two AI music companies its peers made deals with, and escalated instead. That “sue the pirates, license the allies” pivot is effectively setting attribution norms through private agreements that don’t represent independent creators. The risk is that by the time the policy framework is built, the commercial architecture will already be locked in. The rules of the road are effectively already written for the rest of us.

#The trust layer

Attribution and accountability is the trust layer — the system that establishes who made what, how it was made, and whether we can verify any of it.

Without that layer, the other two systems can’t function. You can’t compensate creative labor if you can’t identify the contributions or verify that the audience is real; that’s the Value and Exchange problem. You can’t sustain a healthy creative culture if you can’t verify what’s in it or measure what it means to the people who encounter it; that’s the Cultural Infrastructure problem.

But the trust layer itself is under construction, unevenly built, and resting on some assumptions the technology is already outrunning. This essay has tried to map what’s being built, what’s missing, and where the ground may not hold. Not to discourage the building, but to try to make sure it’s aimed at the right problems.

The decomposition of creative identity into separable, extractable layers is the structural reality that distinguishes this moment from what came before it. The recompilation of those layers into a functioning system of recognition and accountability is the challenge. And the distance between the two — between what’s being decomposed and what we’ve figured out how to recompile — is where the real work lies.

— July 30, 2026

View Sources