Position: No Retroactive Cure for Infringement during Training

Satoru Utsunomiya, Masaru Isonuma, Junichiro Mori, Ichiro SakataarXiv 2026 · arXiv

TL;DR

Post-hoc AI safety measures like machine unlearning and inference-time guardrails do not erase legal liability incurred during unauthorized training data ingestion. Because copyright, contract, and tort liabilities attach the moment data is copied and processed, developers must shift from retroactive sanitization to verifiable, upfront process compliance.

What Problem Does This Paper Address?

The machine learning community frequently relies on an engineering assumption the authors call "output equivalence": the belief that if a model can be patched (via unlearning, fine-tuning, or guardrails) so that its output distribution is statistically indistinguishable from a model trained without problematic data, the system is legally compliant.

This paper exposes a massive epistemic gap between this engineering intuition and legal reality. The authors address whether post-hoc fixes can retroactively cure the legal liabilities established during the upstream data scraping and training phases. They argue that compliance is fundamentally about data lineage (how a model came to know something), not just product behavior (what a model says).

Methods

As a position paper, this work does not introduce a new empirical algorithm; instead, it provides a rigorous interdisciplinary legal-technical analysis. It maps current engineering mitigation techniques—divided into Inference-Time Control (e.g., safety filters, Llama Guard) and Model Modification (e.g., SISA, approximate unlearning, model editing)—against established legal doctrines across the US, EU, and Japan:

  • Completed-Act Doctrine: Demonstrating that copyright infringement is a strict-liability wrong completed the moment data is downloaded ($t=0$) or loaded into RAM.
  • Fixation and Perceptibility Analysis: Evaluating model weights under 17 U.S.C. §101 to show they act as stable, persistent copies where protected expressions remain technically recoverable via forensic extraction attacks.
  • Global Encirclement Theory: Analyzing how overlapping liabilities from breach of contract (violating website Terms of Service) and torts (unfair competition/free-riding) independently restrict data use, bypassing traditional copyright defenses like fair use.
  • Head Start Doctrine: Examining how remedies for unjust enrichment strip away structural market advantages gained from wrongful shortcuts, which can result in court-mandated algorithmic disgorgement (the destruction of the model itself).

To bridge this gap, the authors propose a technical framework for Ex-ante Process Compliance, introducing concepts like a "Glass Pipeline" (using C2PA standards and Zero-Knowledge Proofs for data verification) and "Architectural Reversibility" (a "Git for Models" utilizing high-frequency checkpoint rollbacks instead of unlearning).

Datasets

Not discussed in the paper. (The text references existing controversial corpora like Books3 and The Pile to ground its legal analysis of unlicensed datasets, but it does not introduce or utilize datasets for empirical training).

Key Insights and Ideas

  • The Fallacy of Output Equivalence: Just because an engineered model no longer outputs copyrighted material does not mean the historical act of copying is erased. The law views compliance through data lineage, making downstream filtering beside the point for initial infringement.
  • Infringement as a Fait Accompli: The law treats unauthorized dataset creation analogously to physical theft. A thief who returns a stolen car might receive a lighter sentence, but they cannot un-commit the theft. Similarly, deleting a dataset post-training does not nullify the initial statutory violation.
  • Weights as Ongoing Infringing Copies: Even if an external safety filter blocks a model from displaying a copyrighted work, the expressive value remains embedded within the static weight matrices on a hard drive. Because this data is machine-recoverable via extraction attacks, the weights themselves legally constitute unauthorized copies.
  • The Contractual Trap: Even if developers win the "fair use" argument for copyright, they remain heavily exposed to contract law. Scraping data in violation of browsewrap agreements or website Terms of Service (ToS) creates independent liability that post-hoc model editing cannot fix.
  • The Head Start Advantage: Unlawful data accelerates model convergence and time-to-market. Because this structural advantage is retained in the parameters even if the raw source files are deleted, regulators can demand total model destruction (algorithmic disgorgement) to prevent firms from profiting from misappropriation.

Key Results

Not discussed in the paper. (As a conceptual position paper, it contains no quantitative experiments or statistical benchmarks).

Limitations and Future Work

  • Unsettled Legal Landscape: The authors acknowledge that the global legal framework surrounding AI training is still actively evolving and subject to shifting judicial scrutiny across different jurisdictions.
  • Technical and Economic Friction: Implementing the proposed Ex-ante Process Compliance framework introduces significant overhead. High-frequency checkpointing and surgical re-training ("Git for Models") are computationally expensive, and establishing centralized "Data Trusts" requires massive market restructuring.
  • Future Directions: The authors call for policymakers to shift the legal default from "Opt-Out" to "Opt-In" by establishing a statutory "Binary Beacon" (a mandatory, machine-readable signal declaring AI training permission). They emphasize that future work must focus on building architectures that ensure verifiable data lineage from day one rather than retroactively patching leaks.

Summary updated

Abstract

As generative AI faces intensifying legal challenges, the machine learning community has increasingly relied on post-hoc mitigation -- especially machine unlearning and inference-time guardrails -- to argue for compliance. This paper argues that such post-hoc mitigation methods cannot retroactively cure liability from unlawful acquisition and training, because compliance hinges on data lineage, not the outputs. Our argument has three parts. First, unauthorized copying/ingestion can be a legally complete completed act, and model weights may operate as fixed copies that retain training-derived expressive value, making later filtering beside the point for infringement. Second, contract and tort/unfair-competition rules -- via licenses, terms of service, and anti-free-riding principles -- can independently restrict access and use, often bypassing copyright defenses (e.g., fair use or TDM exceptions). Third, since value from protected inputs can persist in weights, remedies such as unjust enrichment and disgorgement may require stripping gains and, in some cases, reaching the model itself. We therefore argue for a shift from Post-Hoc Sanitization to verifiable Ex-Ante Process Compliance.
BibTeX
@misc{utsunomiya2026position,
  author = {Satoru Utsunomiya and Masaru Isonuma and Junichiro Mori and Ichiro Sakata},
  title = {{Position: No Retroactive Cure for Infringement during Training}},
  year = {2026},
  eprint = {2604.18649},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.18649},
}
Figure