The Provenance Ledger: Auditing AI's Memory and Paying Its Teachers

The problem: a model with no memory of its teachers

Every large language model is a student that has forgotten every teacher it ever had. The training data is ingested, compressed into weights, and discarded. The model can tell you about quantum mechanics, but it cannot tell you which textbook it learned from, whether the author consented, or whether a licensing fee was paid. This is not a feature. It is a structural failure with consequences that are now unfolding in courts, legislatures, and the market.

The U.S. Copyright Office released Part 3 of its AI report in May 2025, concluding that "some uses of copyrighted works for generative AI training will be infringing." The report stopped short of recommending a blanket licensing regime, but it established the legal framework: fair use is not a shield, and the burden of proof is shifting toward AI companies to demonstrate that their training data was lawfully acquired. The lawsuits are already here — $1.5 billion from Anthropic, $3.1 billion from Universal Music, and more coming.

What provenance tracking means

Provenance tracking is simple to describe and hard to build: every training sample is tagged with its source, license, and acquisition method. The model can report which sources contributed to a given output. The training pipeline maintains an auditable chain of custody from creator to weights to inference. This is the foundation of accountability — you can't trust what you can't audit.

The concept is not hypothetical. The Coalition for Content Provenance and Authenticity (C2PA), founded by Adobe, Arm, Intel, Microsoft, and Truepic, has published a technical specification (v2.3, December 2025) for Content Credentials — cryptographically signed manifests that travel inside digital files and record who made them, when, and what tools were used. The manifest is based on X.509 certificates, CBOR (RFC 8949), and JUMBF (ISO 19566-5) — established, battle-tested standards. It is tamper-evident and designed for global, opt-in adoption.

But C2PA solves the content authenticity problem — proving that a file hasn't been tampered with. It doesn't solve the training data provenance problem: proving that a model's weights were derived from specific source materials, and attributing outputs back to those sources. That requires a different architecture, one that connects the content credential to the training pipeline to the inference output.

The attribution breakthrough: Data Shapley

For years, the obstacle to paying creators for training data was not economic but technical. You could estimate the aggregate value of data — roughly 45% of AI value created, according to scaling law analysis — but you couldn't attribute that value to individual contributors. The machine learning community solved this problem by accident.

Researchers building tools to find mislabeled data, prune bad examples, and debug training runs developed what is now called In-Run Data Shapley. By 2025, it could compute every training point's contribution to a model's performance during a single ordinary training run, at negligible overhead. By 2026, federated variants were being marketed for real-time data pricing. The attribution wall had fallen.

The payment formula is now computable. An individual's annual payment is the product of two terms: data's share of AI revenue (measurable from scaling laws) and the individual's fraction of attributed contribution (measurable from Data Shapley). The economics and the engineering have finally met. And the result is counterintuitive: the people with the strongest individual claims are not influencers with large followings but specialists — the rare-disease patient whose case notes taught the model something no textbook contains, the speaker of a low-resource language, the engineer whose niche documentation is the model's only source on a topic. A data market priced by attribution pays expertise and rarity, not volume and virality.

The provenance ledger architecture

Building a working provenance tracking system requires five connected layers. Each layer exists in prototype or production today. The work is in connecting them.

Getting creators paid and in control

The market is already moving. Kled AI signed a $12 million non-exclusive data licensing deal in June 2026, making the marketplace officially profitable. Top contributors earn roughly $7,000 per month. The platform's Proof-of-Data mechanism records a cryptographic hash of each submission on Solana, creating an immutable record of data provenance. Sebastian Thrun (founder of Waymo), Aglaé Ventures (Bernard Arnault's venture fund), and K5 Global backed the $5.5 million seed round.

But the current market is just the beginning. The $34 billion annual gap between data's value to AI (roughly 45% of AI revenue) and what creators are paid (roughly 2%) is the fastest-growing asset class in the world. Data is nonrival — using it to train one model doesn't use it up for the next. The same contribution can sit in the training corpus of every frontier model, every fine-tune, every distilled variant, simultaneously. A rival input's value gets divided across uses. A nonrival input's value gets summed. The liability compounds with each training run.

Creator control means more than payment. It means consent, revocation, and granularity. A creator should be able to: opt in to specific training uses (medical research yes, military applications no); set their own price floor; revoke consent and have their data removed from future training runs (the provenance ledger makes this tractable — you know exactly which weights to retrain); and audit how their data has been used across all models that trained on it. The C2PA manifest carries the license terms. The ledger enforces them.

Integration with the Library of Congress

The U.S. Copyright Office, administered by the Library of Congress, released its three-part AI report between 2024 and 2025. Part 1 recommended federal legislation against unauthorized digital replicas. Part 2 addressed copyrightability of AI-generated outputs. Part 3, released May 2025, addressed the training question directly — concluding that fair use analysis depends on the purpose, nature, and market effect of the use, and that training on copyrighted works without authorization is not categorically protected.

The provenance ledger is the infrastructure the Copyright Office needs to enforce whatever licensing regime Congress ultimately enacts. Here's how integration would work:

The integration is straightforward in principle: when a creator registers a copyright with the Library of Congress, the registration number, deposit metadata, and license terms are written into the C2PA manifest. When the work enters a training pipeline, the manifest is registered on the provenance ledger with the LoC registration number as a foreign key. The AI model's provenance index can then trace any output back to a specific copyright registration. The Copyright Office's public catalog API becomes the source of truth for ownership, and the ledger becomes the enforcement mechanism for licensing.

The semantic ontology layer

Provenance tracking tells you who contributed. A semantic ontology tells you how they relate. This is where the system becomes more than a payment rail — it becomes a knowledge architecture.

The idea is to organize every training sample not just by source and license but by its semantic relationship to every other sample in the corpus. A knowledge graph where nodes are works (articles, books, datasets, code repositories) and edges are relationships (cites, extends, refutes, translates, derives from, complements). This is not a new concept — knowledge graphs have been in production use for years, and 2026 has seen ontologies emerge as the missing layer of meaning that turns generative AI from a fluent talker into a system that actually understands domain context.

When the model generates an output, it doesn't just report which sources contributed — it reports how those sources relate to each other. "This response synthesizes Creator A's textbook, which Creator B cited and Creator C extended, while Creator D published a refutation that the model weighted at 15%." The semantic graph makes the model's reasoning traceable through the network of human knowledge that produced it. It turns the training corpus from a flat list of files into a structured map of human understanding.

This has practical consequences for creators. If your work is a node in the graph, you can see every other work that cites, extends, refutes, or derives from yours. You can discover collaborators working in adjacent domains. You can identify unauthorized derivations. You can track the influence of your work as it propagates through the knowledge graph — not as citation count, but as Shapley-weighted contribution to every model that trained on it.

The future of patents and IP ownership

The U.S. Patent and Trademark Office (USPTO) has been actively engaging with AI. In 2024, it issued guidance on AI-assisted inventions, clarifying that a natural person must have "significantly contributed" to each claim. In 2025, Director Squires signed new guidance on subject matter eligibility under § 101. The patent outlook for AI-related inventions is described as "promising" for 2026. But the USPTO's challenge is not just about patenting AI inventions — it's about building the infrastructure to track IP ownership in a world where AI models are trained on millions of copyrighted works.

The provenance ledger is the IP ownership ledger the USPTO and Copyright Office have been waiting for. Here's the vision:

The trust boundary this enables

With provenance tracking in place, the trust boundary from the previous transmission gains teeth. When a model reports its attribution, the user can verify the claim against the ledger. When a creator contests an output, the system can trace the contribution through the provenance index. When a regulator audits a model, the full chain of custody is available — from creator to C2PA manifest to ledger registration to training run to Shapley score to output attribution.

The hard problems

The path forward

The pieces are converging. C2PA provides the content credentialing layer. The DATA Foundation and Trace provide the ledger infrastructure. Data Shapley provides the attribution method. Kled AI proves the payment market works. The Library of Congress and U.S. Copyright Office are building the legal framework. The USPTO is engaging with AI patent questions. The semantic web community has the ontology tools. What's missing is the integration — the connective tissue that links these systems into a single accountability stack.

The first entity to build that integration wins. Not because it's a competitive advantage, but because it's the infrastructure the entire AI industry needs to survive the legal and ethical crisis it has created. The companies that can prove their training data was lawfully acquired, attribute their outputs to specific sources, and pay creators automatically will be the ones that survive the regulatory wave. The ones that can't will be the ones in court.

The distillation farmers are eating their own crop. The provenance ledger is how we make sure the next generation of AI is grown in soil we can see, from seeds we can trace, by farmers we can pay.