Fair Use Doctrine

Analyzes the characteristics of the original work itself. Published, factual works receive thinner copyright protection than unpublished, highly creative works. Training datasets composed primarily of factual compilations or public domain materials face a lower bar for fair use than those ingesting unpublished manuscripts or artistic creations. Key distinctions:

• Factual vs. creative: factual works are closer to the public domain
• Published vs. unpublished: unpublished works receive stronger protection
• The scope of fair use is narrower for unpublished works

Evaluates both the quantitative and qualitative extent of the material copied relative to the copyrighted work as a whole. Copying an entire work weighs against fair use, but even a small excerpt may be infringing if it constitutes the heart of the work. For AI training, ingesting complete works into a corpus is quantitatively substantial, but the qualitative analysis depends on whether the model reproduces the expressive core of specific works in its outputs. Key tests:

• How much was taken relative to the whole?
• Was the portion taken the most important or distinctive element?
• Copying an entire work is presumptively unfair

Assesses whether the unlicensed use harms the existing or potential market for the original work. This is often considered the most important factor. If the AI-generated output serves as a market substitute for the original—displacing sales, licensing revenue, or derivative work opportunities—it weighs heavily against fair use. Key inquiries:

• Does the use reduce demand for the original?
• Does it impair a traditional, reasonable, or likely-to-be-developed market?
• Widespread, unrestricted copying that substitutes for purchase is not fair use

The legal standard used to determine if an AI output has unlawfully copied protected expression. It compares the 'total concept and feel' of the original and the generated work, filtering out unprotectable elements like facts and ideas.

• Extrinsic Test: An objective analysis by experts dissecting the constituent creative elements.
• Intrinsic Test: A subjective determination by an ordinary observer regarding the overall impression.
• A model regurgitating a near-identical passage from its training data fails this test definitively.

The computational process of identifying AI-generated outputs that are substantially adapted from or directly copied from copyrighted source materials in a training corpus. This moves beyond exact matching to detect paraphrased or transformed infringement.

• Uses perceptual hashing (pHash) for multimedia to find visually similar copies.
• Employs semantic similarity scoring via embeddings to detect text that conveys the same meaning.
• Critical for enforcing licensing terms and providing automated attribution at scale.

A severe legal remedy that mandates the deletion of an entire model when it has been trained on unlawfully collected or infringing data. It is the nuclear option of copyright enforcement in the AI age.

• The FTC has ordered disgorgement of models built on improperly obtained data.
• It is technically complex, often requiring full retraining from a cleansed dataset.
• This remedy creates a powerful incentive for rigorous training data provenance and compliance.

A contractual and technical framework that protects enterprise users from copyright claims arising from retrieval-augmented generation (RAG) outputs. It combines indemnification with architectural safeguards.

• Indemnification Clause: The model provider assumes liability for claims from third-party content retrieved and generated.
• Technical Guardrails: Strict content filtering and source attribution are applied to all retrieved chunks before generation.
• Shields are a key differentiator for enterprise-grade AI platforms like Microsoft Copilot and Google Vertex AI.

A technical specification from the Coalition for Content Provenance and Authenticity that cryptographically binds provenance metadata to digital content. It establishes a verifiable chain of authorship and edit history.

• Uses cryptographic watermarking to sign content at creation.
• Records all subsequent modifications, including AI generation steps, in a tamper-evident manifest.
• Enables platforms to automatically display 'Content Credentials,' clarifying origin and deterring misuse of copyrighted material.