Talkin' 'Bout AI Generation: Copyright and the Generative-AI Supply Chain

Introduction

• Generative AI encompasses technically diverse systems trained on different data and deployed through varied access models, so copyright analysis cannot treat it as a single product.The systems range from chatbots and image generators to coding, music, video, and drug-discovery applications; models may be closed, semi-closed, or open-source. [1]Source: Introduction · p. 4

I. Machine Learning and the Generative-AI Supply Chain

• The generative-AI supply chain traces legal choices across creative works, data conversion, dataset construction, model training, fine-tuning, deployment, generation, and alignment.The stages interact through feedback rather than forming a simple cascade, and choices at each stage can have copyright consequences.Source: Introduction · p. 5

A. Background on Machine Learning

• The Article presents selective technical background on data and machine learning to identify which technologies operate at different supply-chain stages and support later copyright analysis.It explicitly does not aim for completeness, treating the supply-chain framework as essential to the legal analysis developed later.Source: §I · pp. 7–8

1. What is data?

• Machine-learning data are quantified representations of entities or phenomena, organized into datasets and converted into numerical formats that models can store, process, and interpret.For example, text may be represented as word embeddings, while paintings and books must first be computationally converted into data.Source: §1 · pp. 8–10

2. What is machine learning?

• Machine learning uses algorithms to model patterns in data, with discriminative methods predicting or classifying inputs and generative methods estimating distributions that can produce content.Generative modeling does not necessarily create new content, and the familiar deterministic function analogy fits discriminative classification better than generative modeling.Source: §2 · p. 10, a. Discriminative modeling · p. 11
• Generative models can produce multiple reasonable outputs for one input, unlike discriminative models that return a single output from possible labels.For example, a caption such as “cat in a red and white striped hat” can correspond to many reasonable images, while a classifier labels an image as cat or dog.Source: b. Generative modeling · p. 14
• Because generation involves randomness, running the same model on the same input can produce different outputs.This variability is illustrated by multiple generated images for the same caption in Figure 2.Source: b. Generative modeling · p. 14, Figure 2 · p. 15
• Generative models learn output likelihoods from training data and use those probabilities to select among plausible continuations.Words such as “beach” are more likely than “slopes” after “In the summer, I like to go to the,” while “beach” and “pool” may be similarly likely.Source: b. Generative modeling · pp. 14–15

2. Generation Modalities

• Generative AI accepts expressive inputs and produces expressive outputs across modalities, including text-to-text, text-to-image, code, audio, video, and molecular structures.The Article focuses its detailed examples on text and image systems while acknowledging broader applications.Source: B. What Is “Generative AI”? · p. 16, §2 · p. 18, c. Other modalities · pp. 23–24

B. What Is “Generative AI”?

• Generative AI encompasses loosely related systems for text, images, code, music, video, and molecular structures, rather than one uniform technology.The Article therefore focuses on systems whose models are embedded in larger technical and organizational arrangements.Source: B. What Is “Generative AI”? · pp. 16–17, c. Other modalities · pp. 23–24
• The field’s distinguishing features include multiple embedded models, massive web-scraped datasets, transformer and diffusion techniques, and scale across systems, models, data, and computing resources.These features distinguish contemporary generative AI from more traditional generative modeling.Source: B. What Is “Generative AI”? · pp. 16–17, §2 · p. 19, b. Image data and generations · p. 22
• The Article organizes these systems through a supply-chain lens linking development, evaluation, deployment, and evolution across stages involving different people and organizations.This framing is presented as indispensable for analyzing potential legal issues.Source: B. What Is “Generative AI”? · p. 17, §1 · p. 18

1. Generative-AI Systems

• Generative-AI systems combine multiple models and components behind user interfaces, so users generally interact with a system rather than a model directly.Examples include ChatGPT’s interface and DreamStudio’s larger system built around Stable Diffusion.Source: §1 · pp. 17–18

2. Generation Modalities

• Text models such as ChatGPT are trained on sequences to predict the next word, while text-to-image models learn relationships between captions and images to generate prompt-matching images.Image-text datasets are often scraped from the Internet and may contain copyrighted and public-domain material.Source: a. Text data and generations · pp. 19–20, b. Image data and generations · pp. 21–22

3. Machine-Learning Techniques in Generative AI

• Generative-AI systems prominently use transformer architectures and diffusion-based models, while increased dataset, model, and computing scale forms a third major development.The Article introduces these developments because each raises distinct copyright considerations.Source: §3 · p. 25

a. Transformer architecture

• Transformers model relationships within sequences and generate the next token from context; their scalable stacking supports models with billions of parameters.ChatGPT is an example of a transformer-based generative text system.Source: a. Transformer architecture · pp. 25–27

b. Diffusion-based models

• Diffusion is a training algorithm rather than a model architecture: it repeatedly adds noise to training images and trains a neural network to reverse that corruption.Text-to-image diffusion training uses paired images and captions, evaluating whether denoised images match the originals.Source: b. Diffusion-based models · pp. 27–29
• At generation time, diffusion-based systems start from noise and repeatedly remove it, producing a sequence of increasingly prompt-aligned images before the final output.Midjourney can display or animate these intermediate denoising stages in its interface.Source: b. Diffusion-based models · p. 29, Figure 3 · p. 29
• Intermediate images in a diffusion sequence can matter legally because an internal or displayed stage may resemble a copyrighted work more closely than the final generation.The Article connects this possibility to questions about reproduction, display, and transient generations.Source: b. Diffusion-based models · p. 29

4. The Role of Scale

• Scale enables generative-AI models to capture massive datasets and produce powerful content, but their internal operation and the reasons for these capabilities remain poorly understood.State-of-the-art transformer models have billions of parameters and trillions of connections, while datasets such as LAION-5B contain billions of examples.Source: §4 · pp. 30–31

C. The Generative-AI Supply Chain

• The generative-AI supply chain comprises eight connected stages, clarifying which artifacts, production moments, and actors may matter for copyright liability.The stages are expressive-work creation, data creation, dataset collection and curation, pre-training, fine-tuning, deployment, generation, and alignment.Source: C. The Generative-AI Supply Chain · pp. 32–33, Figure 4 · p. 36

1. The Creation of Expressive Works

• Creators and their expressive works enter the supply chain when works are transformed into data objects and circulated online, regardless of creators’ preferences.Most state-of-the-art systems depend on data derived from pre-existing creative works, although synthetic training data is also possible.Source: §1 · pp. 33–34

2. Data Creation

• Datafied creative works remain relevant copyrights, and different generative-AI modalities require corresponding data types such as text, images, audio, or combinations thereof.Examples include copyrighted code in GitHub Copilot, web-scraped text underlying ChatGPT, and text and images used for Stable Diffusion.Source: §2 · pp. 34–35

3. Dataset Collection and Curation

• Web-scale dataset collection and curation separate potential dataset-creator liability from model-trainer liability, while missing provenance makes source verification and removal difficult.A dataset may itself be copyrightable, yet a dataset license does not guarantee that its included works are appropriately licensed; LAION-5B illustrates this uncertainty.Source: §3 · pp. 38–39

4. Model (Pre-)Training

• Model pre-training requires selecting data, architecture, algorithm, and randomness while consuming substantial time, storage, computing power, and human monitoring.BLOOM used 176 billion parameters, 1.6 terabytes of text, 384 GPUs, and 3.5 months of training; estimated computing costs were $2–5 million.Source: §4 · pp. 39–41
• Evaluation metrics attempt to measure whether a model is useful or good, but no single metric can comprehensively capture those qualities.The passage characterizes usefulness and goodness as difficult to define mathematically.Source: §4 · p. 41
• Model parameters occupy a murky boundary between data and programs because they can be loaded, transformed into generations, and executed with prompts.“Model” may refer either to parameter vectors or to software combined with those parameters.Source: §4 · p. 41
• A base model can be shared, deployed in an application, or used by another actor for additional training with new or modified data.This reuse can occur within the original organization or across organizations.Source: §4 · p. 42
• Further training of a base model is called fine-tuning, explaining why the initial stage is typically termed pre-training.A base model is also commonly called a pre-trained model.Source: §4 · p. 42

5. Model Fine-Tuning

• Fine-tuning transforms a trained model into another model by continuing training with additional data, typically from a specific domain.Unlike pre-training, it starts from already-trained parameters and often targets a particular capability, such as scientific-text summarization.Source: §5 · pp. 42–43
• A model may proceed from pre-training directly to deployment or generation, undergo repeated fine-tuning, or receive alignment instead.“Base” and “fine-tuned” describe relative creation processes rather than inherent technical properties.Source: Forks in the supply chain · pp. 43–44, §6 · p. 45
• Fine-tuning can be performed by successive actors, creating chains of models derived from a publicly released base or fine-tuned model.Vicuna, for example, was fine-tuned from Meta’s Llama model using the ShareGPT dataset, after which others could fine-tune it again.Source: Forks in the supply chain · pp. 44–45
• Different participants may have different knowledge, control, and intentions regarding training data and safeguards, so copyright analysis may need to distinguish their conduct.A base-model creator may add anti-infringement guardrails that another party later attempts to remove through fine-tuning.Source: Forks in the supply chain · p. 45

6. Model Release and System Deployment

• Models can be released for direct download or embedded in hosted applications and APIs, with each deployment mode giving users and deployers different levels of control.Released parameters enable independent embedding, fine-tuning, or alignment, while hosted services generally limit direct parameter changes.Source: §6 · pp. 45–47
• Public release makes prohibited-use policies difficult to enforce because users obtain direct access to model parameters.The Llama model’s leak illustrates the difficulty of controlling access after release.Source: §6 · p. 46
• Hosted deployment can modify prompts and filter inputs or outputs, allowing deployers to steer behavior without necessarily changing model parameters.Output filters intended to prevent similarity to training examples are described as imperfect.Source: §6 · pp. 47–48
• Training-data access affects model quality: open-source models often rely on open data and may perform worse in poorly represented domains or demographics.The cited examples report degradation for scientific domains and for diffusion models trained on Creative-Commons images.Source: §6 · pp. 48–49

7. Generation

• Users choose prompts and deployed systems, then receive randomly selected candidate generations; these choices shape how the user-system relationship may be characterized in copyright analysis.The relationship may be viewed as user-driven, system-driven, or a partnership, with the best fit depending on the system, prompt, marketing, and interface.Source: §7 · pp. 49–50
• Prompts themselves may be expressive or contain copyrighted works, and deployers may retain them for training or alignment.A prompt can include an entire book or copyrighted image, creating additional copyright considerations beyond the resulting generation.Source: §7 · p. 51
• Generation can feed back into dataset creation, while retrieval-augmented generation appends selected data examples to prompts to guide outputs.These loops make the boundaries between expressive works, data creation, dataset curation, and generation porous.Source: §7 · pp. 51–52
• External plugins and services can supply current information or other content to shape generations, potentially bringing additional organizations and copyright issues into the supply chain.The resulting chain may include news articles or short stories integrated through external services.Source: §7 · p. 53

8. Model Alignment

• Model alignment uses human feedback, often through reinforcement learning with human feedback, to encourage preferred generations and discourage others.Companies may perform alignment before deployment, using simulated user feedback and complementary filtering to steer outputs toward safety, helpfulness, legal compliance, or organizational policies.Source: §8 · pp. 53–55

II. Tracing Copyright Through the Supply Chain

• The Article applies established copyright doctrines systematically to each supply-chain stage, covering authorship, infringement, secondary liability, defenses, remedies, and copyright-like rights.Its stated goal is careful and systematic analysis rather than a dramatically new account.Source: §II · pp. 55–56

A. Authorship

• Copyrightable training data must be fixed and human-authored, while datasets can mix copyrighted and uncopyrighted material across domains.Birdsong recordings and synthetic training data may be uncopyrightable, whereas photographs, illustrations, music, writing, and software may be copyrighted; datasets themselves may qualify as compilations through original selection or arrangement.Source: A. Authorship · pp. 56–58
• Models may lack copyrightable originality when they apply standard algorithms and architectures, but curated data, novel architectures, or training choices can imbue some models with copyrightable expression.A model may also be treated as either a compilation of training data or a derivative work that transforms underlying works, with doctrinal consequences for pre-trained models.Source: Pre-Trained/Base Models · pp. 59–60

Fine-Tuned Models and Aligned Models

• Fine-tuned and aligned models typically raise the same categorization issues and may be derivative works of their base models, while training choices and RLHF prompts can supply creative human input.User prompts may be retained as training data, allowing copyrighted material—including entire books—to enter fine-tuning or alignment datasets, subject to possible express or implied licenses and separate infringement questions.Source: Fine-Tuned Models and Aligned Models · pp. 60–61

A. Authorship

• U.S. copyright law does not recognize computers as authors, so an autonomously generated output may lack copyright under Thaler v. Perlmutter.Potential human claimants include training-data authors, supply-chain entities, users, or no one, and the paper finds no one-size-fits-all allocation.Source: Deployed Services · pp. 62–63
• A generation is not literally a collage: models encode training examples in parameters rather than explicitly selecting, coordinating, and arranging them during generation.If an output is identical to a training example, it is a copy rather than a new copyrightable work; if it is an infringing derivative work, no separate copyright extends to the unlawfully used material.Source: Deployed Services · pp. 63–64
• When an output is sufficiently distinct from training data, app developers may own outputs requiring minimal user input, while substantial prompt-based creative control can support user ownership.Ownership may also be influenced by providers’ practical decisions not to assert copyright over outputs.Source: Deployed Services · pp. 64–65
• Short prompts such as “Scary lighthouse” may provide insufficient originality or control for user authorship, leaving some generations uncopyrightable despite containing material that would qualify if made manually.The paper notes that this conclusion is limited: elaborate, iterative prompt engineering can produce generations attributable to users, and future interfaces may provide greater control.Source: Deployed Services · pp. 65–66
• User-provided sketches can contribute potentially copyrightable material that is transformed into more-complete AI-generated images.The cited example describes users uploading sketches and further controlling the generated images with text prompts.Source: Deployed Services · p. 66
• Long-context and retrieval-augmented systems may incorporate entire books or retrieved documents as copyrighted inputs guiding generation.These system capabilities expand the authorship analysis beyond prompts to copyrighted textual materials supplied or retrieved during generation.Source: Deployed Services · p. 66
• Users who deliberately craft prompts may claim copyright in resulting AI outputs, just as artists can claim copyright in accidental variations they adopt as their own.The section frames this as an authorship question: if an accidental spatter can qualify for copyright, a deliberately engineered prompt may also support a claim.Source: Deployed Services · pp. 66–67

Reproduction right

• Every stage of the generative-AI supply chain requires a potentially infringing reproduction, although whether infringement occurs depends on substantial similarity.Datasets, models, prompts, and generations can be stored as copies, while the storage device—not the dataset itself—is the material object containing the copy.Source: B. The Exclusive Rights · p. 67, The Reproduction Right · p. 68
• Transient-copy analysis usually matters more for inputs than outputs because datasets, models, applications, prompts, and generations are typically stored longer than 2.4 seconds.How long training examples and intermediate generation values remain in memory depends on algorithms, architecture, processors, and generation speed.Source: The Reproduction Right · p. 68

Adaptation right

• The adaptation right covers new works based on existing works, making models and generations potentially derivative even when they are not exact reproductions.A dataset is more appropriately classified as a compilation, while whether a prompt or generation is derivative depends on its relationship and similarity to existing works.Source: The Adaptation Right · p. 69
• Models may implicate the adaptation right merely by referencing training works, but a narrower approach would require the ability to generate substantially similar outputs.Reliable attribution of training examples to resulting generations remains an open research question.Source: The Adaptation Right · p. 70

Distribution right

• The distribution right generally requires interaction between parties, whereas copying within one hosting entity is internal and not public distribution.Relevant interactions include sharing datasets, downloading models, producing generations for users, retrieving external content, and sharing artifacts as-is.Source: The Distribution Right · p. 71

Display and performance rights

• Display and performance rights generally apply only to perceptible generations, especially when services show or play outputs publicly, while models and datasets are not ordinarily perceptible.Transient intermediate images or live audio and video can implicate these rights even without reproduction or distribution.Source: The Display and Performance Rights · p. 72

Training data

• Training datasets contain complete literal copies of millions of copyrighted works, while recognizable or partially transformed copies may still support substantial similarity.Modified or noised works can remain substantially similar when the original is recognizable; comparisons focus on protected creative choices rather than unprotectable features.Source: C. Substantial Similarity · pp. 73, 85

Memorization

• Evidence of memorization includes Stable Diffusion reproducing recognizable photographs and GPT-4 filling in fictional character names from copyrighted books.The GPT-4 example is suggestive rather than conclusive because its training data are not public and the names could have been learned from summaries or related material.Source: C. Substantial Similarity · p. 75, Figure 6 · p. 75

Models and memorization

• Models may themselves be substantially similar copies when they can generate a copyrighted work, regardless of whether that generation is ever produced.This theory treats the model as the copy, rather than merely a tool that causes an infringing output.Source: C. Substantial Similarity · p. 76

Generations

• Generations range from nearly identical memorized works to outputs dissimilar from every training work, with the legally difficult cases lying between those extremes.Similarity must be assessed work by work, and no automated metric definitively resolves substantial similarity across massive datasets.Source: C. Substantial Similarity · p. 78

Snoopy effect

• The “Snoopy effect” describes outputs that reproduce recognizable characters, prevalent works, or artists’ styles even when prompts do not name the source.Examples include Darth Vader, Dr. Seuss-like technical directions, Dorothea Lange-like photographs, and an Indiana Jones-like archaeologist.Source: Figure 7 · p. 79, C. Substantial Similarity · p. 80, Figure 8 · p. 81

Limits on similarity

• Scènes à faire, de minimis copying, and filtration can limit infringement when similarities reflect genre tropes, insignificant details, or unprotected elements.An “ice princess” output may resemble Frozen through standard wintry-glamour tropes, while facial features or hairstyles may be irrelevant to a photographer’s protected creativity.Source: C. Substantial Similarity · pp. 82, 84, 85

Copying in fact

• Copyright’s “copying in fact” inquiry asks whether the defendant copied from the plaintiff’s work rather than creating independently or using another source.Courts typically rely on direct evidence or indirect evidence of access and probative similarity.Source: D. Proving Copying · pp. 85–86

Digital evidence

• Digital file identity can strongly evidence copying, but differing encodings do not establish independent creation because content-comparison algorithms risk false positives and false negatives.Exact bit-for-bit identity is unlikely to arise from independent digitization, whereas the same work can produce different files.Source: D. Proving Copying · p. 86

Training-data evidence

• Training-data access is difficult to prove because datasets are large, derivative copies are hard to detect, and companies increasingly disclose little about what their models trained on.Exact copies are theoretically searchable, but non-exact copies and undisclosed datasets create substantial practical obstacles.Source: D. Proving Copying · pp. 87–88

Burden of proof

• The evidentiary burden is asymmetrical: a plaintiff may point to one similar work in the dataset, while a defendant must show that no copied work appears there.For alleged generation infringement, similarity between the generation and plaintiff’s work may support an inference that similar material was present in training even without identifying it.Source: D. Proving Copying · p. 87

Model opacity

• Generative-AI models are not human-interpretable, so proving copying will typically require showing that the model produced a generation substantially similar to the allegedly copied work.Uninterpretability often leaves access and similarity as the principal available evidence.Source: D. Proving Copying · p. 88

Absence from training data

• If the plaintiff’s work is absent from training data, independent generation may appear dispositive, but proving that neither the work nor its derivatives were included remains difficult.The evidentiary challenge persists even when the defendant documents an apparently independent creative process.Source: D. Proving Copying · p. 89

E. Direct Infringement

• Direct infringement turns on who most proximately caused the infringing copy, distinguishing actors who choose specific content from providers that merely offer general-purpose tools.A service that curates infringing content or makes copies for users may be direct, while a service that lets users choose what to upload or download may instead face indirect liability.Source: E. Direct Infringement · p. 90
• Most generative-AI supply-chain stages involve direct infringement because selecting copyrighted training sources is volitional conduct, even without knowledge of specific copyrighted works.Direct liability is strict, so curators, model trainers, fine-tuners, and aligners act at their own risk when choosing datasets containing copyrighted material.Source: E. Direct Infringement · pp. 89–90, Pre-Trained, Fine-Tuned, and Aligned… · p. 91
• When a model and prompt come from the same actor, subconscious-copying doctrine provides an established analogy for treating the resulting infringing generation as the actor’s responsibility.The model’s opaque internal representations may draw on prior works, while the prompter could have avoided generation or checked whether the output resembled copyrighted material.Source: Pre-Trained, Fine-Tuned, and Aligned… · p. 91
• For hosted generation services, courts could treat the user, the provider, or both as direct infringers, depending on who selected or caused the infringing output.The choice is partly factual and partly policy-driven; treating both parties only as indirect infringers would leave no direct infringer for indirect liability to attach.Source: Pre-Trained, Fine-Tuned, and Aligned… · pp. 93, 95

F. Indirect Infringement

• Indirect liability requires an underlying direct infringement and can move responsibility upstream toward actors with greater control, knowledge, financial interest, or intent.The framework includes vicarious, contributory, and inducement theories, with the analysis proceeding backward from service users toward content creators.Source: F. Indirect Infringement · pp. 95–96
• A generation service may face vicarious or inducement liability based on control, financial incentives, or intentional promotion of infringement, while generalized knowledge usually does not establish contributory liability.Specific notice of an infringing output changes the analysis: after notice, the provider must take steps to prevent similar future generations.Source: Generation via a Hosted Deployed Service · p. 97
• Model trainers with contractual relationships to downstream users may incur vicarious or contributory liability and therefore have an ongoing duty to avoid licensing models to blatant infringers.Open or semi-closed models raise a separate possibility of direct liability if their embedded representations are treated as infringing reproductions, which would defeat the Sony defense.Source: Model Pre-Trainers, Model Fine-Tuners,… · p. 99
• Dataset curators are ordinarily direct infringers because datasets contain copies of expressive works, while content creators are generally too remote for secondary liability absent a close nexus or knowing contribution to blatant infringement.Secondary liability could nevertheless reach dataset or content contributors who knowingly supply material for an infringing model, even when their own contributions are non-infringing.Source: Training Dataset Creators/Curators and… · pp. 99–100
• Stability AI’s donated computing resources for Stable Diffusion’s training group and dataset creator could support findings of informal agreement, material contribution, or direct financial interest in infringement. [492]A court could treat the parties as more than arms-length despite no formal legal obligation to provide anything in return.Source: Training Dataset Creators/Curators and… · p. 100

Section 512 overall

• Most Section 512 safe harbors likely do not apply to generative AI because model trainers and deployers, rather than users, choose the relevant data, models, and operations.The statute’s transmission, caching, storage, and search-engine frameworks generally describe different intermediary functions than those performed by generative-AI systems.Source: G. Section 512 · p. 101

Sections 512(a) and 512(b)

• Sections 512(a) and 512(b) do not fit model training or deployment because storage is not transient or temporary, activity is not user-directed, and generation modifies or combines content.Caching also presents a sequencing problem: relevant model storage often occurs before any user request, rather than between an initial request and later transmissions.Source: Section 512(a): Transmission · pp. 101–102, Section 512(b): Caching · p. 102

Section 512(c)

• Section 512(c) is a closer but generally unsuccessful fit because datasets, models, and incorporated systems are stored at providers’ or curators’ direction, except potentially user prompts or fine-tuning datasets.A request to generate content may differ from uploading material intended for storage and later viewing by other users.Source: Section 512(c): User-Directed Storage · p. 103

Section 512(d)

• Section 512(d) generally does not describe generative-AI systems because datasets and models usually contain copied works rather than links to external online locations, though generation-time plugins may sometimes function like search engines.Plugins could potentially produce outputs that link to infringing content, even though they may also include infringing material directly in generations.Source: Section 512(d): Search Engines · pp. 103–104

Notice and takedown

• Notice-and-takedown rules under Sections 512(c) and 512(d) remove safe-harbor protection when providers receive sufficiently specific infringement notices and fail to disable access.The notice must identify both the copyrighted work and the location of the infringing material, making the regime a mechanism for conveying actionable knowledge.Source: Notice and Takedown · p. 104

Fair use

• Fair use outcomes for generations vary: transformation may favor fair use, but memorized or minimally transformed outputs, substantial copying, and market substitution may weigh against it.Factors one, three, and four can point strongly either way, while factor two depends on the model and the nature of its training works.Source: H. Fair Use · pp. 105–109
• Training non-generative-AI systems has a stronger fair-use argument when it involves nonexpressive uses such as stylometry, sentiment analysis, or plagiarism detection that do not compete with authors.The rationale is that these uses do not involve human consumption of expression, so intermediate training, fine-tuning, alignment, and deployment likewise do not compete with authors.Source: H. Fair Use · pp. 109–110

Models and generations

• Generative-AI models may copy more material than any single generation and may disproportionately capture works’ distinctive “hearts.”The amount and qualitative significance of material embedded in a model remain difficult conceptual and empirical questions.Source: H. Fair Use · p. 110
• Whether a licensing market exists for generative-AI training is unresolved because fair-use findings can prevent the market from developing, while commercial actors increasingly seek licenses.Getty’s litigation against Stability AI and The New York Times’s lawsuit against Microsoft and OpenAI are described as efforts to force licensing negotiations.Source: H. Fair Use · p. 111
• Fine-tuning and alignment can change fair-use analysis by adding copyrighted data and steering a model toward more or fewer infringing outputs.Commercial hosted services and paid model or API licensing are commercial uses, whereas freely released models may be non-commercial.Source: H. Fair Use · pp. 111–112
• A model may qualify as fair use while some outputs do not, although an unfair model is unlikely to produce only fair outputs.Generative systems can produce human-visible copyrighted expression, so the non-copyright-implicating rationale associated with Google Books no longer applies categorically.Source: H. Fair Use · pp. 110, 112

Training datasets

• Training datasets differ from models and generations because they usually preserve complete works without new expression, though assembling them may serve a transformative training purpose.Datasets often copy wholesale to retain information useful for training, while licensing markets for datasets and AI training are nearly indistinguishable.Source: H. Fair Use · pp. 112–113
• A dataset can be unfair for publicly exposing copyrighted works independently of training, making the shadow-library Books3 dataset very likely unfair.A trainer’s knowledge or reason to know that a dataset is infringing may also influence a court’s fair-use analysis.Source: H. Fair Use · pp. 113–114

Licensing

• Specific copyright-owner licenses fully defend authorized uses, but unauthorized dataset or model licenses are ineffective and mistaken works are harder to remove from trained models.Specific permissions cover only a small fraction of current training data, and models trained solely on specifically licensed works are rare and often lower quality.Source: §I · pp. 114–115
• Creative Commons terms create distinct constraints: attribution is technically unsupported, non-commercial terms exclude commercial links, no-derivatives effectively prohibits generative AI, and share-alike raises compatibility issues.Generation-time plugins can add unlicensed data; because no current generative-AI models or generations use Creative Commons licenses, the authors call CC licensing a current dead-end.Source: §I · pp. 116–118

Implied licenses

• Online publication can imply consent to search indexing and archival copies, but the license generally does not extend to unrelated publication uses, AI training, or AI-generated outputs.The implied-license analysis depends on scope: training resembles covered archiving, but training and generation are newer uses, and generation is more plausibly analyzed under fair use than consent-based implied license.Source: J. Implied Licenses · pp. 118–120
• Voluntary online publication can still support fair use for AI training, while trespass and computer-access laws usually target the scraper and allow scraping of public material absent additional site harm.These noncopyright restrictions generally do not follow scraped data to model trainers or deployers, whereas copyright does.Source: J. Implied Licenses · p. 120

Damages

• Lost-license-fee damages depend on a licensing market, but no established market yet exists for licensing AI training on books, making valuation difficult.Actual damages may also be difficult to prove, and individual works may have uncertain licensing value.Source: Damages and Profits · pp. 121–122
• A generation may be treated like an anthology, potentially exposing defendants to all profits unless they prove the share attributable to other sources; conversely, a single work’s inclusion in a full model may not establish attributable profits.Mass claims become more significant when plaintiffs collectively own a large fraction of the training data, even though individual apportionment remains difficult.Source: Damages and Profits · pp. 122–123
• Statutory damages are awarded per infringed work, ranging from $750 to $30,000 absent special adjustments, making billions of works potentially expose models to trillions of dollars.Registration requirements limit availability and create uneven exposure: books are commonly registered, while many visual works and websites are not.Source: Damages and Profits · pp. 124–125

Injunctions and fee shifting

• Attorney’s fees create especially strong settlement pressure for smaller defendants, while injunctions depend on irreparable harm, inadequate damages, hardship, public interest, and the system’s mix of infringing and noninfringing uses.Courts may protect expressive noninfringing uses by limiting an injunction to infringing generations and requiring content filters rather than shutting down the model entirely.Source: Attorney’s Fees · pp. 126–128

Injunctions

• Separating fair uses from infringing outputs is harder in generative-AI systems than on content-hosting services, and filtering remains a difficult, unsolved technical problem.Even giving defendants lists of works to avoid may not make reliable filtering feasible.Source: Attorney’s Fees · pp. 128–129

Destruction remedies

• Treating a model as an infringing copy would make destruction an exceptionally harsh remedy that could effectively discard all compute used to train it.Deleting the model, rather than storing it pending legal or licensing changes, would make the remedy irreversible.Source: Destruction · p. 129
• Destruction orders could extend beyond infringing models to dual-use systems and other service components capable of reproducing infringing works, raising severe free-expression concerns.A court might, for example, order destruction of a style-transfer system that could generate an infringing derivative work from a copyrighted image.Source: Destruction · pp. 129–130

• Section 1202 prohibits providing false copyright management information or removing or altering it.Copyright management information includes titles, authors, copyright owners, performers, and licensing information.Source: L. Copyright Management Information · p. 130
• Training datasets can remove copyright management information by separating images from webpage credits, and model generations may lack the original information.The analysis analogizes this decontextualization to cases involving reproduced photographs without accompanying photographer credits.Source: L. Copyright Management Information · pp. 130–131
• Getty alleges that Stable Diffusion both removes or alters watermarks during training and produces distorted watermarks constituting false information.Source: L. Copyright Management Information · p. 131
• Section 1202 liability requires a connection to infringement, including intent to induce, enable, facilitate, or conceal infringement for falsification claims.The analysis questions whether CMI treatment independently facilitates or conceals infringement when the underlying conduct would not otherwise be actionable.Source: L. Copyright Management Information · p. 131
• Statutory damages range from $2,500 to $25,000 per section 1202 violation, potentially multiplying across datasets, models, or watermarked outputs.Courts may reduce or remit damages for innocent violations.Source: L. Copyright Management Information · p. 132

Overview of the Right of Publicity

• The right of publicity protects recognizable aspects of identity, including a person’s name, voice, signature, photograph, or likeness, primarily against commercial uses.Core applications include endorsement, merchandising, and digitally recreating a person in expressive media.Source: Overview of the Right of Publicity · pp. 132–133
• Copyright preemption often bars publicity claims based on ordinary exploitation of copyrighted works unless the claim protects different subject matter or adds an extra element.Using a vocal sample in a new song was preempted, while using it in an advertisement could supply the required additional element.Source: Overview of the Right of Publicity · p. 134

Incorporation in the Generative-AI Supply Chain

• Generative AI creates publicity issues through identity incorporation and advertising, illustrated by claimed imitation of George Carlin’s voice and promotion using his name.The Carlin dispute ultimately involved a human-written episode, but the example distinguishes incorporating identity from using identity to advertise.Source: Incorporation and Advertising · p. 135
• Fake celebrity endorsements generated by AI are actionable chiefly because they use a person’s identity without permission, not because the media was generated rather than photographed or recorded.Examples include cloned Taylor Swift audio and a deepfake Tom Hanks dental advertisement.Source: Incorporation and Advertising · p. 135, Incorporation in the Generative-AI… · p. 136
• Publicity law remains unsettled for generative AI because secondary liability is less developed, Section 230’s applicability is disputed, and AI generations may or may not count as third-party content.Transformative-use and public-interest defenses can protect some later-stage generations, while earlier supply-chain stages often lack recognizable identity use.Source: Incorporation in the Generative-AI… · pp. 136–138
• Claims about AI systems imitating an artist’s style may be preempted when they mirror copyright claims, while models specifically developed to imitate a person may support advertising, merchandising, or contributory-impressment theories.A user’s prompt may not be attributable to the service, creating an additional doctrinal hurdle.Source: Incorporation in the Generative-AI… · p. 137

N. Hot News Misappropriation

• Hot-news misappropriation addresses competitors’ systematic copying of fresh information when free-riding threatens incentives to produce it.The doctrine originated in the AP–INS dispute and distinguishes copying from the public’s independent discussion or reporting of news.Source: N. Hot News Misappropriation · pp. 138–139
• Modern hot-news claims are governed by state law and face copyright preemption unless the copied information falls outside copyright’s scope or the claim includes an extra element.The plaintiff must identify a legally distinct basis beyond the ordinary elements of copyright infringement.Source: N. Hot News Misappropriation · p. 139
• The New York Times’s claims against Microsoft and OpenAI resemble hot-news disputes because the Times gathers time-sensitive information at substantial cost.The Times also alleges misappropriation of Wirecutter recommendations, but copyright preemption makes that theory difficult because the reviews are fixed works and the asserted elements overlap with infringement.Source: N. Hot News Misappropriation · p. 139
• Generative-AI training and deployment usually occur too slowly for breaking-news value to remain fresh, making the Times’s hot-news theory unlikely to succeed.The paper distinguishes AI systems from same-day news retransmission because competition for readership alone is probably insufficient after preemption.Source: N. Hot News Misappropriation · pp. 139–140

III. Which Way from Here?

• Generative AI cannot be classified categorically because legality depends on dataset curation, training algorithms, deployment, and user prompts.Copyright doctrines such as substantial similarity, indirect infringement, fair use, and remedies also contain open-ended tests that permit different fact-sensitive outcomes.Source: §III · p. 140

A. Possible Outcomes

• A no-liability regime would make services, users, models, and usually training datasets categorically legal, but could destabilize copyright for style-emulation and retrieval systems.The paper warns that developers could add generative components to launder nearly reproduced copyrighted works.Source: No Liability · p. 141
• A generations-only regime would treat systems as creative tools while placing responsibility for external infringement on users, potentially softened by indemnification.Services could receive strong Sony-like protection through no substantial similarity, no indirect infringement, and/or fair use.Source: Liability for Generations Only · pp. 141–142
• Courts could impose liability on providers whose systems are specifically designed to increase infringement, while preserving responsibility for each actor’s own targeted conduct.A separate Napster-like regime would require services to respond to known infringements, with duties resembling familiar notice-and-takedown expectations.Source: Liability for Generations Only · pp. 142–143
• Notice-based filtering is technically harder for generative AI than for hosting because notices concern future, variable generations rather than existing files, and removing training influence from models remains unsolved.Retraining from scratch may be the only safe option but is often infeasible for individual notices.Source: Liability for Generations Only · pp. 143–144
• Courts might protect distributed model parameters under Sony because of substantial noninfringing uses, or instead require trainers to prevent infringement through safe deployment and guardrails.The stricter approach would make open-source models especially risky and divide systems between licensed or carefully controlled proprietary models and conservative open models.Source: Liability for Generations Only · pp. 144–145
• Datasets are comparatively easier removal targets than models because identified works can be searched, hashed, traced through provenance, and deleted from future distributions.Open or leaked datasets may still circulate despite deletion from the provider’s copy.Source: Liability for Generations Only · p. 145
• A licensing regime would require models to train on licensed data and negotiate model and generation uses, leaving infringement mainly to quality-control failures.This regime would make licensing the central component of copyright compliance.Source: Infringing Models · pp. 145–146

B. Lessons

• Copyright touches every supply-chain stage, from training data through alignment, because each stage can use copyrighted works and create legal consequences.The paper contrasts copyright’s pervasive role with other AI issues that mainly concern outputs.Source: B. Lessons · p. 146
• Copyright concerns cannot be localized because choices by datasets, architectures, algorithms, fine-tuning, deployment, and prompts can affect another actor’s downstream liability.Different actors may control each step, so an output’s resemblance to a protected work can reflect decisions distributed across the chain.Source: B. Lessons · p. 147
• Design choices affect copyright exposure, including whether data are unlicensed, how much models memorize, how users control prompts, and how systems respond to infringement notices.The paper treats architectural choices as legally relevant alongside explicit copyright decisions.Source: B. Lessons · p. 147
• Fair use is not a silver bullet because potentially infringing generations reopen fair-use questions at earlier stages of the supply chain.The possibility that outputs may infringe undermines the assumption that training materials and models will remain unseen by users.Source: B. Lessons · p. 147
• Courts should continue making ordinary, case-by-case judgments about similarity, copying, and transformation, allowing some systems and generations to infringe while others do not.The paper cautions against categorical rules driven by the novelty of the technology.Source: B. Lessons · pp. 147–148
• Analogies to search engines, websites, libraries, or authors can help but should not be treated as definitive because generative AI can converge with several technologies and uses.Excluding competing analogies risks ignoring relevant similarities in a diverse, generative system.Source: B. Lessons · p. 148

IV. Conclusion

• The conclusion is that generative AI has no yes-or-no copyright answer because its systems differ and copyright law contains too many open decision points.The supply-chain framework is offered as a mechanism for reasoning about that diversity and its legal consequences.Source: §IV · pp. 148–149