Derivative Work Detection

The legal and technical standard for determining whether an AI output has copied protected expression. This analysis compares the total concept and feel of two works rather than just surface-level elements.

• Extrinsic Test: Analyzes specific, objective similarities in plot, sequence, and structure
• Intrinsic Test: Evaluates whether an ordinary observer would perceive substantial taking
• Filtration: Separates unprotectable elements (ideas, facts, scènes à faire) from protected expression before comparison

A fingerprinting algorithm that generates a compact digest of multimedia content based on its perceptual features rather than exact bit patterns. This enables detection of visually or audibly similar copies even after modification.

• Robustness: Survives compression, resizing, and minor color adjustments
• Hamming Distance: Measures similarity between hashes; a low distance indicates near-duplicate content
• Applications: Used in content ID systems to flag AI outputs that closely mirror copyrighted training images or audio

An imperceptible, cryptographically secure signal embedded directly into AI-generated content. This enables reliable detection and attribution of the output's origin, distinguishing between human-authored and model-generated derivative works.

• Invisible Embedding: Modifies pixel values or token probabilities in ways undetectable to humans but recoverable by detectors
• Tamper Resistance: Designed to survive common transformations like cropping, re-encoding, or paraphrasing
• C2PA Integration: Works alongside the Coalition for Content Provenance and Authenticity standard to bind provenance metadata

A privacy auditing technique that determines whether a specific data record was part of a model's training set. In copyright contexts, this helps identify if proprietary content was ingested without authorization.

• Shadow Model Training: Trains surrogate models on known datasets to learn the behavioral signatures of memorized data
• Confidence Score Analysis: Exploits the tendency of models to be more confident on training data than unseen data
• Legal Application: Provides statistical evidence that a model likely trained on copyrighted material, supporting infringement claims

A process that cross-references generated claims against a trusted knowledge base to detect hallucinations and unlicensed derivative outputs before they reach the end user.

• Real-Time Grounding: Compares each generated sentence against a vector database of authorized source documents
• Similarity Thresholding: Flags outputs exceeding a cosine similarity threshold with copyrighted reference material
• Attribution Tracing: When similarity is detected, traces the lineage back to the specific source document for licensing verification

A legal remedy requiring the deletion of models trained on unlawfully collected or infringing data. This effectively forces the destruction of the tainted algorithmic asset rather than merely removing the infringing output.

• Model Deletion: Complete removal of model weights trained on infringing data
• Machine Unlearning: Technical process to surgically remove the influence of specific data points without full retraining
• Precedent: Increasingly discussed in FTC enforcement actions and EU AI Act compliance frameworks as a deterrent against unauthorized data ingestion

The primary legal standard for determining if an AI output constitutes an infringing derivative work. The test compares the total concept and feel of the generated content against the copyrighted original, filtering out unprotectable elements like ideas, facts, and scenes a faire. In AI litigation, this involves computationally analyzing the extrinsic similarity of specific expressive elements and the intrinsic similarity as perceived by an ordinary observer.

A core computational technique for derivative work detection that generates a compact, fixed-size digest of multimedia content based on its perceptual features rather than raw bits. Unlike cryptographic hashes, pHash survives common transformations like resizing, compression, and color correction. Detection systems compare the Hamming distance between hashes to flag outputs that are visually or audibly substantially similar to copyrighted source materials in the training corpus.

The documented chain of custody for all datasets used in model training. Effective derivative work detection depends on complete provenance records that establish the legal status of every ingested work. Without rigorous provenance tracking, it becomes impossible to determine whether a generated output is a novel creation or a memorized reproduction of a copyrighted training sample. This forms the foundation for all downstream copyright compliance verification.

An imperceptible, cryptographically secure signal embedded directly into AI-generated content that enables reliable detection of its origin. In the context of derivative work detection, watermarking serves a dual purpose: it marks outputs as AI-generated to prevent passing off as human-created works, and it can encode provenance metadata that traces the generation path back to specific model versions and training data subsets.

A privacy and copyright attack that reconstructs representations of sensitive training data by exploiting access to a model's parameters and confidence scores. This attack vector demonstrates the risk that proprietary content can be extracted from models even without direct dataset access. Derivative work detection systems must account for the possibility that outputs may contain memorized fragments of copyrighted works that were never intended to be reproduced.

A legal remedy requiring the deletion of models trained on unlawfully collected or infringing data, effectively forcing the destruction of the tainted algorithmic asset. When derivative work detection systems identify systematic infringement, algorithmic disgorgement represents the most extreme compliance outcome. This remedy is increasingly discussed in regulatory frameworks and litigation as a proportional response to models that have ingested copyrighted works without authorization.