PaperlayerView exact revision on arXiv

Source-linked AI summary

Report of the 1st Workshop on Generative AI and Law

A. Feder Cooper, Katherine Lee, James Grimmelmann, Daphne Ippolito, Christopher Callison-Burch, Christopher A. Choquette-Choo, Niloofar Mireshghallah, Miles Brundage, David Mimno, Madiha Zahrah Choksi, Jack M. Balkin, Nicholas Carlini, Christopher De Sa, Jonathan Frankle, Deep Ganguli, Bryant Gipson, Andres Guadamuz, Swee Leng Harris, Abigail Z. Jacobs, Elizabeth Joh, Gautam Kamath, Mark Lemley, Cass Matthews, Christine McLeavey, Corynne McSherry, Milad Nasr, Paul Ohm, Adam Roberts, Tom Rubin, Pamela Samuelson, Ludwig Schubert, Kristen Vaccaro, Luis Villa, Felix Wu, Elana Zeide

arXiv:2311.06477v3cs.CY

TL;DR

The report addresses how law and generative AI challenge one another amid limited shared understanding across computer science and legal disciplines. Drawing on the GenLaw workshop, it synthesizes shared terminology, technical characteristics, legal issues, and research priorities. Its central outcome is an initial framework for interdisciplinary work, while its glossary and taxonomy remain provisional and its analysis is scoped rather than exhaustive.

  • Problem

    Law and ML communities use important terms differently, complicating interdisciplinary analysis of generative-AI issues.

    Source: §3 · p. 3

  • Method

    The report synthesizes two-day workshop discussions among legal and ML experts into conceptual, technical, legal, and research frameworks.

    Source: §1 · p. 2, §2 · p. 2, §7 · p. 16

  • Results

    The workshop produced an initial shared knowledge base, account of generative AI’s distinctive aspects, taxonomy of legal issues, and research agenda.

    Source: §7 · p. 16

  • Takeaways & Limitations

    Interdisciplinary progress requires common vocabulary, technical and legal understanding, and coordinated research on how generative AI interacts with law.

    Source: §7 · p. 16

    The report is non-exhaustive and does not provide detailed analyses of specific systems, lawsuits, or regulatory policy proposals.

    Source: §3.1 · p. 4, §6.3 · p. 15

Abstract

from arXiv · show

This report presents the takeaways of the inaugural Workshop on Generative AI and Law (GenLaw), held in July 2023. A cross-disciplinary group of practitioners and scholars from computer science and law convened to discuss the technical, doctrinal, and policy challenges presented by law for Generative AI, and by Generative AI for law, with an emphasis on U.S. law in particular. We begin the report with a high-level statement about why Generative AI is both immensely significant and immensely challenging for law. To meet these challenges, we conclude that there is an essential need for 1) a shared knowledge base that provides a common conceptual language for experts across disciplines; 2) clarification of the distinctive technical capabilities of generative-AI systems, as compared and contrasted to other computer and AI systems; 3) a logical taxonomy of the legal issues these systems raise; and, 4) a concrete research agenda to promote collaboration and knowledge-sharing on emerging issues at the intersection of Generative AI and law. In this report, we synthesize the key takeaways from the GenLaw workshop that begin to address these needs. All of the listed authors contributed to the workshop upon which this report is based, but they and their organizations do not necessarily endorse all of the specific claims in this report.

1 Introduction

The GenLaw workshop brought machine-learning and legal experts together to examine generative AI’s legal challenges, focusing initially on U.S. intellectual-property and privacy issues. Its report synthesizes discussions into a shared conceptual foundation, technical clarification, legal taxonomy, and collaborative research agenda.

1 Introduction

  • The workshop convened ML and law experts to discuss legal challenges raised by generative-AI technology, with initial emphasis on U.S. law, intellectual property, and privacy.Other discussed topics included free speech, products liability, and transparency.Source: §1 · p. 2
  • The two-day format combined a public ICML session with keynotes, panels, talks, posters, and subsequent roundtables for deeper discussion.Approximately forty participants joined the second-day roundtables.Source: §1 · p. 2
  • The report organizes workshop takeaways around five needs: explaining significance and difficulty, building shared language, clarifying capabilities, mapping legal issues, and setting a research agenda.These headings reflect participants’ consensus about the area’s most urgently needed contributions.Source: §1 · p. 2
  • The report synthesizes interdisciplinary reflections rather than detailed system analyses, lawsuit accounts, or regulator policy proposals.It targets scholars and practitioners already familiar with some technical or legal background.Source: §1 · p. 2

2 The Impact of Generative AI on Law

Generative AI is significant for law because it combines broad generativity with leverage, adaptability, ease of mastery, accessibility, and transferability. These properties make its legal and policy challenges unusually broad while requiring shared technical and legal understanding.

2 The Impact of Generative AI on Law

  • Generative AI’s generativity combines leverage, adaptability, ease of mastery, accessibility, and transferability across users and applications.The report characterizes generative AI as enabling unanticipated change through broad and varied contributions.Source: §2 · pp. 2–3
  • Generative-AI services are broadly available and often inexpensive and near-real-time, although creating cutting-edge models still requires substantial data, compute, and expertise.This creates a distinction between concentrated model creation and widely accessible inference services.Source: §2 · p. 3
  • Because models, prompts, and prompting techniques can be shared readily, users can build on one another’s innovations and improve systems rapidly.The report describes this combination as making generative AI unusually generative.Source: §2 · p. 3
  • The report compares generative AI’s anticipated legal and policy challenges in scope, scale, and complexity to those raised by computers and the Internet.It presents generative AI as potentially the third supremely generative technology after those examples.Source: §2 · p. 3
  • Addressing these challenges requires shared vocabulary, technical literacy for lawyers, legal literacy for technologists, and a common research agenda.The report presents this framework as a starting point for interdisciplinary collaboration.Source: §2 · p. 3

3 Developing a Shared Knowledge Base

The report argues that interdisciplinary progress requires shared concepts, careful metaphors, and up-to-date understanding of how generative-AI systems are built and used. It proposes glossaries and other knowledge-building practices while emphasizing that terminology and business models remain unsettled.

3 Developing a Shared Knowledge Base

  • Terminology differences between law and ML—including privacy, pre-training, and harms—created misunderstandings that impeded collaboration.Participants often reached common understandings only after false starts and clarification.Source: §3 · p. 3
  • A shared baseline should clarify ambiguous or overloaded terms through glossaries, cross-disciplinary metaphors, and continuing attention to technical developments and deployments.These practices are presented as complementary ways to improve communication across disciplines.Source: §3 · p. 4

3.1 Identifying and Defining Terms

  • The initial glossary defines technical terms and critical concepts for non-experts, while emphasizing that legal concepts such as privacy extend beyond removing personally identifiable information.Its definitions are intentionally concise rather than exhaustive.Source: §3.1 · p. 4
  • The glossary is a starting point because terminology will evolve as technologies and controversies change.The organizers plan to update it as the field develops.Source: §3.1 · p. 4

3.2 Crafting Useful Metaphors

  • Metaphors can clarify complex concepts, but they may also simplify or distort them, so understanding where an analogy fails remains instructive.The report discusses anthropomorphic terms such as models “learning” or “memorizing” as technical terms rather than human-like processes.Source: §3.2 · pp. 4–5

Memorization.

  • Machine-learning memorization refers to technically defined properties of model weights or outputs, not deliberate human remembering or intent.Stylistic similarity in text-to-image systems is not equivalent to memorization in this technical sense.Source: Memorization · p. 5

3.3 Understanding Evolving Business Models

  • Understanding business models requires tracking the diverse actors and arrangements involved in datasets, diagnostics, training, deployment, hosted services, integrations, and open models.The report notes that this landscape is still evolving and includes both proprietary and open or mixed offerings.Source: §3.3 · pp. 5–6

4 Pinpointing Unique Aspects of Generative AI

Generative AI differs from earlier systems through flexible open-ended outputs, modern training pipelines, interconnected supply-chain choices, and massive scale. These developments help explain both its broad capabilities and the need for more precise legal analysis.

4 Pinpointing Unique Aspects of Generative AI

  • The report frames these technical developments as the answer to what makes Generative AI appear “magical” and legally distinctive.The section organizes that explanation around flexible models, modern training pipelines, supply-chain design, and scale.Source: §4 · pp. 6–7

4.1 The Transition to Very Flexible Generative Models

  • Generative AI combines open-ended content generation with general-purpose models that handle many tasks across modalities.Its applications include image captioning, music and speech generation, programming assistance, protein folding, drug design, and materials science.Source: §4.1 · p. 7

4.2 Developments in the Training Pipeline: Pre-Training and Fine-Tuning

  • Pre-training builds broad capabilities from enormous datasets, while fine-tuning adapts base models to particular domains, formats, or goals.Pre-training is expensive and usually occurs rarely; fine-tuning uses smaller datasets and can occur many times, although the boundary between the stages is practice-dependent.Source: §4.2 · pp. 7–8

4.3 Generative-AI Systems and the Supply Chain

  • Generative-AI systems involve many supply-chain actors and intervention points whose technical and sociotechnical design choices can carry distinct legal implications.These choices include training data, deployment, filters, rate limits, access controls, terms of use, API policies, and interface design.Source: §4.3 · p. 8

4.4 The Massive Scale of Generative-AI Models

  • Scaling pre-existing techniques, datasets, architectures, and computing systems has enabled the quality and variety of contemporary generations.Modern models use substantially more data and model capacity than earlier systems, while the cost of training a state-of-the-art model can reach hundreds of thousands or millions of dollars.Source: §4.4 · pp. 8–9

5 A Preliminary Taxonomy of Legal Issues

The report offers an interim, non-exhaustive taxonomy of legal issues arising from Generative AI, beginning with inherited AI/ML and speech-related concerns. It focuses on intent, privacy, misinformation and disinformation, and intellectual property, where generative capabilities create distinctive questions.

5 A Preliminary Taxonomy of Legal Issues

  • The taxonomy is explicitly preliminary and scoped to privacy and intellectual property, so it does not exhaust legal or policy concerns.The report distinguishes legal issues from the broader set of capabilities, consequences, risks, and harms associated with Generative AI.Source: §5 · p. 9

5.1 Intent

  • Generative AI can produce human-like harms without human intention, making established approaches to intent and responsibility difficult to apply.Respondeat superior may shift responsibility to a user, but it does not eliminate the underlying question of whether a tort occurred or what intentions matter.Source: §5.1 · p. 10

5.2 Privacy

  • Privacy analysis is complicated by the tension between computationally quantifiable definitions and law’s contextual reliance on social norms and reasonable expectations.Large web-scraped datasets may contain private information that models memorize, recombine, and reveal through generations or adversarial prompts.Source: §5.2 · pp. 10–11

5.3 Misinformation and Disinformation

  • Generative AI can produce plausible false content at scale, creating or amplifying legal concerns involving defamation, elections, impersonation, consumer protection, deepfakes, and other speech-related harms.Training data may contain misinformation, while fine-tuning or alignment can also be used deliberately to skew models toward misleading outputs.Source: §5.3 · p. 11

5.4 Intellectual Property

  • Generative AI further complicates intellectual property law by challenging human authorship and inventorship, the idea-expression boundary, human volition, market effects, trade secrecy, and scraping.The prompt and generation may be viewed as intertwined, while memorized training data and intentionally elicited outputs raise additional infringement and secrecy questions.Source: §5.4 · pp. 11–12

6 Toward a Long-Term Research Agenda

The agenda calls for sustained interdisciplinary research linking generative-AI design choices with legal analysis, standards, governance, and evaluation. It highlights unresolved questions about dataset licensing, model unlearning, and metrics that must track changing capabilities and contexts.

6 Toward a Long-Term Research Agenda

  • Technical design choices and legal scholarship should inform each other, creating a long-term agenda that can influence policy, education, and industrial practice.The report presents the agenda as a starting point for ongoing collaboration rather than a complete list of topics.Source: §6 · pp. 12–13

6.1 Centralization and Decentralization

  • Open, permissively licensed datasets require clear data provenance and coordinated technical, legal, and organizational innovation, but current openly licensed training efforts have mixed generation quality.The proposal aims to avoid alleged legal issues associated with web-scraped data while addressing collection, licensing, model-training, and stewardship challenges.Source: §6.1 · p. 13
  • Centralization versus decentralization is a legal, technical, business, and competition-policy question spanning data, compute, models, and infrastructure.Potential responses include antitrust enforcement, subsidies, open-access requirements, public-option infrastructure, export restrictions, and structural separation.Source: §6.1 · p. 13

6.2 Rules, Standards, Reasonableness, and Best Practices

  • Generative-AI creators’ and users’ duties may require evolving combinations of rules, standards, reasonableness, and best practices as capabilities and safety techniques change.What counts as reasonable is contextual, including differences between large and small actors.Source: §6.2 · p. 14
  • As technology stabilizes, concrete standards may define reasonableness, while interim best efforts and self-assessments can guide future standards-setting.Effective standards will differ by modality and system capabilities, including systems that use plugins or other APIs.Source: §6.2 · p. 14

6.3 Notice and Takedown̸ = Machine Unlearning

  • Removing a training example from a trained model is difficult because its influence is dispersed, impact and removal are ill-defined, and practical unlearning and attribution methods remain immature.Removal may require tracing and negating influence or retraining the entire model; current strategies are generally not computationally feasible for deployed systems.Source: §6.3 · pp. 14–15, §6.4 · p. 15

6.4 Evaluation Metrics

  • Evaluation metrics must reconcile scalable machine-learning measurement with legal decisions that depend on facts, context, and concepts such as memorization and removal.Different modalities and goals may require different definitions, and operationalizing societal or legal concepts mathematically remains difficult.Source: §6.4 · p. 15
  • Evaluation is dynamic because extraction methods, model use, fine-tuning, supply-chain structure, and alignment can change what should be measured.Fine-tuning may cause removed data to effectively resurface, although that possibility is described as speculative.Source: §6.2 · p. 14, §6.4 · p. 15

7 Conclusion and the Future of GenLaw

The conclusion broadens GenLaw’s agenda beyond copyright and calls for interdisciplinary evaluation, continually updated resources, and sustained engagement among technical, legal, policy, and public audiences. The glossary supports this work by organizing shared machine-learning, generative-AI, and legal terminology.

7 Conclusion and the Future of GenLaw

  • The report’s main topics are shared knowledge, Generative AI’s distinctive characteristics, a taxonomy of legal issues, and open research questions.These topics organize the workshop’s reported contributions.Source: §7 · p. 16
  • Copyright disputes cover only a small part of Generative AI’s potential legal questions, so the research agenda must evolve with rapidly changing technology and adoption.The report explicitly frames this as an overarching assessment of participants’ views.Source: §7 · p. 16
  • Useful evaluation methodologies require both machine-learning expertise and legal considerations, while also accounting for broader social, economic, educational, and interactional changes.The report rejects disciplinary boundaries as sufficient for designing comprehensive evaluations.Source: §7 · p. 16
  • GenLaw’s educational and engagement resources will need frequent updating because generative-AI systems and products continue to evolve.The intended engagement spans technologists, legal experts, policymakers, and the general public.Source: §7 · p. 16
  • GenLaw is developing a nonprofit home for research, education, and interdisciplinary discussion, with participation across 25 institutions and plans for broader global engagement.Its activities include accessible writing, expert resources, event recordings, external-resource collections, and a glossary.Source: §7 · p. 16

A Glossary

  • The glossary separates machine-learning and Generative AI terms from legal terms and alphabetizes entries within those sections.This organization supports the report’s goal of developing a common conceptual language.Source: A Glossary · p. 24

Alignment

  • Alignment means tuning a pre-trained model so its outputs follow a developer’s policy, whose goals may include instruction-following, human values, helpfulness, factuality, politeness, and copyright avoidance.Alignment can refer either to the tuning process or to the resulting state of the model.Source: Alignment · p. 26

Application Programming Interface (API)

  • An API lets users send company-specified queries to a company-operated model or system, typically under Terms of Service and additional security layers.This differs from open-sourcing a checkpoint for local use.Source: Application Programming Interface (API) · p. 26

Architecture

This glossary section defines core model-architecture concepts, explaining how architectural scale, attention, pre-training, checkpoints, and context windows shape model use and performance.

Architecture

  • Larger model families typically add layers and wider internal representations, which can improve capability while increasing training and inference time.The LLaMA family illustrates architectures that vary primarily in layer count and intermediate-vector size.Source: Architecture · p. 26

Attention

  • Attention selectively weights input tokens and describes each token’s influence on the token generated by a Transformer.It is a sub-component of transformer architecture.Source: Attention · p. 26

Base Model

  • A base model is pre-trained on a large general-purpose dataset and typically adapted through alignment, fine-tuning, or in-context learning before user exposure.GPT-3 and InstructGPT illustrate the transition from a base model to an instruction-following model.Source: Base Model · p. 27

Checkpoint

  • A checkpoint stores trained model parameters in long-term memory; public release of a checkpoint and loading code is described as the minimum bar for open-source status.Checkpoints are often written during training at regular intervals.Source: Checkpoint · p. 27

Context Window

  • A context window, or prompt length, is the maximum accepted input length, and exceeding it may degrade generation performance.The limit arises because training examples are bounded by that maximum length.Source: Context Window · p. 27

Data Curation and Pre-Processing

Data curation and pre-processing prepare training examples independently of model choice, while dataset selection depends on the intended task and data format.

Data Curation and Pre-Processing

  • Data pre-processing transforms raw examples into task-useful formats, such as removing HTML markup and navigation headers from crawled webpages.The transformation operates on the data before model training.Source: Data Curation and Pre-Processing · p. 27
  • Curation and pre-processing are model-independent, so different models can be trained on the same dataset, while dataset choice depends on the target task.Image generation from natural-language descriptions requires aligned image-text pairs.Source: Data Curation and Pre-Processing · p. 27, Datasets · p. 27

Datasets

This section establishes shared technical vocabulary for datasets, models, generation, training, and related machine-learning concepts. It also distinguishes generative systems from traditional predictive models and highlights why their outputs complicate evaluation.

Datasets

  • Datasets are collections of examples divided into training and testing portions, with test data measuring generalization to unseen but similar examples.They may be built from raw sources or assembled from existing datasets, and collection-based datasets often retain separate licenses for each constituent dataset.Source: Datasets · pp. 27–28

Decoding

  • Decoding algorithms generate language-model outputs by repeatedly selecting the next word from probabilities conditioned on the preceding prompt.Greedy, beam-search, and top-k are examples; decoding may also be called inference, generation, or sampling.Source: Decoding · p. 28, Language Model · p. 30

Diffusion-Based Modeling

  • Diffusion describes a model-training process, commonly used for image generation, rather than a standalone model architecture.Diffusion models typically involve an underlying neural network trained through diffusion processes.Source: Diffusion-Based Modeling · p. 28

Embedding

  • Embeddings represent data numerically to capture statistical relationships or semantic similarity, but they are not identical to the entities they represent.Similar vectors can reflect similar meanings according to a chosen distance metric.Source: Embedding · p. 28

Generation

  • Generative models produce human-interpretable outputs such as sentences or images, and unlike traditional regression or classification models, can generate multiple valid outputs for one prompt.This multiplicity makes performance evaluation difficult, while even high-quality generations may contain subtly incorrect details.Source: Generation · pp. 28–29

Fine-Tuning

Language Model

  • A language model predicts probabilities for the next token, and its decoding process can support tasks such as autocomplete, translation, and classification.Modern neural language models base probabilities on complex relationships in training data rather than simple sequence counts.Source: Language Model · pp. 29–30

Loss

  • Training minimizes a loss by repeatedly comparing model outputs with target outputs and updating parameters, while an objective specifies the mathematical reward or loss being optimized.Parameters are learned values saved in checkpoints, whereas hyperparameters are manually selected settings such as learning rate or architecture properties.Source: Loss · p. 30, Model · p. 31, Objective · p. 31

Pre-Training and Fine-Tuning

This section explains how modern generative-AI systems are trained and adapted, emphasizing the distinction between broad pre-training and task-specific fine-tuning. It also introduces tokenization, transformers, scale, human feedback, and the resulting supply-chain perspective.

Pre-Training and Fine-Tuning

  • Instruction tuning uses specially constructed examples of requests and responses to improve a model’s ability to follow instructions.Such interactions are relatively rare online compared with question-and-answer content.Source: Pre-Training and Fine-Tuning · pp. 31–32
  • Pre-training creates a general-purpose base model, while fine-tuning adapts an existing checkpoint to a task using smaller application-specific datasets and lower cost.The boundary between the two is not well-defined because systems may use more than two training stages.Source: Pre-Training and Fine-Tuning · p. 32

Regurgitation

  • Reinforcement learning incorporates human feedback about whether generations are good or useful, including feedback collected through system interactions or paid testers.ChatGPT’s thumbs-up and thumbs-down feedback is one example of this process.Source: Reinforcement Learning · p. 32

Scale

  • Scale refers to model parameters, training-data size, or computational requirements, all of which influence the resources needed for training.Scaling can also increase training-example fidelity, such as longer text sequences or higher-resolution images.Source: Scale · p. 32

Supply Chain

  • Generative-AI systems can be understood as supply chains involving multiple stages and actors in their creation, deployment, and use.The report points to prior work for a fuller account of this supply-chain framing.Source: Supply Chain · p. 32

Tokenization

  • Tokenization breaks inputs into vocabulary items with numeric IDs and learned embeddings, while subword tokenization handles rare or misspelled words.The number of tokens determines the effective context-window size.Source: Tokenization · p. 32

Transformer

  • Transformers process text, image, or music segments through layers that produce vector representations and learn how much attention to assign to input parts.Their training objective rewards reproducing target examples, with parameters updated through repeated training examples.Source: Transformer · p. 33, Training · p. 33

Training

This section describes how training encodes patterns into model parameters and clarifies how web data and openness shape generative-AI systems. It distinguishes practical availability from legal rights to reuse information.

Training

  • Stochastic gradient descent trains neural networks in batches by comparing outputs with targets and modifying parameters so correct outputs become more likely.The updates can be understood as leaving “echoes” of training examples encoded in model parameters.Source: Training · p. 33

Vector Representation

  • Common Crawl is a public web crawl used at least partly to train most open-source and many closed-source language models.Web crawls are expensive and compute-intensive, so most companies keep their own crawls private.Source: Web Crawl · p. 33

A.2 Open versus Closed Software

  • Openness has separate practical and legal dimensions: information must be publicly available and accompanied by rights permitting reuse.Because licenses vary, open versus closed is not a binary distinction and must be analyzed by the specific rights granted.Source: §A.2 · p. 33

Closed Dataset

This section distinguishes incomplete public information about datasets from several different meanings of a closed model or software system. It emphasizes that release status can vary across components and rights.

Closed Dataset

  • Many ostensibly open models rely on non-public datasets that cannot be fully reproduced because curation, preprocessing, or constituent sources are undisclosed.GPT-4 and PaLM are examples of models trained on non-public datasets.Source: Closed Dataset · p. 34
  • A model may be closed because it is inaccessible, available only through a restricted API or application, or publicly inspectable without a license permitting free use.These forms differ in how much users can inspect, reconstruct, or build upon the model.Source: Closed Model · p. 34

Closed Model

  • Closed-source software withholds source code, and systems can combine open and closed components, such as releasing a model while withholding its dataset.Older open-source licenses may not prevent such combinations.Source: Closed Software · p. 34, Open Dataset · p. 34

Open Dataset

The appendix distinguishes different meanings of openness for datasets, models, and software, then introduces core intellectual-property and privacy concepts relevant to Generative AI.

Open Dataset

  • Open datasets may be downloadable, restricted by licenses or Terms of Service, or accessible only through paid or institutional arrangements.Super-GLUE combines multiple constituent licenses, while some datasets require institutional affiliation or payment.Source: Open Dataset · p. 34

Open Model

  • Open models may provide downloadable checkpoints and inference software without releasing training data, training code, or development details.The machine-learning usage of “open-source” therefore differs from broader software-community expectations about inspectability, modification, and redistribution.Source: Open Model · p. 34, Open Software · p. 34

Open Software

  • Machine-learning systems contain separable dataset-building, model-training, and inference software components, each of which can be independently open-sourced.This decomposition makes “openness” a property that may apply differently across system components.Source: Open Software · p. 35

A.3 Legal Concepts in Intellectual Property and Software

  • Intellectual-property law covers exclusive rights in creative and inventive works, including copyright, patents, trademarks, trade dress, and trade secrets.Copyright protects original creative expression rather than facts or ideas, while patents protect qualifying inventions through claims defining their scope.Source: The Field of Intellectual Property (IP) · p. 35, Copyright · p. 35, Claims · p. 35

Harm

  • Legal remedies and applicability depend on recognized harm, with damages potentially covering lost profits, reputational damage, or licensing fees.Standing and remedies may be limited to identifiable harms suffered by plaintiffs, and legally cognizable harm need not match intuitive notions of harm.Source: Harm · p. 35, Damages · p. 35

Copyright

  • Copyright questions distinguish protected expression from unprotected ideas and may involve infringement, fair use, transformative use, licenses, and non-expressive computational uses.Relevant analyses include whether material was used without permission, whether a use is limited or transformative, and whether computation such as text mining displays the original work.Source: Copyright Infringement · p. 35, Fair Use · p. 35, Idea vs. Expression · p. 36

A.4 Privacy

  • Privacy concepts include anonymization, consent, personally identifiable information, and data-protection obligations such as those imposed by the CCPA and GDPR.Differential privacy seeks to prevent statistics from revealing whether an individual record was in a dataset, but added noise can reduce accuracy or model performance and creates challenges for unstructured generative-AI data.Source: Differential Privacy · pp. 36–37, The California Consumer Privacy Act… · p. 36, Consent · p. 36

The General Data Protection Regulation (GDPR)

The appendix introduces privacy and data-protection concepts spanning identifiable information, organizational privacy practices, privacy harms, and rights to remove outdated information.

Personally Identifiable Information (PII)

  • Personally identifiable information is data that can identify an individual, including contact, financial, and biometric information.Organizations collecting such information must safeguard its confidentiality and integrity under relevant data-protection laws such as the GDPR.Source: Personally Identifiable Information… · p. 37

Privacy Policy

  • A privacy policy explains how an organization collects, uses, stores, and protects personal information and informs individuals about related rights.It communicates the organization’s data-processing practices rather than defining privacy protection as a single technical measure.Source: Privacy Policy · p. 37

Privacy Violation

  • Privacy violations involve unauthorized or inappropriate intrusion into personal information or activities and may produce autonomy, reputational, or financial harms.Examples include data breaches, surveillance, identity theft, and sharing sensitive information without consent.Source: Privacy Violation · p. 37

The Right to be Forgotten

  • Some legal systems recognize a right to be forgotten, allowing individuals to request removal of personal information when public interest is outweighed by information that is outdated or misleading.The GDPR includes a form of this right.Source: The Right to be Forgotten · p. 37

The General Data Protection Regulation (GDPR)

  • A tort is a civil wrongdoing causing harm or injury to a person or property for which tort law provides remedies and compensation.This concept supplies a general legal framework for responding to harmful conduct.Source: Tort · p. 37

B Metaphors

The report uses metaphors to explain how generative-AI systems are trained and generate outputs, while emphasizing where those analogies diverge from human or artistic processes.

Models are trained

  • Model training resembles reinforcement through rewards, but unlike dog training it typically lacks a progression from easier to harder skills.The reward is computed from similarity between model outputs and desired outputs, and the formula generally remains unchanged during training.Source: B Metaphors · p. 38

Models learn like children do

  • The “models learn like children” metaphor highlights imitation but obscures the different mechanisms involved in model scaling and human abstraction.Increasing model size has no child-development parallel, while children can extract novel and abstract structures beyond statistical patterns.Source: B Metaphors · p. 38

Generations are collages

  • The collage metaphor is misleading because models encode patterns from many examples rather than explicitly selecting, coordinating, and arranging source works during generation.A generation may nevertheless resemble a collage of several examples, even though its production process differs from an artist’s.Source: B Metaphors · p. 38

Large language models are stochastic parrots

  • The stochastic-parrot metaphor captures models’ tendency to output frequent training data but is criticized for undervaluing their competencies and prompting debate about human-like understanding.Passing exams such as the Bar Exam or GRE [77] may still be interpreted as parroting training data rather than human-like skill.Source: B Metaphors · p. 38

Large language models are noisy search engines

  • The noisy-search-engine metaphor distinguishes language-model prediction from database retrieval because models lack direct access to training data and often produce unattributed or incorrect mixtures of information.Some outputs may quote memorized entries, but that is not described as typical.Source: B Metaphors · p. 39
Loading 2311.06477v3…