AI Model Training Data Rights: Copyright, Fair Use, and Licensing

The legal question of whether training AI models on copyrighted works constitutes infringement has moved from academic debate to courtroom reality. With billions of dollars in damages at stake and several cases now decided at summary judgment, AI companies face immediate decisions about data sourcing, licensing strategy, and risk mitigation. The early rulings do not point in one direction: courts have treated model training itself as strongly transformative while drawing a hard line at the use of pirated source copies.

This guide provides a practical framework for understanding copyright risks in AI training, evaluating fair use defenses, implementing compliant data sourcing practices, and navigating the evolving regulatory landscape.

The Legal Framework: Copyright Law and AI Training

Does Training on Copyrighted Data Constitute Infringement?

In the Office’s view, expressed in its May 2025 pre-publication report, building a training dataset using copyrighted works “clearly implicate[s] the right of reproduction”---making the act of copying presumptively infringing unless a defense like fair use applies. That framing is influential but not settled law: the document is a pre-publication version rather than a final report, and its institutional standing is contested, as the Register of Copyrights was removed shortly after its release. The practical point holds regardless: copying copyrighted works into a training dataset requires either authorization or a valid legal defense.

The copyright infringement analysis follows established principles:

What constitutes copying: Making intermediate copies of copyrighted works during data collection, preprocessing, and storage all constitute reproduction under Section 106 of the Copyright Act, regardless of whether those copies persist after training completes.

Rights holder standing: Copyright owners can sue for infringement even if their works represent a tiny fraction of a massive training dataset. Class action lawsuits aggregate claims from thousands of copyright holders, multiplying potential exposure.

Chain of liability: Companies face liability not only for works they directly copied, but also for works obtained from third-party datasets if those datasets were assembled through unauthorized copying (such as the LAION-5B dataset built from scraped images).

The Four-Factor Fair Use Test

Fair use provides the primary defense for unauthorized use of copyrighted training data. Courts evaluate four statutory factors:

Factor 1: Purpose and Character of Use

This factor examines whether the use is “transformative”---whether it adds new meaning, expression, or purpose rather than merely superseding the original work.

In the 2025 summary-judgment rulings, courts have found that training a general-purpose model on text can be strongly transformative---one judge described it as “exceedingly transformative.” The Copyright Office’s pre-publication analysis draws a related distinction:

• More transformative: Training general-purpose foundation models on large, diverse datasets to enable a wide range of outputs across different contexts and applications
• Less transformative: Training models to generate outputs “substantially similar” to specific training data or that “share the purpose of appealing to a particular audience”

The Copyright Office cautioned against treating AI training as inherently transformative simply because it involves computational analysis rather than human reading, noting that AI systems create perfect copies with instant analysis capabilities. Early case law has been more receptive to the transformative-use argument for the training step itself, while courts have separately scrutinized how the underlying copies were obtained.

Commercial purpose weighs against fair use but is not dispositive if other factors favor the defendant.

Factor 2: Nature of the Copyrighted Work

This factor considers the type of work copied:

• Highly expressive creative works (novels, artwork, music, photography): Courts afford these stronger protection, disfavoring fair use
• Factual or functional works (databases, technical documentation, news articles): These receive thinner protection, favoring fair use
• Published vs. unpublished works: Using unpublished works disfavors fair use, as authors have the right to control first publication

Most AI training involves published, creative works---a combination that ordinarily weighs against fair use, though it has not proven decisive in the recent text-training rulings.

Factor 3: Amount and Substantiality of Use

AI training typically involves copying entire works rather than excerpts or portions. While wholesale copying ordinarily weighs against fair use, the Copyright Office acknowledged that copying complete works “may be necessary” for certain types of training, particularly for general-purpose models.

Courts will evaluate whether copying entire works was reasonably necessary to achieve the transformative purpose, or whether training on smaller portions or samples would have sufficed.

Factor 4: Effect on the Market

The Copyright Office broadly interpreted this factor to encompass:

• Direct substitution: Whether AI-generated outputs replace sales of original works
• Market dilution: Whether proliferation of AI-generated content in similar styles undermines market value even without direct substitution
• Lost licensing opportunities: Whether the unauthorized use deprives copyright holders of potential licensing revenue

The existence or likely emergence of licensing markets for training data weighs against fair use. As licensing marketplaces proliferate and major publishers negotiate training data deals, courts may view unauthorized use as market displacement rather than fair use.

This factor has emerged as the decisive battleground. In Kadrey v. Meta, the court treated market dilution as the dispositive question and granted the defendant summary judgment only because the plaintiffs failed to develop a market-harm record---while signaling that a better-supported dilution theory could prevail in a future case. The lesson for both rights holders and developers is that Factor 4 outcomes are record-dependent, not foreordained.

Current Litigation Landscape: What Courts Are Deciding

The 2025 summary-judgment rulings do not line up neatly for either side. One court rejected fair use for a non-generative legal-research tool; two others found model training transformative while splitting on how the source copies were acquired and whether market harm was proven. The throughline is that outcomes turn on the specific use and the specific record, not on a categorical rule that AI training is or is not infringement.

The June 2025 Fair Use Rulings: A Split Emerges

Two decisions issued days apart in the Northern District of California reframed the field, both finding the training step transformative while diverging on the issues that decided each case.

Bartz v. Anthropic (Judge Alsup, June 23, 2025): The court issued a split summary-judgment ruling. It held that using books to train large language models is “exceedingly” and “spectacularly” transformative and therefore fair use, and that digitizing lawfully purchased print books (a format change, with the originals discarded) was also fair use. But it held that downloading and retaining pirated copies---more than seven million books pulled from pirate libraries to build a permanent central library---was not fair use and constituted infringement, sending the piracy question toward trial on damages.

Kadrey v. Meta (Judge Chhabria, June 25, 2025): The court granted Meta summary judgment on the training claim, finding that training its models on books (including pirated copies) was highly transformative. The ruling turned on Factor 4: the plaintiffs failed to prove market dilution, which the court treated as dispositive on this record. The court was careful to add that this outcome reflected a weak record, not a safe harbor---signaling that a well-developed market-dilution theory could succeed in a future case, and observing that such training will, in many cases, be unlawful.

Implications: Read together, these rulings establish that the training step has meaningful fair use support, but that the provenance of the source copies and the strength of the market-harm record can be decisive. A developer who trains on lawfully obtained materials stands on far firmer ground than one who builds its corpus from pirated libraries.

Thomson Reuters v. Ross Intelligence: Fair Use Rejected for a Competing Research Tool

In February 2025, a Delaware federal court issued an early decision rejecting fair use for AI training data. Thomson Reuters sued Ross Intelligence for training a legal research tool on Thomson Reuters’s proprietary Westlaw headnotes without authorization. The ruling was a revised decision that reversed the court’s own 2023 order, which had sent the fair use question to a jury.

Key holdings:

• Not transformative: The court found Ross’s use “not transformative” because the tool served the same purpose as Westlaw’s original content---legal research. Despite the technological sophistication, the functional purpose remained identical.
• Market harm: The court emphasized this as “the most important factor,” finding that Ross’s product directly competed with Westlaw and could harm both existing and derivative markets for legal research platforms.
• Direct infringement: The court granted partial summary judgment on direct copyright infringement.

Posture and limits: The ruling is now on interlocutory appeal to the Third Circuit, so it is best read as persuasive but not settled. Its reach is also narrower than it first appears: Ross’s tool was a non-generative legal-research product trained on Westlaw headnotes, and the court emphasized that Ross’s output competed directly with the source. Companies building generative models on broad, diverse corpora should be cautious about over-reading Ross, whose facts differ materially from the foundation-model context.

Andersen v. Stability AI: Artists’ Case Proceeds to Trial

In August 2024, Judge William Orrick allowed visual artists’ claims to proceed against Stability AI, Midjourney, DeviantArt, and Runway AI, with a jury trial now scheduled for April 5, 2027.

Key findings:

• Plausible infringement claims: The court found artists had “reasonably argued” that Stable Diffusion was built “to a significant extent on copyrighted works” and was “created to facilitate that infringement by design.”
• Storage theory: The court accepted the theory that AI systems storing copies of training data---even as learned patterns in model weights---may constitute copyright infringement.
• Discovery phase: The case proceeding to discovery could expose internal communications about training data sourcing and corporate decision-making regarding copyright compliance.

Implications: This decision validates the theory that training AI on scraped artwork without authorization may constitute infringement, even when the model doesn’t store pixel-perfect copies.

OpenAI Litigation: Consolidated in an MDL

OpenAI faces numerous copyright lawsuits from authors, news publishers, and other content creators:

Authors’ lawsuits: Sarah Silverman, Paul Tremblay, Ta-Nehisi Coates, Michael Chabon, and other authors allege OpenAI trained ChatGPT on their copyrighted books without permission.

The New York Times case: The Times seeks “billions of dollars” in damages for unauthorized use of articles to train GPT models. Earlier in the litigation, the court denied OpenAI’s motion to compel evidence about the Times’s own use of generative AI.

Current posture: The U.S. cases have been consolidated in a multidistrict litigation---In re OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-03143 (S.D.N.Y.), before Judge Sidney H. Stein. OpenAI’s motion to dismiss was largely denied, and the court ordered the production of roughly 20 million de-identified ChatGPT logs---a discovery development with significant implications for proving (or rebutting) infringement and output similarity.

International suits: Parallel suits have followed abroad, including in Canada and India.

GitHub Copilot: Code Training Under Scrutiny

Developers sued GitHub, Microsoft, and OpenAI in November 2022, alleging Copilot was trained on billions of lines of open-source code without complying with licensing terms.

Case status: The court (Judge Jon Tigar, N.D. Cal., No. 4:22-cv-06823) dismissed most claims but allowed two to proceed:

• Open source license violation
• Breach of contract

The court also dismissed the plaintiffs’ DMCA Section 1202(b) claims. That dismissal is now on interlocutory appeal to the Ninth Circuit (No. 24-6136), and the district court proceedings are stayed pending the appeal.

Key issue: Whether training on open-source code violates license conditions requiring attribution, even when the AI doesn’t reproduce code verbatim.

Implications: Even permissive open-source licenses impose conditions that AI training may violate. Companies training on open-source code must evaluate license compliance, not just copyright law.

Anthropic’s $1.5 Billion Settlement: Resolving the Piracy Claim

In September 2025, Anthropic agreed to pay $1.5 billion to settle authors’ claims---approximately $3,000 per work for roughly 500,000 copyrighted books used to train Claude. The settlement resolves the piracy liability that Judge Alsup found in Bartz; it does not resolve the fair use question on the training itself, which Anthropic won.

After an initial procedural denial without prejudice on September 8, 2025, the court granted preliminary approval on September 25, 2025, finding the settlement “fair, reasonable, and adequate.” The final fairness hearing was held on May 14, 2026, with a claims rate of approximately 93 percent. Final approval remains pending as of publication; the settlement is to be funded in installments.

Significance:

• Damages signal: The settlement amount---four times the statutory minimum of $750 per work---reflects market expectations for copyright liability even in negotiated resolutions, particularly where pirated source material is involved.
• Potential exposure: Before settlement, Anthropic faced potential statutory damages “in the tens of billions of dollars” for willful infringement tied to the pirated copies.
• Scope of resolution: Because the settlement addresses the piracy claim rather than the fair use of training, it confirms that how a company acquires its source copies can drive liability even where the training itself is defensible.

Risk Assessment: Evaluating Copyright Exposure by Data Source

Not all training data carries equal copyright risk. Companies should assess exposure based on source, authorization status, and intended use.

High-Risk Data Sources

Web-scraped content without permission:

• Risk level: Highest
• Why: Direct copying of copyrighted works without authorization or license
• Examples: Scraping news articles, blog posts, creative writing, artwork, photographs from websites
• Current status: Multiple lawsuits target this practice; courts viewing skeptically

Pirated or illegally obtained works:

• Risk level: Highest (willful infringement)
• Why: Copyright law provides enhanced damages (up to $150,000 per work) for willful infringement
• Examples: Books obtained from shadow libraries (Library Genesis, Sci-Hub), leaked datasets
• Current status: Bartz v. Anthropic held that downloading and retaining pirated copies is not fair use even where the training itself is transformative, and that holding underlies the $1.5 billion Anthropic settlement

Licensed works used beyond scope:

• Risk level: High
• Why: Breach of contract claims plus copyright infringement if license doesn’t permit AI training
• Examples: Subscribing to database for research but using it for commercial AI training
• Current status: Thomson Reuters case illustrates liability even when defendant had legitimate access

Moderate-Risk Data Sources

Public datasets of uncertain provenance:

• Risk level: Moderate to High
• Why: Many popular datasets (CommonCrawl, LAION-5B) contain copyrighted works obtained through scraping
• Examples: Image-text pairs, web corpora, code repositories
• Current status: Downstream users may face liability for infringement in dataset assembly
• Mitigation: Investigate dataset creation methodology; prefer datasets with documented copyright clearance

Open-source code with restrictive licenses:

• Risk level: Moderate
• Why: Many open-source licenses require attribution, notices, or share-alike provisions that AI training may violate
• Examples: GPL, AGPL, Creative Commons ShareAlike (CC BY-SA)
• Current status: GitHub Copilot litigation tests whether training violates license conditions
• Mitigation: Implement license compliance tracking; provide attribution mechanisms

User-generated content from platforms:

• Risk level: Moderate
• Why: Platform terms may authorize AI training, but users retain copyright in their content
• Examples: Reddit posts, Stack Overflow answers, social media content
• Current status: Platform terms protect the platform, not necessarily downstream AI developers
• Mitigation: Review platform terms; consider direct user consent mechanisms

Lower-Risk Data Sources

Licensed commercial datasets:

• Risk level: Low to Moderate
• Why: Explicit license authorization reduces copyright risk, though license compliance remains essential
• Examples: Shutterstock AI licensing, publisher agreements, specialized data vendors
• Cost: Substantial (millions to hundreds of millions for large-scale training)
• Mitigation: Negotiate broad license terms covering training, fine-tuning, and commercial deployment

Public domain works:

• Risk level: Low
• Why: No copyright protection eliminates infringement liability
• Examples: Works published 1930 and earlier (pre-1931), U.S. government works, Creative Commons Zero (CC0) works
• Limitations: Public domain datasets don’t represent contemporary culture; insufficient for most commercial applications
• Mitigation: Verify public domain status; beware of copyright restoration for foreign works

Permissively licensed content:

• Risk level: Low
• Why: Licenses explicitly permit broad use, including commercial applications
• Examples: Creative Commons Attribution (CC BY), MIT License, Apache License 2.0
• Limitations: Still requires license compliance (attribution, notices)
• Mitigation: Implement attribution systems; maintain license records

Company-owned or commissioned content:

• Risk level: Minimal
• Why: Company owns copyright or has explicit authorization from creators
• Examples: Internal documents, commissioned works with work-for-hire agreements
• Limitations: Volume typically insufficient for foundation model training
• Mitigation: Document ownership or assignment agreements

Special Consideration: Synthetic Data

Risk profile: Variable, requires careful analysis

Synthetic data---artificially generated by AI models rather than collected from real-world sources---has been promoted as a copyright-safe alternative. However, synthetic data is “no silver bullet” for several reasons:

Indirect infringement risk: If the model generating synthetic data was itself trained on copyrighted works without authorization, the synthetic data may “retain in substantial part the initial real-world data” and carry infringement liability downstream.

Training requirements: Producing useful synthetic data requires training a generator model on real-world data, creating the same copyright questions synthetic data purports to avoid.

Output similarity: Models trained on synthetic data may still generate outputs similar to copyrighted works from the original training data, particularly if the synthetic data preserves stylistic or structural patterns.

Practical value: Synthetic data quality degrades with each generation, limiting its utility for producing cutting-edge models.

Licensing Options: Building Compliant Training Datasets

As copyright litigation intensifies, licensing markets for training data have emerged rapidly. Companies have three primary licensing strategies:

Commercial Data Licensing

Major content providers now offer training data licenses:

News publishers:

• Providers: News Corp, Financial Times, Axel Springer, Le Monde
• License structure: Typically negotiated deals with major AI companies
• Cost range: Tens to hundreds of millions of dollars for comprehensive access
• Recent deals: OpenAI agreements with multiple publishers; Google’s deals with news organizations

Stock media platforms:

• Providers: Shutterstock, Getty Images, Adobe Stock
• License structure: API access to images, videos, and metadata with explicit AI training rights
• Cost range: Shutterstock reported $104 million in AI licensing revenue in 2023, projecting $250 million by 2027
• Advantages: High-quality, commercially licensed content with clear rights

Publisher licensing:

• Providers: HarperCollins and other major publishers exploring AI licensing
• License structure: Author-approved licensing programs for book content
• Status: Emerging market with unclear pricing and terms
• Challenges: Complex rights (author vs. publisher ownership) and collective action problems

Code repositories:

• Providers: GitHub, GitLab potentially developing formal licensing programs
• Current status: Most code training relies on open-source licenses rather than commercial agreements
• Future development: Expect evolution toward opt-in commercial licensing as litigation clarifies risks

Creative Commons and Open Licensing

Creative Commons licenses provide graduated permission levels:

CC0 (Public Domain Dedication):

• Permission: No restrictions on any use, including commercial AI training
• Attribution: Not required but may be appreciated
• Best for: Maximum flexibility with zero compliance overhead
• Examples: Many government datasets, scientific databases, Common Corpus

CC BY (Attribution):

• Permission: Use, modification, and commercial application permitted
• Requirements: Provide attribution to original creator
• AI training considerations: Attribution requirement applies when publicly sharing works or adaptations; unclear whether trained models require attribution
• Examples: Many academic papers, educational resources, Wikipedia content

CC BY-SA (Attribution-ShareAlike):

• Permission: Use and commercial application permitted
• Requirements: Attribution plus derivatives must use same license
• AI training considerations: Unclear whether trained models constitute “derivatives” requiring ShareAlike; conservative approach treats models as subject to license
• Risk: Could require open-sourcing entire model if ShareAlike applies

CC BY-NC (Attribution-NonCommercial):

• Permission: Noncommercial use only
• Requirements: Attribution; all uses (including training) must be noncommercial
• AI training considerations: Commercial AI companies generally cannot use NC-licensed content, as both training and deployment involve commercial purposes
• Clarity: Most restrictive but clearest application---commercial AI training prohibited

CC BY-ND (Attribution-NoDerivatives):

• Permission: Redistribution permitted, modification prohibited
• Requirements: Attribution; no modifications or derivatives
• AI training considerations: Question whether training creates prohibited “derivatives”; conservative approach avoids these works
• Practical application: Often incompatible with AI training, which inherently transforms works

Important limitations: Creative Commons itself states that “using a more restrictive CC license in an effort to prevent AI training is not an effective approach” because copyright law may permit AI training regardless of license restrictions. However, violating license terms creates contract-based liability even if copyright fair use might apply.

Public Domain Resources

Works published 1930 and earlier (pre-1931):

• Status: Definitively in U.S. public domain (the cutoff advances one year each January 1; 1930 works entered the public domain on January 1, 2026)
• Advantages: No copyright restrictions, no licensing costs
• Limitations: Historical works don’t reflect contemporary language, culture, or knowledge
• Examples: Classic literature, historical photographs, vintage artwork

U.S. government works:

• Status: Not subject to copyright under 17 U.S.C. § 105
• Advantages: Vast repositories of technical, scientific, and administrative content
• Limitations: Excludes contractor-produced works; state government works may have copyright
• Examples: Federal agency reports, NASA images, legislative materials

Common Corpus and curated public domain datasets:

• Description: Approximately 2 trillion tokens (roughly 2.27 trillion) across multiple languages, drawn from public domain and permissively licensed sources
• Advantages: Specifically curated for AI training with legal compliance
• Size: Significant but smaller than proprietary datasets
• Limitations: Represents only a portion of human knowledge; skews historical

Expiration and restoration: Copyright terms are complex. Works published 1929-1963 may have entered public domain through non-renewal, but determining status requires research. Foreign works may have copyright restored under GATT implementation, creating unexpected liability for seemingly public domain content.

Cost Analysis: Licensing vs. Litigation Risk

Licensed Training Data Costs

Foundation model training (100B+ parameters):

• Commercial licensing costs: $50M to $500M+ depending on data volume, exclusivity, and content type
• Breakdown:
  • News content: $10M-100M+ per major publisher
  • Stock media: $5M-50M for image/video datasets
  • Code repositories: Variable, often relies on open-source rather than commercial licensing
  • Academic publishers: Emerging market, likely $10M-100M range

Specialized or fine-tuning datasets:

• Commercial licensing costs: $100K to $10M depending on domain specificity and volume
• Examples: Medical imaging data, financial news, technical documentation
• Structure: Often per-dataset pricing rather than comprehensive agreements

Open-source and public domain:

• Direct costs: Zero for most resources (CC BY, CC0, public domain)
• Compliance costs: $50K-500K for legal review, license tracking, attribution systems
• Limitations: Insufficient volume/quality for leading-edge models

Copyright Litigation Exposure

Statutory damages:

• Standard range: $750-$30,000 per infringed work
• Willful infringement: Up to $150,000 per work
• Innocent infringement: Minimum $200 per work (rarely applied to commercial entities)

Scale calculations: If a foundation model is trained on 1 million copyrighted works (illustrative only):

• Minimum exposure (innocent): $200 per work × 1M works = $200 million
• Standard exposure: $750-$30,000 per work × 1M works = $750 million to $30 billion
• Willful infringement: $150,000 per work × 1M works = $150 billion

These figures are arithmetic illustrations of the statute’s per-work mechanics, not predictions; courts may question per-work aggregation across millions of dataset works.

Actual damages alternative: Plaintiffs may elect actual damages (provable economic harm) plus defendant’s profits attributable to infringement instead of statutory damages. In practice, proving causation between specific training data and company profits presents challenges, making statutory damages more common in copyright litigation.

Anthropic settlement signal: The roughly $3,000-per-work figure suggests a middle-ground resolution between the statutory minimum ($750) and the willful maximum ($150,000). As a class settlement rather than a merits ruling, it sets no binding precedent, but it offers a real-world reference point for negotiated valuations where pirated source copies are at issue.

Defense costs:

• Motion to dismiss: $200K-$500K for comprehensive briefing and arguments
• Discovery through summary judgment: $2M-$10M including document production, depositions, expert witnesses
• Trial: $5M-$20M+ depending on complexity and duration
• Appeals: $1M-$3M per appellate level

Total defense costs for major litigation: $10M-$35M+ even if defendant ultimately prevails, not counting settlement amounts or judgments.

Risk-Adjusted Analysis

For companies training foundation models, licensing costs of $50M-$500M compare favorably to litigation exposure of hundreds of millions to billions in damages plus tens of millions in defense costs.

Break-even calculation: Licensing becomes cost-effective if it eliminates:

• 10% probability of $500M judgment, or
• 50% probability of $100M judgment, or
• Certainty of $50M defense costs plus settlement

Most AI companies’ training practices face materially higher litigation risk than these thresholds, making proactive licensing economically rational for all but the smallest-scale operations---and the recent rulings sharpen one point in particular: clean provenance for source copies is among the cheapest forms of risk reduction available.

Strategic considerations beyond pure cost:

• Speed to market: Litigation delays product launches and fundraising
• Reputation risk: Copyright infringement allegations damage brand and customer relationships
• Enterprise sales: Many enterprise customers require vendor copyright indemnification
• Regulatory attention: Copyright violations attract regulatory scrutiny in the EU (AI Act) and U.S.

Compliance Framework: Building Defensible Training Data Programs

Data Sourcing Checklist

Implement systematic review for all training data sources:

• Document data provenance: Maintain records showing source, acquisition date, and legal basis for use
• Verify copyright status: Confirm works are public domain, licensed, or subject to fair use analysis
• Review licenses: For licensed content, verify license permits AI training and commercial use
• Implement filtering: Exclude high-risk categories (pirated works, recent creative content without authorization)
• Check opt-out lists: Respect robots.txt exclusions and explicit opt-out requests (required in EU under AI Act)
• Assess fair use: For unlicensed copyrighted works, document fair use analysis for each category
• Third-party datasets: Investigate assembly methodology and copyright compliance for datasets obtained from third parties

Documentation Requirements

Courts and regulators increasingly require transparency about training data. Maintain comprehensive documentation:

Dataset documentation:

• Source and acquisition methodology
• Copyright status of included works
• License terms and compliance measures
• Opt-out mechanisms and responses
• Data processing and filtering applied

Fair use analysis:

• Written analysis applying four-factor test to each data category
• Business justification for using copyrighted works
• Market analysis of licensing alternatives
• Technical guardrails to prevent infringing outputs

EU AI Act compliance (for companies operating in the EU):

• Public summary of training data contents, using the AI Office’s mandatory template (required for general-purpose AI models from August 2, 2025)
• Documentation of copyright reservations respected
• Measures to remove illegal content
• Synthetic data generation details if applicable

Technical Safeguards

Implement technical measures to reduce infringement risk:

Input filtering:

• Block pirated or unauthorized sources at data collection stage
• Implement allowlists of authorized sources rather than blocklists
• Respect robots.txt and meta tag exclusions
• Honor DMCA takedown requests for specific works

Output filtering:

• Monitor generated outputs for substantial similarity to training data
• Implement guardrails preventing reproduction of copyrighted works
• Use perplexity filters or similarity detection to flag potential infringement
• Provide content filtering options for enterprise customers

Attribution systems:

• For CC BY licensed training data, develop attribution mechanisms
• Consider optional source attribution for generated content
• Implement transparency features showing training data categories

Audit capabilities:

• Maintain ability to identify whether specific works were included in training
• Implement logging for data usage and model training runs
• Develop response protocols for copyright holder inquiries

Organizational Practices

Cross-functional compliance team:

• Legal counsel for copyright analysis
• Engineering leadership for technical implementation
• Data acquisition team for sourcing decisions
• Compliance officer for ongoing monitoring

Board-level governance:

• Regular reporting on copyright compliance and litigation risk
• Executive approval for material training data decisions
• Budget allocation for licensing and compliance infrastructure

Vendor management:

• Due diligence on third-party data providers
• Contractual representations about copyright compliance
• Indemnification provisions for data provider breaches

Internal training:

• Educate engineers and data scientists on copyright compliance
• Establish clear escalation procedures for legal questions
• Foster culture of compliance rather than “move fast and break things”

Robots.txt and Opt-Out Mechanisms

Legal Status of Robots.txt

The robots.txt protocol is a voluntary technical standard without inherent legal force in the United States. Website owners place robots.txt files at domain roots to communicate preferences about automated crawling, but U.S. law does not explicitly require respecting these files.

However, practical and legal considerations favor compliance:

EU legal requirements: Under the EU AI Act and Digital Single Market directives, ignoring opt-out mechanisms is explicitly illegal. Non-compliance risks regulatory enforcement and fines.

Terms of service: Violating website terms of service that incorporate robots.txt restrictions may constitute breach of contract or Computer Fraud and Abuse Act (CFAA) violations under certain circumstances.

Fair use implications: Courts may view deliberately ignoring explicit opt-outs as evidence weighing against fair use, particularly on Factor 4 (market harm) when copyright holders attempt to preserve licensing markets.

Industry practice: Major AI companies including OpenAI and Google have published robots.txt guidance and claim to respect exclusions, making non-compliance a competitive and reputational disadvantage.

Implementing Opt-Out Respect

Technical implementation:

• Check robots.txt before scraping any domain
• Implement user-agent-specific rules (GPTBot, CCBot, etc.)
• Honor both robots.txt and meta tag exclusions
• Maintain allowlist of domains confirming permission

Common AI crawler user agents:

• GPTBot (OpenAI)
• Google-Extended (Google AI training)
• CCBot (Common Crawl)
• Bytespider (ByteDance)
• Anthropic-AI (Anthropic)

Retroactive compliance: For training data already collected, some companies offer:

• Removal upon request from copyright holders
• Do-not-track mechanisms for future training runs
• Opt-out registries for artists, writers, and content creators

Documentation: Maintain records of opt-out compliance for regulatory and litigation defense purposes.

Strategic Recommendations for AI Companies

For Early-Stage Startups

Immediate priorities:

1. Audit existing training data: Document sources and legal basis for all data currently used
2. Implement filtering: Remove highest-risk categories (pirated works, recent unlicensed creative content)
3. Shift to lower-risk sources: Prioritize public domain, permissive licenses, and commercial licensing for critical datasets
4. Document fair use analysis: Prepare written analysis for any remaining unlicensed copyrighted content
5. Budget for licensing: Allocate capital for data licensing in proportion to litigation risk

Funding stage considerations:

Pre-seed/seed stage: Use public domain and permissively licensed data (CC0, CC BY) exclusively if possible; consider pre-trained base models from vendors who handle copyright compliance

Series A/B stage: Begin commercial licensing relationships; implement comprehensive compliance framework; prepare for enterprise customer due diligence

Series C+ stage: Robust licensing portfolio; industry leadership in compliance practices; potential advocacy for favorable regulatory frameworks

For Enterprise AI Deployments

Companies deploying rather than developing models face different risks:

Vendor due diligence:

• Require vendors to disclose training data sources and copyright compliance
• Obtain contractual representations about copyright compliance
• Negotiate indemnification provisions for copyright claims
• Audit vendor compliance practices periodically

Fine-tuning and customization:

• Apply same compliance framework to any custom training data
• Ensure proprietary data licensing permits AI fine-tuning
• Document employee-generated content as company-owned

Open-source models:

• Investigate training data sources for open models
• Assess reputational and legal risks even if development liability lies with model creators
• Prefer models with documented compliant training data

For International Operations

EU-specific requirements:

The EU AI Act entered into force on August 1, 2024, and its obligations for general-purpose AI models---including the training-data-summary requirement---began applying on August 2, 2025. The AI Office published the mandatory training-data-summary template in July 2025. Models already on the market before August 2, 2025 have until August 2, 2027 to come into compliance. Key obligations include:

• Public summary requirement: Publish a summary of training data contents, modalities, sizes, and sources, using the AI Office template
• Copyright compliance documentation: Demonstrate respect for copyright reservations and opt-outs
• Illegal content filtering: Document measures to remove illegal content from training data
• Synthetic data disclosure: If using synthetic data, explain generation methodology

Multi-jurisdictional strategy:

• Comply with strictest requirements (generally EU) for all markets
• Document compliance for each jurisdiction separately
• Monitor regulatory developments in key markets (UK, California, Canada)
• Consider jurisdiction-specific model variants if necessary

Looking Ahead: Regulatory and Legal Developments

Pending Legislation

AI-specific copyright legislation: Multiple proposals in Congress address AI training data rights, though passage timelines remain uncertain

State-level AI regulation: California, New York, and other states are considering AI bills that may address training data

Copyright Office Guidance Evolution

The May 2025 Copyright Office report is a pre-publication version and not the final word. Expect:

• A final version and revised guidance as case law develops
• Potential regulatory rulemaking on specific AI copyright issues
• Congressional requests for additional analysis as legislation advances

Industry Self-Regulation

AI industry associations are developing:

• Voluntary codes of practice for training data sourcing
• Licensing standards and best practices
• Opt-out registries and attribution systems

Companies participating in industry-led initiatives may receive favorable consideration from regulators and courts.

Litigation Timeline

Major cases proceeding through 2027:

• Andersen v. Stability AI: Jury trial April 5, 2027
• OpenAI MDL (S.D.N.Y.): Consolidated discovery underway, including the ordered production of roughly 20 million de-identified ChatGPT logs
• Thomson Reuters v. Ross: On interlocutory appeal to the Third Circuit
• GitHub Copilot: District proceedings stayed pending Ninth Circuit appeal of the DMCA dismissal
• Anthropic settlement: Final approval pending after the May 2026 fairness hearing

Court decisions in these cases will continue to clarify fair use boundaries and establish damages reference points.

Key Takeaways

1. Copying is presumptively infringing, but the case law is split: Reproducing copyrighted works to build a training dataset implicates the reproduction right, yet 2025 rulings found the training step itself transformative---so the question is whether a fair use or other defense applies, not whether copying occurred

2. Provenance of source copies matters as much as the training: Bartz v. Anthropic treated model training as fair use while holding that the use of pirated copies was infringing; clean acquisition of source material is now a central risk variable

3. Factor 4 is record-dependent: Kadrey v. Meta turned on the plaintiffs’ failure to prove market dilution, with the court signaling a stronger record could change the outcome---market-harm evidence, not categorical rules, drives results

4. Litigation costs are substantial: Even successful defense costs $10M-$35M+; resolutions range from roughly $3,000 per work to $1.5 billion in the aggregate (the Anthropic settlement)

5. Licensing markets are maturing: Commercial licensing options now exist across content types; costs range from millions to hundreds of millions but compare favorably to litigation exposure

6. Risk-based approach required: Assess copyright exposure by data source; prioritize lawfully obtained, licensed data for commercial applications; reserve fair use arguments for genuinely transformative uses

7. Technical safeguards are essential: Implement filtering, monitoring, and attribution systems to reduce infringement risk and demonstrate good faith

8. Documentation is critical: Courts and regulators demand transparency; maintain comprehensive records of data sourcing, legal analysis, and compliance measures

9. International compliance matters: The EU AI Act’s training-data transparency obligations for general-purpose models apply from August 2, 2025 (with a 2027 grace period for pre-existing models); U.S. companies with EU operations must comply

10. Monitor legal developments: Case law and regulatory guidance are evolving rapidly; quarterly compliance reviews are advisable

When to Seek Legal Counsel

Consult experienced AI copyright counsel when:

• Selecting training data sources for new models or significant training runs
• Evaluating fair use arguments for specific datasets
• Negotiating commercial licensing agreements with publishers or data vendors
• Responding to copyright holder inquiries or takedown requests
• Facing litigation or regulatory investigation
• Operating in multiple jurisdictions with varying copyright regimes
• Acquiring companies or technologies with uncertain training data provenance
• Developing enterprise sales where customers require copyright indemnification

---

Need AI Training Data Compliance Guidance?

Astraea Counsel advises AI companies on training data rights, copyright compliance, fair use analysis, and licensing strategies. We help AI startups navigate copyright risk while building compliant training datasets. Explore our AI & Emerging Tech services.

Related Resources

• California Frontier AI Law (SB 53) - State frontier-AI regulation framework
• Federal AI Regulation Landscape - Pending federal AI legislation
• AI & Emerging Technology Practice - Comprehensive AI legal counsel
• Regulatory Compliance Services - Navigate AI compliance requirements
• Contact Us - Discuss your training data compliance needs

---

Disclaimer: This article provides general information only and does not constitute legal advice. Copyright law application to AI training involves complex, fact-specific analysis. Consult qualified legal counsel for advice on your specific situation.

---

Sources

1. U.S. Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version, May 9, 2025), available at https://www.copyright.gov/ai/

2. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 1:20-cv-00613 (D. Del. Feb. 11, 2025) (Bibas, J., sitting by designation) (on interlocutory appeal to the Third Circuit)

3. Bartz v. Anthropic PBC, No. 3:24-cv-05417 (N.D. Cal. June 23, 2025) (Alsup, J.) (split summary judgment); related $1.5 billion class settlement (preliminary approval Sept. 25, 2025; final fairness hearing May 14, 2026)

4. Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417 (N.D. Cal. June 25, 2025) (Chhabria, J.)

5. Andersen v. Stability AI Ltd., No. 3:23-cv-00201 (N.D. Cal. Aug. 12, 2024) (Orrick, J.) (jury trial set for Apr. 5, 2027)

6. In re OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-03143 (S.D.N.Y.) (Stein, J.)

7. Doe v. GitHub, Inc., No. 4:22-cv-06823 (N.D. Cal.) (Tigar, J.); DMCA § 1202(b) dismissal on interlocutory appeal to the Ninth Circuit, No. 24-6136

8. 17 U.S.C. § 504(c) (statutory damages); 17 U.S.C. § 105 (U.S. government works); 17 U.S.C. § 104A (restoration of copyright in certain foreign works)

9. European Union, Regulation (EU) 2024/1689 (AI Act), in force Aug. 1, 2024; general-purpose AI obligations applicable Aug. 2, 2025; AI Office training-data-summary template (July 2025), available at https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai

10. Creative Commons, Understanding CC Licenses and AI Training: A Legal Primer (May 15, 2025), available at https://creativecommons.org/

11. PleIAs, Common Corpus (open dataset, ~2.27 trillion tokens), available at https://huggingface.co/datasets/PleIAs/common_corpus

12. BakerHostetler, Case Tracker: Artificial Intelligence, Copyrights and Class Actions, available at https://www.bakerlaw.com/services/artificial-intelligence-ai/case-tracker-artificial-intelligence-copyrights-and-class-actions/

13. Electronic Frontier Foundation, No Robots(.txt): How to Ask ChatGPT and Google Bard to Not Use Your Website for Training (Dec. 2023), available at https://www.eff.org/