Substantial Similarity Test

A structured judicial method for separating unprotectable ideas from protectable expression in software and AI cases. The process unfolds in three stages:

• Abstraction: Deconstruct the work into levels of generality, from specific code to high-level functions.
• Filtration: Eliminate unprotectable elements—scenes a faire, merger doctrine material, public domain components, and functional processes.
• Comparison: Compare only the remaining protected expression for substantial similarity.

This test, established in Computer Associates v. Altai, is the dominant framework for analyzing non-literal copying in complex systems.

A holistic infringement analysis that evaluates whether an ordinary observer would recognize the aesthetic appeal and overall impression of the original work in the allegedly infringing output.

Key considerations include:

• Selection and arrangement of elements rather than isolated components
• Pacing, mood, and character in literary or audiovisual works
• Compositional choices in visual art

This doctrine is particularly relevant when AI models generate outputs that capture the distinctive style or structural essence of a copyrighted work without verbatim copying.

A two-part analysis used primarily in the Ninth Circuit to evaluate substantial similarity:

Extrinsic Test (Objective)

• Expert analysis of specific, identifiable criteria
• Breakdown of discrete elements: plot, theme, dialogue, mood, setting, pace, characters, sequence of events
• Relies on analytic dissection and expert testimony

Intrinsic Test (Subjective)

• Ordinary reasonable person's impression
• No expert assistance or analytical dissection permitted
• Focuses on whether the total concept and feel are substantially similar

Both prongs must be satisfied for a finding of infringement.

A legal doctrine holding that trivial or microscopic taking of copyrighted material does not constitute actionable infringement. In AI contexts, this applies when:

• The copied portion is so small that an ordinary observer would not recognize the appropriation
• The quantitative and qualitative significance of the copied material is negligible
• The fragment constitutes unprotectable factual data rather than creative expression

This defense is frequently invoked when models incidentally reproduce short phrases or common code patterns that fall below the threshold of substantial similarity.

A controversial judicial principle establishing an inverse relationship between the degree of access to a copyrighted work and the level of substantial similarity required to prove infringement.

Practical implications:

• High access + lower similarity = May still find infringement
• Low access + higher similarity = Required to prove copying

In AI training contexts, where models have ingested vast corpora, courts may find pervasive access to copyrighted works, potentially lowering the similarity threshold needed to establish infringement.

A copyright principle that excludes from protection elements that are standard, stock, or common to a particular genre or topic. These elements are considered indispensable to the treatment of a given subject matter.

Examples in AI outputs:

• Generic code patterns and standard algorithms
• Common visual tropes in image generation
• Conventional narrative structures and character archetypes

When evaluating AI-generated content for substantial similarity, courts filter out scenes a faire material as unprotectable, focusing only on original creative choices that go beyond genre conventions.

A judicial test examining whether a new work merely supersedes the original or adds something new with a further purpose or different character. Key considerations include:

• Does the AI output comment on, criticize, or parody the source?
• Has the raw training data been abstracted and recombined into a distinct creative expression?
• Is the output a market substitute for the original work?

The computational process of identifying AI-generated outputs that are substantially similar to or directly adapted from copyrighted source materials. This involves comparing total concept and feel through techniques like embedding similarity scoring and n-gram overlap analysis. Detection systems must distinguish between unprotected ideas and protected expression to flag potential infringement before distribution.

A fingerprinting algorithm that generates a compact digest of multimedia content based on perceptual features rather than exact bit patterns. Unlike cryptographic hashes, pHash survives common transformations like resizing, compression, and color adjustment. This makes it critical for detecting visually similar AI-generated images that may infringe on copyrighted visual works, even when outputs are not pixel-for-pixel copies.

A legal remedy requiring the deletion of models trained on unlawfully collected or infringing data. When substantial similarity is proven across a class of outputs, courts may order the destruction of the tainted algorithmic asset itself. This goes beyond content takedown to address the root cause—model weights influenced by infringing training data—and represents the most severe remedy in AI copyright litigation.

A contractual and technical indemnification framework that protects enterprise users from copyright infringement claims arising from retrieval-augmented generation outputs. Unlike black-box models, RAG systems cite sources explicitly, enabling attribution chains that demonstrate provenance. Providers offering copyright shields assume liability for outputs that fail the substantial similarity test, shifting risk from the deployer to the model vendor.