Training Governance for Creative AI

Regulatory Framework

The legal status of creative AI training is being defined simultaneously across at least four jurisdictions, and the resulting framework already constrains how training pipelines must be built. In the United States, 17 USC 107 fair use is the primary doctrinal lens for training on copyrighted creative works. The four-factor analysis weighs the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality of the portion used, and the effect on the market for the original. The pending litigation matters not only for its outcomes but for the discovery it has produced: Andersen v. Stability AI established that training on identifiable artist works without consent is at least colorably actionable, Getty Images v. Stability AI introduced the question of whether watermark reproduction in outputs is direct evidence of memorization, and New York Times v. OpenAI placed verbatim regurgitation of training text at the center of the fair-use balance. Each of these cases turns on whether the model encoded generalizable patterns or specific protected expression.

In the European Union, Articles 3 and 4 of the Copyright in the Digital Single Market Directive (2019/790) define the text-and-data-mining exceptions. Article 3 provides a non-overridable TDM exception for research organizations and cultural heritage institutions. Article 4 provides a broader TDM exception that applies unless the rights holder has expressly reserved the right in a machine-readable manner. Commercial creative AI training depends on Article 4, and Article 4 depends on the training pipeline being able to honor opt-out reservations on a per-work basis. The General Data Protection Regulation Article 22 layers an additional constraint when training data includes personal data, and the European Data Protection Board's Opinion 28/2024 on AI models clarified that training, deployment, and output stages each carry distinct lawful-basis obligations and that legitimate interest balancing must account for memorization risk explicitly.

Japan's Copyright Act Article 30-4, often cited as the most permissive global regime for AI training, permits use of copyrighted works for information analysis but only when the use is not for the purpose of enjoying the expression of the work and when it does not unreasonably prejudice the rights holder's interests. The unreasonable-prejudice qualifier has been read by Japanese commentators to exclude training that produces models capable of reproducing the work itself. China's Interim Measures for the Management of Generative Artificial Intelligence Services impose direct obligations on training data lawfulness, intellectual property respect, and output controls, with the Cyberspace Administration of China requiring filings that include training data provenance. United States Executive Order 14110, while procedural in form, directs federal agencies to develop guidance on training data provenance, watermarking, and dual-use concerns, and the resulting NIST and Copyright Office guidance is rapidly converging on per-example provenance as a prerequisite for federal use.

Architectural Requirement

The architectural requirement that follows from this framework has three components. First, the training pipeline must be able to control the depth at which a specific work influences the model. Fair use and the TDM exceptions tolerate generalizable learning but not memorization, which means the pipeline must be able to demonstrate, per work, that the influence on model parameters was bounded below the memorization threshold. Second, the pipeline must produce provenance records sufficient to answer per-example questions in litigation, in regulatory review, and in marketplace platform compliance audits. Third, the pipeline must honor rights metadata, including TDM opt-out reservations, license terms, and personal-data flags, in a manner that affects training rather than only being filtered upstream.

These requirements are not satisfied by curating training data alone. Two pipelines trained on identical, fully-licensed data can produce models with radically different memorization profiles depending on epoch count, batch composition, learning rate schedule, and parameter capacity. The same work in the same dataset can be learned generalizably in one configuration and memorized in another. A compliance posture that asserts only that all training data was licensed does not answer the question that fair-use balancing, Article 4 TDM enforcement, Article 30-4 unreasonable-prejudice analysis, and EDPB Opinion 28/2024 memorization analysis are all asking. The training process itself must be the locus of compliance.

Why Procedural Compliance Fails

The dominant procedural responses to creative AI training rights have been licensing aggregation, opt-out registries, and post-hoc output filtering. Each addresses a real concern, and none is sufficient on its own. Licensing aggregation acquires the right to use a corpus of works for training but does not address how those works are learned, leaving the licensee exposed to claims that the licensed works were memorized rather than learned from. Opt-out registries, including the emerging implementations of the EU Article 4 reservation mechanism, address which works enter the pipeline but do not control depth of influence on the works that remain. Output filtering, in which generated outputs are checked against a database of training works, catches verbatim reproduction at inference time but cannot retroactively un-memorize the model and does not address the structural infringement claim that the model itself is a derivative work.

Procedural compliance also fails because it produces the wrong evidentiary record. In the Andersen, Getty, and New York Times litigation, the most damaging evidence has not been data-acquisition contracts; it has been demonstrations that the model itself reproduces specific protected expression. A licensing record establishes that the input was lawful at the moment of acquisition but does not establish that the output was non-infringing at the moment of generation. A pipeline that cannot produce per-example depth-of-influence evidence cannot rebut a memorization claim with anything stronger than aggregate statistics, and aggregate statistics do not satisfy a court asking whether this specific output infringed this specific work.

The EDPB's Opinion 28/2024 sharpened this point in the GDPR context. The Opinion held that the lawful basis analysis for training on personal data must account for the risk that the trained model itself constitutes processing of personal data, which depends on whether the model memorized the data or learned generalizably from it. A pipeline that cannot demonstrate the depth of influence on a per-record basis cannot support a legitimate-interest balancing test under Article 6(1)(f) and cannot establish the safeguards that would make Article 22 automated-decision review tractable. Procedural compliance treats training as a black box; the regulatory framework increasingly requires that the box be opened.

What AQ Primitive Provides

The AQ training-governance primitive treats training depth as a controlled variable. Gradient updates are routed by a depth-selective mechanism that distinguishes structural and stylistic learning, which is permitted to flow to deep layers and form foundational model knowledge, from specific-content learning, which is restricted to surface layers and bounded in its parameter influence. The classification between structural and specific is not a manual annotation; it is derived continuously during training from the entropy profile of the model's representation of each example. When a representation collapses toward the example itself, indicating memorization, the gradient depth for that example is reduced and the influence is bounded.

Memorization detection operates as a closed-loop control on the training process. Each training example carries an entropy signature that the system tracks across epochs. Examples whose signature trends toward memorization are progressively routed to shallower depth, and in the limit are dropped from the loss surface for parameters above a threshold. This produces a model in which generalizable principles are deeply learned and specific works are not, by construction rather than by hope.

Rights-metadata-aware routing integrates licensing, TDM opt-out, and personal-data flags into the gradient routing decision. A work with a permissive license and no TDM reservation is eligible for the standard depth schedule. A work covered by an Article 4 reservation is excluded entirely. A work licensed for stylistic training but not for content reproduction is routed to surface depth with hard caps on parameter influence. Personal data is handled by an additional routing layer that enforces EDPB Opinion 28/2024 safeguards, including memorization bounds tight enough to support the legitimate-interest balancing test required by GDPR Article 6(1)(f) and to render Article 22 review tractable.

Provenance tracing produces the per-example record that the regulatory framework requires. For each training example, the system records the depth schedule applied, the entropy trajectory observed, the rights metadata that governed the routing decision, and the cumulative parameter influence. At inference time, the system can map a generated output back to the training examples that most influenced it, the depth at which they influenced it, and the rights metadata under which they were trained. This is the evidentiary record that fair-use analysis under 17 USC 107, Article 4 TDM enforcement, Article 30-4 unreasonable-prejudice analysis, and EDPB Opinion 28/2024 review all require, produced as a byproduct of training rather than reconstructed under litigation pressure.

The primitive also supports the operational compliance regimes that creative AI platforms face. China's Generative AI Service regulations require filings that include training data provenance; the per-example record is the filing substrate. United States Executive Order 14110 directs agencies toward provenance-based assurance; the per-example record satisfies the emerging NIST and Copyright Office guidance directly. Marketplace platforms imposing rights-compliance requirements on AI-generated content can verify the provenance trace as part of their listing review, shifting compliance from declaration to demonstration.

Compliance Mapping

The training-governance primitive maps to each component of the regulatory framework at the architectural level. For 17 USC 107 fair use, the per-example depth-of-influence record provides direct evidence on the third statutory factor (amount and substantiality used) and the fourth factor (market effect), because depth-bounded learning demonstrably does not produce a market substitute for the original work in the way that memorized reproduction does. The Andersen, Getty, and New York Times memorization arguments are met not by denial but by structural evidence that the architecture made memorization unavailable for the works in question.

For EU Copyright Directive Article 4, the rights-metadata-aware routing honors TDM reservations on a per-work basis, and the provenance trace demonstrates compliance with the reservation at the level the Directive contemplates. For Article 3 research-organization use, the depth record supports the non-commercial purpose limitation. For GDPR Article 22 and EDPB Opinion 28/2024, the memorization bounds and per-record influence record support the safeguards analysis and the legitimate-interest balancing test, addressing the Opinion's specific concern that training memorization can render the model itself a personal-data processing artifact.

For Japan's Copyright Act Article 30-4, the depth bounds support the unreasonable-prejudice qualifier by demonstrating that the model is not capable of reproducing the work itself. For China's Generative AI Service regulations, the provenance record provides the training-data lineage that the Cyberspace Administration filings require. For US Executive Order 14110 and the resulting NIST guidance, the per-example provenance record provides the assurance substrate that the EO directs agencies to require for federal use.

Adoption Pathway

Adoption proceeds in three phases corresponding to the platform's existing training infrastructure and the regulatory cadence it faces. In the first phase, the platform instruments its existing training pipeline with the entropy-tracking and provenance-recording layers of the primitive, operating in observe-only mode. The output is a per-example memorization profile for the platform's current models and a baseline against which depth-routed retraining can be measured. This phase produces immediate value for litigation posture and for EDPB Opinion 28/2024 documentation without changing the model.

In the second phase, the platform enables depth-selective routing for new training runs. Rights metadata is integrated into the data pipeline, including TDM reservations, license terms, and personal-data flags, and gradient routing is conditioned on those metadata. The resulting models are demonstrably depth-bounded for the works that require it, and the platform can offer marketplace partners and licensees a provenance trace as part of model delivery. China filing obligations and EO 14110 federal-use requirements become satisfiable at the model level rather than by external attestation.

In the third phase, the platform's entire training estate operates under training governance, and the provenance trace becomes the platform's primary compliance artifact. Fair-use posture is grounded in per-example depth evidence rather than aggregate licensing claims. Article 4 TDM compliance is verifiable against the platform's reservation-honoring record. Article 30-4 unreasonable-prejudice analysis is supported by the architecture itself. The platform's creative AI development continues, with the rights framework encoded in the training process rather than applied as a post-hoc filter, and the model lineage that emerges supports the licensing, provenance, and accountability obligations that the next phase of creative AI regulation will impose.