Transformative Use Analysis

The core inquiry examines whether the new work is transformative—adding new expression, meaning, or message—rather than merely superseding the original. Courts assess if the use serves a different purpose (e.g., indexing for search vs. artistic display) or character (commercial vs. non-profit educational).

• Transformative threshold: Does the AI output alter the original with new aesthetics, insights, or understandings?
• Commerciality: Purely commercial use weighs against fair use, but is not dispositive if the work is highly transformative.
• Generative AI context: Training a model on copyrighted data to create a statistical representation of language patterns may be considered transformative, while generating outputs that compete with the original market is not.

This factor considers the degree of creativity in the original work and whether it is published or unpublished. Creative works (novels, films, art) receive stronger protection than factual compilations (databases, news reports).

• Published vs. unpublished: Use of unpublished works is heavily disfavored; the author's right to control first publication is paramount.
• AI training implications: Models trained on highly creative, expressive corpora face greater scrutiny than those trained on factual, scientific, or functional texts.
• Thin copyright: Works with limited creative expression (e.g., simple data tables) receive only thin protection, making transformative use easier to establish.

Courts evaluate both the quantitative volume and qualitative significance of the portion taken relative to the copyrighted work as a whole. Copying the heart of a work—its most distinctive or memorable elements—weighs against fair use even if the amount is small.

• Quantitative analysis: Copying an entire work is presumptively unfair, though complete copying may be justified for certain transformative purposes (e.g., search indexing, parody).
• Qualitative analysis: Extracting the most expressive, creative, or commercially valuable segments undermines fair use claims.
• Training data context: Ingesting full texts for language model pre-training raises substantiality concerns, though the model's internal representation is a statistical abstraction rather than a stored copy.

The most influential factor examines whether the new use harms the actual or potential market for the original work. If the AI-generated output serves as a market substitute—displacing demand for the original—fair use is unlikely.

• Market substitution test: Does the generative output fulfill the same demand as the original? A summary that replaces the need to read the book is harmful; a brief excerpt in a critical review is not.
• Derivative market harm: Courts consider harm to licensing markets the copyright holder might reasonably develop (e.g., translation rights, film adaptation).
• AI-specific concerns: Generative models that produce outputs competing with training data sources—such as generating code from open-source repositories or artwork from artist portfolios—face significant market harm arguments.

Key legal decisions shaping how transformative use applies to machine learning systems:

• Authors Guild v. Google (2015): Google's book-scanning project was held transformative because it created a searchable index and text-mining corpus rather than serving as a reading substitute. This precedent strongly supports AI training as fair use.
• Andy Warhol Foundation v. Goldsmith (2023): The Supreme Court narrowed transformative use, holding that a new work must have a distinct purpose or character—not merely a different aesthetic. Commercial licensing of AI outputs that serve the same market as originals faces heightened scrutiny.
• Pending litigation: Cases against Stability AI, OpenAI, and Microsoft are actively testing whether model training and output generation constitute transformative use or mass infringement.

No single factor is determinative; courts weigh all four factors together in a case-by-case, equitable balancing test. The transformative use analysis is the heart of this inquiry but must be considered alongside the other statutory factors.

• Interplay of factors: A highly transformative use may outweigh commercial purpose and substantial copying. Conversely, minimal transformation combined with market harm is fatal.
• Burden of proof: The party asserting fair use bears the burden of demonstrating all factors support the defense.
• Enterprise risk management: Organizations deploying generative AI should conduct multi-factor fair use assessments for training data and output generation, documenting the analysis to support legal defensibility.

A legal standard applied after a finding of factual copying to determine if the copying is legally actionable. Courts compare the total concept and feel of two works, filtering out unprotectable elements like ideas, facts, and scènes à faire. In AI litigation, this test determines whether an output is an infringing derivative work or a sufficiently distant transformation of the training data.

Computational methods for identifying AI outputs that are substantially similar to copyrighted source materials. Techniques include: embedding similarity comparisons in latent space, n-gram overlap analysis for text, and perceptual hashing for images. These tools provide quantitative evidence for the substantial similarity and market effect prongs of fair use analysis.

The documented chain of custody establishing the legal rights and licensing status of all content ingested during model training. Robust provenance records enable courts to trace whether an AI output derives from a specific copyrighted input. Without this lineage data, transformative use analysis becomes speculative, as the relationship between training data and generated output cannot be definitively established.

A legal remedy requiring the deletion of models trained on unlawfully collected or infringing data. When a court finds that an AI system's use of copyrighted material is not transformative and thus infringing, disgorgement may be ordered. This forces the destruction of tainted model weights, representing the most severe consequence of failing the transformative use test.

A contractual and technical indemnification framework protecting enterprise users from infringement claims arising from retrieval-augmented generation. By grounding outputs in licensed, auditable data sources rather than model weights, RAG architectures shift the legal analysis: the retrieval mechanism's purpose and the cited nature of the source material become central to the transformative use inquiry.