Initial gather — 29 March 2026

Landmark Court Rulings #

• Bartz v. Anthropic: Landmark Ruling on Fair Use vs. Infringement (ArentFox Schiff) — June 2025: AI training on lawfully purchased books = “exceedingly transformative” fair use. Training on pirated copies = NOT fair use. First legal boundary between scraping and learning.
• The Bartz v. Anthropic Settlement (Kluwer Copyright Blog) — $1.5B settlement (largest copyright settlement in US history). Acquisition method matters enormously: partial fair-use victory on lawful copies, massive liability for pirated training data.
• Thomson Reuters v. Ross Intelligence: Court Shuts Down AI Fair Use Argument (Reed Smith) — February 2025: first US court decision on AI fair use. Rejected fair use because output directly competed with the copyrighted product. Commercial substitution = no fair use.
• Judge Allows NYT Copyright Case Against OpenAI to Go Forward (NPR) — Judge preserved core infringement claims. Could define whether mass news ingestion for chatbot training survives fair use scrutiny.
• NYT v. OpenAI Reshapes Data Governance and eDiscovery Strategy (Nelson Mullins) — Discovery process may force disclosure of exactly which copyrighted works were used in training. Threatens to expose AI companies’ data practices.
• Two Courts Rule on Generative AI and Fair Use — One Gets It Right (EFF) — Contrasts Bartz (fair use for transformative training) with Thomson Reuters (no fair use for competitive substitution). Decisive question: does the output compete in the same market as the original?

Regulatory Frameworks #

• EU AI Act 2026: New Rules for Training Data and Copyright (Scalevise) — From August 2026: every GPAI provider must publish training dataset summaries, respect copyright opt-outs, and label AI-generated content. First binding training data transparency mandate.
• US Copyright Office Part 3 Report: Generative AI Training (US Copyright Office) — May 2025: commercial use of vast copyrighted troves for competing expressive content “goes beyond established fair use boundaries.” Stops short of recommending legislation.
• Copyright Office Weighs In on AI Training and Fair Use (Skadden) — The 108-page report treats AI training as non-inherently-transformative because models absorb “the essence of linguistic expression” — a significant departure from search-engine precedents.
• Where AI Regulation Is Heading in 2026 (OneTrust) — Converging landscape: EU AI Act full applicability August 2026, US state laws in CA/CO/NY, federal AI Transparency Act requiring training dataset disclosure from January 2026.

Opt-Out Mechanisms and Their Limitations #

• Why AI Opt-Out Systems Don’t Work (Copyright Alliance) — Structurally flawed: models already trained before creators learn about opt-out; robots.txt routinely ignored; works exist across multiple sites making per-copy reservation impossible.
• The EU AI Act and Copyrights Compliance (IAPP) — Article 53 requires GPAI providers to implement “appropriate technical mechanisms” for opt-out. First regulatory enforcement of opt-out as legal obligation.
• AI and the Commons — Creative Commons Preference Signals (Creative Commons) — Developing machine-readable “Preference Signals” for granular training preferences (non-commercial only, attribution required). Beyond binary opt-in/opt-out.

Data Licensing Marketplace #

• The Hidden Economy Behind AI: Data Licensing Takes Center Stage (Kaptur) — Market projected ~$460M (2025) to multi-billion by 2030. Perplexity: 37% of known deals, OpenAI: 29%. Shutterstock supplies images to Google, Meta, Amazon, Apple at $25-50M per deal.
• AI Content Licensing Lessons from Factiva and TIME (Digital Content Next) — Microsoft’s Publisher Content Marketplace as template for structured licensing: negotiated access, usage reporting, revenue-share.
• The False Hope of Content Licensing at Internet Scale (ProMarket, Stigler Center) — Argues licensing cannot scale to billions of works. Creates oligopolistic market favouring incumbents. Mandatory licensing risks becoming a tax on innovation.

AI-Generated Content Ownership #

• Copyright Ownership of Generative AI Outputs Varies Around the World (Cooley LLP) — Global patchwork: US denies copyright to purely AI-generated works, UK grants it to “the person who made the arrangements,” most jurisdictions unsettled.
• Who Owns AI Content? ChatGPT, Claude, Midjourney & Gemini Rights Compared (Terms.law) — All major platforms assign output ownership to users, but only Microsoft (Copilot) and Anthropic (enterprise) offer IP indemnification. Indemnity terms, not ownership clauses, are the real enterprise differentiator.
• All the Liability, None of the Protection (Paddo.dev) — AI-generated code: uncopyrightable by the developer, yet potentially infringing on training sources. Worst-of-both-worlds for enterprise users.

Training Data Transparency #

• Bringing Transparency to the Data Used to Train AI (MIT Sloan) — Researcher-built tool generating machine-readable summaries of dataset provenance. Prerequisite for EU AI Act and US transparency law compliance.
• Open Source AI Models: How Open Are They Really? (Hunton Andrews Kurth) — Models like DeepSeek R1 release weights but not training data: cannot be reproduced, audited, or verified for copyright compliance.
• Understanding CC Licenses and AI Training (Creative Commons) — CC licences have limited application to AI training because copyright law often already permits it. Restrictive CC licences are not an effective opt-out strategy.

Synthetic Data as Escape Valve #

• How Generative AI Is Revolutionizing Training Data with Synthetic Datasets (Dataversity) — Gartner predicts synthetic data >95% of image/video training data by 2030. Driven by copyright risk avoidance (70% reduction in privacy sanctions) and cost advantages.

Litigation Trajectory #

• AI and Copyright: The Cases and the Consequences (EFF) — Expanding copyright to require licensing would entrench Big Tech dominance (only they can afford it), shut out small developers, and undermine fair use.
• AI Trends for 2026: Copyright Litigation Shifts from Training Data to Outputs (Morrison Foerster) — 2026 frontier: liability shifts from “was training fair use?” (increasingly settled as yes) to “do AI outputs infringe?” Fundamentally changes risk calculus.

Cross-links #

• AI Impact on Society EFF argues licensing regimes entrench Big Tech dominance — distributional effects of copyright expansion.
• AI Impact on Society Regulatory fragmentation across jurisdictions (EU AI Act, US state laws, federal transparency act).
• Open vs Closed AI Ecosystems EU transparency mandates affect open-weight vs closed models differently. Licensing costs create barriers favouring closed labs. “Open” models don’t disclose training data.
• Vibe Coding Approaches AI-generated code sits in a “copyright void” — unprotectable yet potentially infringing.
• Applications of Vibe Coding Enterprise IP indemnity (only Microsoft and Anthropic offer it) is a key procurement factor. Enterprises bear infringement liability with no copyright protection.

Meta-observations #

• Emerging theme: Fair use is splitting along functional lines. General-purpose transformative training = fair use. Training that produces a direct market substitute = not. The decisive question is “does your output compete?” not “did you copy?”
• Emerging theme: Acquisition method matters as much as use. Bartz v. Anthropic drew a bright line: lawfully purchased = fair use, pirated = infringement. Data provenance is now a critical compliance concern.
• Emerging theme: Litigation is migrating downstream — from training to outputs. Next wave of risk falls on deployers, not just model builders. Major implications for enterprise adoption.
• Gap: Regulatory convergence is real but asymmetric. EU opt-out enforcement (Article 53) has no US equivalent — creating compliance divergence for global AI companies.
• Quality signal: The licensing market has a scaling paradox. Individual deals work for large publishers, but cannot scale to billions of works. Either compulsory licensing or fair use reaffirmation will be needed.
• Keyword suggestion: “copyright void” — the worst-of-both-worlds for AI-generated code (unprotectable + potentially infringing). Underappreciated enterprise risk.
• Source to watch: ProMarket (Stigler Center) — contrarian, data-backed analysis of IP market failures. High signal.