Derivative Work Doctrine

This factor examines whether the new work is transformative—adding new expression, meaning, or message—rather than merely supplanting the original. In AI training, courts assess whether the model's purpose is fundamentally different from the original creative works.

• Key test: Does the model generate content that competes with the original, or does it serve a distinct analytical function?
• Commercial vs. non-profit: Commercial use weighs against fair use, but transformativeness can override this.
• Landmark case: Authors Guild v. Google, Inc. (2015) found that scanning books to create a searchable index was transformative because it served a different purpose than reading the books themselves.

This factor considers whether the original work is factual or creative. Creative works receive stronger protection than factual compilations, making AI training on artistic content legally riskier.

• Published vs. unpublished: Unpublished works receive stronger protection against unauthorized use.
• Factual compilations: Training on databases of facts or scientific papers faces a lower bar for fair use than training on novels or music.
• Implication for AI: Models trained predominantly on creative works—fiction, visual art, music—face heightened scrutiny under this factor.

Courts evaluate both the quantitative and qualitative portion of the original work used. In AI training, this factor is complicated by the ingestion of entire works into training datasets.

• Whole-work copying: Training typically requires complete ingestion, which historically weighs against fair use.
• Qualitative heart: Even small excerpts can infringe if they capture the 'heart' of the work.
• Counter-argument: Intermediate copying that is necessary for a transformative purpose may be permitted, as established in Sega v. Accolade (1992) for reverse engineering.

Often considered the most important factor, this assesses whether the new use harms the actual or potential market for the original. AI-generated outputs that substitute for licensed content create significant legal exposure.

• Direct substitution: If a model can generate outputs that replace demand for the original—e.g., producing images in a specific artist's style—market harm is established.
• Licensing market: The emergence of AI training licenses (e.g., stock photo platforms offering ML licenses) demonstrates a viable market that unlicensed training may undermine.
• Burden of proof: Rights holders must demonstrate harm, but courts increasingly recognize the loss of licensing revenue as cognizable injury.

The transformative use doctrine is the central battleground for AI derivative work analysis. A use is transformative when it employs the original for a distinct purpose or imbues it with new character.

• Warhol Foundation v. Goldsmith (2023): The Supreme Court narrowed transformativeness, ruling that a new work must have a fundamentally different purpose, not merely a different aesthetic.
• AI training implications: The Court's emphasis on commercial purpose and market substitution directly threatens arguments that all ML training is inherently transformative.
• Open question: Whether training a model constitutes a transformative 'intermediate copying' step remains unresolved across circuits.

Even if training is deemed fair use, model outputs face independent scrutiny. The substantial similarity test evaluates whether generated content is substantially similar to protected expression in specific training examples.

• Filtering expression from ideas: Copyright protects expression, not ideas or facts. Outputs replicating unprotected elements do not infringe.
• Memorization risk: Large models can memorize and regurgitate training data verbatim, creating direct infringement liability for the deployer.
• Mitigation: Techniques such as differential privacy, output filtering, and deduplication of training data reduce the risk of producing substantially similar outputs.

A legal defense permitting limited use of copyrighted material without permission. Courts weigh four factors:

• Purpose and character of the use (transformative vs. commercial)
• Nature of the copyrighted work
• Amount and substantiality of the portion used
• Effect on the potential market

In AI training, the transformative use argument is central—whether a model's internal weights represent a new expressive purpose or merely a compressed copy.

The process of identifying which specific training examples most influenced a model's output. Techniques include:

• Influence functions that trace predictions back to training samples
• Nearest-neighbor retrieval in embedding space
• Data Shapley values for equitable contribution scoring

Critical for defending against infringement claims by demonstrating whether a generated output derives from a specific copyrighted source or represents generalized learning.

A documented, auditable trail establishing the origin, custody, and transformation history of every data point in a training corpus. Includes:

• Source URL and collection timestamp
• Licensing terms and consent status
• All preprocessing and filtering operations applied

Provenance records are the first line of defense in demonstrating that training data was lawfully acquired and properly licensed.

Technical protocols allowing rights holders to exclude their content from AI training pipelines. Common implementations:

• robots.txt directives targeting AI crawlers specifically
• HTML meta tags signaling no-AI-training permissions
• Do Not Train (DNT) registry submissions

The legal weight of these signals under the derivative work doctrine remains unsettled, but they establish a clear expression of the rights holder's intent.

The technical process of removing the influence of specific data from a trained model without full retraining. Approaches include:

• Exact unlearning via data partitioning and selective retraining
• Approximate unlearning using gradient-based scrubbing
• Differential privacy guarantees of removal

Directly relevant when a copyright holder successfully asserts rights over training data and demands removal of their work's influence from model weights.

Embedding verifiable identifiers into model weights or outputs to prove intellectual property ownership. Techniques include:

• Backdoor watermarks triggered by specific inputs
• Weight-space fingerprints embedded during training
• Output steganography detectable by the owner

Serves as a deterrent against unauthorized derivative models and provides evidence in IP disputes over model lineage.