Fair Use Doctrine

Examines how the copyrighted work is being used. The central question is whether the new work is transformative—adding new expression, meaning, or message—rather than merely superseding the original.

• Commercial vs. Nonprofit: Nonprofit educational uses are favored, but commercial use is not presumptively unfair.
• Transformative Use: A critical inquiry; parody, criticism, and search engine indexing have been found transformative.
• Key Case: Campbell v. Acuff-Rose Music, Inc. (1994) established that a commercial parody can be fair use if it comments on the original.

A cornerstone of copyright law determining whether a new creation—such as an AI model trained on copyrighted data—constitutes a transformative use or an infringing copy. Courts assess whether the new work adds new expression, meaning, or message. In AI contexts, the central question is whether a trained model's weights represent a derivative of the training corpus. The Supreme Court's Warhol Foundation v. Goldsmith decision narrowed transformativeness, requiring that the new work serve a fundamentally different purpose rather than merely conveying a different aesthetic.

The process of identifying the specific source or subset of training data responsible for a model's particular output. This is technically challenging due to the distributed nature of neural network learning. Key approaches include:

• Influence functions: Tracing a prediction back to training examples that most shaped it
• Data Shapley values: Quantifying each training sample's marginal contribution
• Near-neighbor retrieval: Identifying training samples closest to a generated output in embedding space Accurate attribution is critical for copyright compliance and debugging model behavior.

A technical or legal process allowing rights holders to exclude their data from AI training. Common implementations include:

• robots.txt directives: Website-level signals blocking crawlers from scraping content
• Do Not Train (DNT) headers: Proposed HTTP standard for machine-readable opt-out signals
• Dataset de-listing requests: Formal takedown procedures for datasets like LAION
• Model unlearning: Post-hoc removal of data influence from already-trained models The EU AI Act mandates that providers respect opt-outs, making this a critical compliance infrastructure component.

A documented trail establishing the origin, custody, and transformations of a dataset. For AI governance, provenance answers: Who collected this data? Under what license? Has it been filtered or augmented? Key standards include:

• W3C PROV: A family of specifications for representing provenance on the web
• Data Cards: Structured transparency documents describing dataset motivation and composition
• Cryptographic hashing: Ensuring dataset integrity across pipeline stages Provenance is the foundation for auditing whether training data was lawfully acquired.

The technical process of removing the influence of specific training data from a trained model without full retraining. This is directly relevant when a rights holder successfully asserts that their copyrighted material was used without permission. Approaches include:

• SISA (Sharded, Isolated, Sliced, Aggregated): Training on data shards so only affected shards need retraining
• Gradient-based scrubbing: Applying inverse updates to approximate removal
• Exact unlearning via retraining: The only provably complete method, but computationally prohibitive Unlearning remains an open research problem with no production-ready solution for large models.

The robots.txt standard allows website operators to signal which parts of their site should not be crawled. Originally designed for search engines, it has become a de facto mechanism for expressing data preferences to AI scrapers. Key considerations:

• No legal force: robots.txt is a voluntary protocol, not a binding contract
• Crawler-specific rules: Different User-agent directives can target specific AI crawlers (e.g., GPTBot, CCBot)
• Sitemap integration: Complementary XML sitemaps guide legitimate crawlers to permitted content Major AI developers including OpenAI and Google now publish dedicated crawler user-agent tokens.