Global Divergence in AI Copyright Liability: A Comparative Analysis of Fair Use, Text and Data Mining, and Fair Dealing in the US, EU, and Caribbean (2025),

 Global Divergence in AI Copyright Liability: A Comparative Analysis of Fair Use, Text and Data Mining, and Fair Dealing in the US, EU, and Caribbean (2025),

 

By. Dr. Abiola Inniss Ph.D. LLM, ACIarb

 

Abstract

This article analyzes the diverging legal frameworks governing AI training and copyright in the United States, European Union, and Caribbean as of late 2025. In the United States, the judiciary has established a "conditional fair use" doctrine (Bartz v. Anthropic, Kadrey v. Meta), where training is transformative but liability arises from illicit data sources ("shadow libraries") or market substitution. The European Union enforces a statutory compliance regime under the AI Act, permitting text and data mining (TDM) only where rights holders have not exercised machine-readable opt-outs (e.g., C2PA). In the Caribbean, notably Barbados, legislative reforms prioritize creator sovereignty, rejecting broad TDM exceptions in favor of collective licensing. The analysis concludes that global AI liability is fractured: permitted as fair use in the US (with caveats), conditional on protocols in the EU, and largely infringing in the Caribbean, creating a complex compliance landscape for global AI development.

Executive Summary

As the global digital economy transitions into late 2025, the legal frameworks governing the intersection of Artificial Intelligence (AI) training and copyright law have fractured into distinct, often competing, regulatory paradigms. This report offers an exhaustive analysis of the judicial, legislative, and regulatory developments in the United States, the European Union, and the Caribbean regarding the legality of ingesting copyrighted works for the purpose of training generative AI models.

In the United States, 2025 has been a year of judicial refinement rather than legislative overhaul. The U.S. Copyright Office, in its pivotal Part 3 Report, declined to recommend a statutory blanket license, leaving the courts to define the boundaries of Section 107 "Fair Use".¹ The resulting case law, specifically the summary judgment rulings in Bartz v. Anthropic and Kadrey v. Meta, has established a "bifurcated" liability model. While the act of intermediate copying for the purpose of pattern recognition is generally viewed as transformative and fair, this defense collapses if the underlying data source is illicit (e.g., "shadow libraries"). Thus, US liability is increasingly defined by data provenance and output substitution rather than the training process itself.²³

In the European Union, the focus has shifted from judicial interpretation to strict regulatory compliance following the full implementation of the AI Act in August 2025. The EU regime is predicated on a statutory "Text and Data Mining" (TDM) exception under the Digital Single Market (DSM) Directive, which is heavily conditioned on a "machine-readable opt-out" mechanism. The legal battleground in Europe has moved to the technical layer, with German courts ruling in late 2025 that AI training is permitted under TDM exceptions unless rights holders have strictly adhered to emerging technical protocols like C2PA and TDM Reservation Protocols, as identified by the EUIPO.⁴⁵

In the Caribbean, the region is navigating a complex transition between colonial-era "Fair Dealing" statutes and the urgent need for digital sovereignty. Barbados has emerged as the legislative vanguard, passing the Copyright Bill, 2025 to empower Collective Management Organizations (CMOs) and extend protections, signaling a pro-creator stance that implicitly rejects a broad US-style fair use for AI.⁶⁷ Conversely, Jamaica and Trinidad and Tobago have focused on judicial regulation, issuing Practice Directions to curb AI "hallucinations" in court while retaining conservative copyright frameworks that likely classify unlicensed commercial AI training as infringement.⁸⁹

This article explores these diverging paths, analyzing the economic and legal ripple effects of a world where data training is "fair use" in California, a "conditional exception" in Berlin, and potentially "infringement" in Bridgetown.

 

1. The United States: The Judicial Refinement of Fair Use

The United States legal system, relying on the flexibility of common law, spent 2025 rigorously testing the applicability of the "Fair Use" doctrine (17 U.S.C. § 107) to the mass ingestion of copyrighted works. Unlike other jurisdictions that sought statutory fixes, the US approach has been defined by high-stakes litigation in federal courts, primarily the Northern District of California and the Southern District of New York.

1.1 The Policy Baseline: The US Copyright Office Part 3 Report

On May 9, 2025, the U.S. Copyright Office released its long-awaited report, Copyright and Artificial Intelligence: Part 3 – Generative AI Training. This document serves as the foundational policy text for the US government's stance on AI and copyright in 2025.¹

The Report effectively ended speculation regarding an immediate legislative "fix" such as a compulsory licensing scheme. Instead, the Office reaffirmed that existing copyright principles, specifically the fair use doctrine, should govern the training of AI models. However, the Office’s analysis was far from a blanket endorsement of the AI industry's practices. The Report emphasized that fair use is a context-specific, fact-intensive inquiry that cannot be determined in the abstract.¹

Crucially, the Report questioned the "transformativeness" of generative AI models in instances where the model's purpose is to generate content that competes directly with the training data. The Office noted that while the technical act of copying for analysis (intermediate copying) has historically been protected, the "purpose and character" analysis for Generative AI is complicated by the fact that the output often serves the same intrinsic purpose as the input—creative expression.¹ This "market substitution" concern became the central theme of subsequent litigation.

1.2 The "Bifurcated" Fair Use Doctrine: Bartz and Kadrey

The most significant legal development of 2025 was the emergence of a "bifurcated" fair use analysis, crystallized by two summary judgment rulings in June 2025: Bartz v. Anthropic PBC and Kadrey v. Meta Platforms, Inc. These cases dismantled the monolithic view of "AI training" and instead separated the act of training from the source of the data.

1.2.1 Bartz v. Anthropic: The Piracy Exception

In Bartz v. Anthropic, Judge William Alsup of the Northern District of California addressed whether Anthropic's use of copyrighted books to train its Claude models constituted fair use. The plaintiffs alleged that Anthropic utilized "The Pile," a massive dataset that included "Books3," a constituent dataset sourced from "shadow libraries" (pirated ebook repositories like Bibliotik).²

The court’s ruling on June 23, 2025, was nuanced. Judge Alsup agreed with Anthropic on the technical aspect of training, finding that the computational analysis of texts to extract statistical correlations and linguistic patterns was "highly transformative." This use served a fundamentally different purpose than the original books (which were meant for human entertainment), thus satisfying the first factor of fair use regarding the training process itself.¹⁰

However, the court denied summary judgment on the issue of the data source. Judge Alsup ruled that "Anthropic had no entitlement to use pirated copies for its central library" and that "Creating a permanent, general-purpose library was not itself a fair use excusing Anthropic's piracy".² This ruling established a critical precedent: Fair use cannot cleanse the taint of piracy. Even if the use (training) is transformative, the possession and ingestion of illicitly acquired data remains an infringing act. This decision forced AI developers across the US to audit their datasets and purge components derived from shadow libraries to avoid liability.²

1.2.2 Kadrey v. Meta: The "Market Competition" Theory

Two days later, on June 25, 2025, Judge Vince Chhabria issued a ruling in Kadrey v. Meta Platforms, Inc. that further tightened the fair use defense. Similar to Bartz, the plaintiffs (authors) alleged that Meta trained its LLaMA models on copyrighted books without permission. Meta argued that its use was transformative and therefore fair use as a matter of law.¹¹

While acknowledging the transformative nature of extracting information from books to build a language model, Judge Chhabria introduced a "market competition" theory that expanded potential liability. The court denied Meta's motion for total exoneration, explicitly noting that "fair use... typically doesn't apply to copying that will significantly diminish the ability of copyright holders to make money from their works".¹¹

The court emphasized that if an AI model creates a "substitute" for the original work, the transformative defense is weakened. This ruling shifted the burden of proof onto AI companies to demonstrate that their models do not negatively impact the market for the original works—a difficult bar to clear given the capability of LLMs to generate summaries, sequels, or stylistically similar prose that could theoretically displace the original author.¹¹

1.3 The "Substitution" Frontier: NYT v. OpenAI and Perplexity

Following the precedents set in the summer of 2025, the legal focus in late 2025 shifted from the training phase to the output phase, specifically regarding news and non-fiction content where the risk of substitution is acute.

1.3.1 The New York Times v. OpenAI

As of December 2025, this case remains in a contentious discovery phase. The core allegation is that ChatGPT regurgitates New York Times articles verbatim, serving as a direct market substitute for the newspaper's subscription service. In a significant development in late 2025, Judge Ona Wang ordered OpenAI to produce 20 million chat logs.¹² This discovery order aims to empirically test the "regurgitation" rate of the model. If the plaintiffs can prove that the model consistently outputs substantial portions of copyrighted articles, the "fair use" defense likely fails under the fourth factor (market effect), regardless of how transformative the training process was.

1.3.2 The New York Times v. Perplexity

On December 5, 2025, the New York Times opened a new front in the AI copyright war by suing Perplexity AI.¹³ This lawsuit differs from the OpenAI case as it targets "Retrieval-Augmented Generation" (RAG) rather than just model training. The Times alleges that Perplexity's search engine creates "substitutes" for its journalism by scraping real-time content and summarizing it for users, bypassing the paywall and depriving the publisher of ad revenue and subscriptions.¹³

This case challenges the "fair use" status of RAG systems. While search engines have historically enjoyed fair use protection (e.g., Authors Guild v. Google), the Times argues that Perplexity does not just index content but replaces it. A ruling against Perplexity could imperil the business models of AI search assistants that rely on real-time web summarization.¹³

1.4 Visual Arts and Compressed Copies: Andersen v. Stability AI

In the realm of visual arts, the case of Andersen v. Stability AI proceeded toward a September 2026 trial date, but significant rulings in 2025 shaped the landscape.¹⁴ The court allowed claims to proceed regarding "compressed copies," accepting the plausible theory that the AI model's weights might contain compressed representations of the training images, thereby constituting a derivative work or a form of storage infringement.¹⁴

This theory attacks the "abstraction" defense used by AI companies (which argues that models only store mathematical patterns, not images). If plaintiffs can prove that the model can reconstruct the original training images (memorization), the "fair use" defense for training becomes significantly harder to maintain, as it resembles storage and distribution rather than just learning.¹⁴

1.5 Synthesis of US Jurisprudence

By the end of 2025, the US legal position on AI training and fair use can be summarized as Conditional Fair Use:

1. Training Process: The intermediate copying of data for the purpose of pattern recognition is generally considered transformative and fair use.¹⁰
2. Data Source: The use of illicitly acquired data (e.g., shadow libraries) is infringement and is not shielded by fair use (Bartz).²
3. Market Effect: If the AI model generates outputs that serve as market substitutes for the original works (e.g., verbatim regurgitation, detailed summaries), the fair use defense is likely to fail (Kadrey, NYT).¹¹¹³

 

2. The European Union: The Statutory Compliance Regime

While the US navigated the nuances of common law, the European Union spent 2025 operationalizing a comprehensive statutory framework. The EU's approach is characterized by explicit legislative exceptions for "Text and Data Mining" (TDM) coupled with rigorous transparency and opt-out obligations mandated by the AI Act.

2.1 The AI Act and the DSM Directive: A Regulatory Pincer

The EU AI Act entered into full force in mid-2025, with key provisions regarding General-Purpose AI (GPAI) models becoming applicable on August 2, 2025.¹⁵ The AI Act does not create a new copyright exception; rather, it enforces the existing framework established by the Directive on Copyright in the Digital Single Market (DSM Directive) (2019/790).

The interaction between these two laws creates the EU's unique "Opt-Out" regime:

• Article 3 (DSM): Provides a mandatory exception for TDM for the purposes of scientific research by research organizations and cultural heritage institutions. This exception cannot be overridden by contract or opt-outs.¹⁶
• Article 4 (DSM): Provides an exception for TDM for any other purpose (including commercial AI training), unless the rightsholder has expressly reserved their rights in an "appropriate manner, such as machine-readable means".¹⁶

The AI Act bridges the gap by mandating that providers of GPAI models must (a) put in place a policy to comply with EU copyright law, and specifically (b) identify and respect the opt-outs exercised under Article 4(3) of the DSM Directive.¹⁷

2.2 The "Machine-Readable" Standard and Technical Protocols

The critical legal battleground in the EU in 2025 was the definition of "machine-readable means." Because Article 4 allows commercial TDM unless an opt-out is present, the technical validity of the opt-out determines the legality of the training.

In July 2025, the European Commission published the GPAI Code of Practice, which established the standards for compliance. Simultaneously, the EUIPO (European Union Intellectual Property Office) Study on Generative AI and Copyright provided the technical blueprint for these standards.⁴

The EUIPO study and the subsequent consultation identified a hierarchy of protocols:

1. Robots Exclusion Protocol (robots.txt): While widely used, the study noted its limitations (lack of granularity). However, the GPAI Code of Practice explicitly requires signatories to respect robots.txt as a baseline.¹⁷
2. TDM Reservation Protocol (TDM Rep): A W3C community group specification designed specifically for Article 4 compliance, allowing granular rights reservation at the file level.⁴
3. C2PA (Coalition for Content Provenance and Authenticity): A cryptographic standard for asserting provenance and rights, endorsed as a "state-of-the-art" method for expressing opt-outs in media files.⁴
4. Rights Reservation Databases: The study proposed "federated rights reservation databases" (e.g., via the Liccium Trust Engine or Valuenode) to allow rights holders to register opt-outs centrally, addressing the difficulty AI crawlers face in parsing every individual website.⁴

Implication: In 2025, an EU rightsholder who merely states "No AI Training" in a footer or Terms of Service without implementing a machine-readable protocol (like robots.txt or C2PA metadata) has arguably failed to validly opt-out under Article 4(3). Consequently, scraping their content for AI training would be legal under the TDM exception.⁴⁵

2.3 Judicial Validation: Germany as the Testbed

Germany, often the bellwether for EU copyright jurisprudence, provided the first major court rulings interpreting these provisions in late 2025.

2.3.1 Kneschke v. LAION (Hamburg Regional Court, September 2025)

In this landmark case, a photographer sued LAION (a non-profit that compiles datasets for AI training) for including his images in the LAION-5B dataset. The Hamburg Regional Court ruled in favor of LAION, finding that the creation of the dataset qualified as TDM for scientific research under Section 60d of the German Copyright Act (which implements Article 3 of the DSM Directive).¹⁸

Crucially, the court held that the creation of a dataset for AI training is intrinsically an act of "text and data mining." This validated the legal basis for non-profit research organizations to aggregate data for AI development without needing licenses, provided they fall under the scientific research definition.¹⁸

2.3.2 GEMA v. OpenAI (Munich Regional Court, November 2025)

In November 2025, the Munich Regional Court issued a ruling in GEMA v. OpenAI. GEMA, the German music collecting society, argued that OpenAI's use of lyrics constituted infringement. The court affirmed that the TDM exception (Article 4) generally covers the ingestion of data for AI training.¹⁹

However, the court introduced a critical caveat: Memorization equals Infringement. The court ruled that if the AI model is capable of reproducing the training data (e.g., outputting full song lyrics), it has exceeded the scope of the TDM exception. TDM allows for the extraction of patterns, not the reproduction of the work itself. Therefore, "overfitting" (where a model memorizes its training data) strips the developer of the TDM protection, rendering them liable for reproduction infringement.¹⁹

2.4 Synthesis of EU Jurisprudence

By late 2025, the EU position is a Statutory Opt-Out Regime:

1. Commercial Training: Permitted by default under Article 4 DSM, provided the developer respects machine-readable opt-outs (robots.txt, C2PA).⁵¹⁷
2. Scientific Research: Permitted unconditionally under Article 3 DSM for qualifying institutions.¹⁶
3. Liability Trigger: Liability attaches if the developer ignores valid opt-outs OR if the model generates infringing outputs (memorization/reproduction).¹⁹
4. Penalty: Violations are subject to the AI Act's massive fines (up to 7% of global turnover) and copyright damages.¹⁵

3. The Caribbean: Fair Dealing and Digital Sovereignty

In the Caribbean, the legal landscape in 2025 is marked by a tension between the legacy of colonial "Fair Dealing" laws and the modern necessity of digital sovereignty. Unlike the US "fair use" (which is open-ended and flexible) or the EU "TDM exception" (which is specific and statutory), Caribbean copyright laws generally require specific statutory exceptions for any use that is not authorized by the copyright holder.

3.1 The Regional Context: Fair Dealing vs. Fair Use

Most Caribbean jurisdictions (including Barbados, Jamaica, and Trinidad & Tobago) operate under a "Fair Dealing" framework derived from UK law. Fair dealing is traditionally limited to specific purposes: research, private study, criticism, review, and news reporting. It does not typically include a broad "transformativeness" test, nor does it automatically extend to commercial data processing.²⁰

In 2025, the lack of a specific "Text and Data Mining" exception in most Caribbean laws means that commercial AI training likely falls outside the scope of fair dealing, rendering it potentially infringing absent a license.²⁰

3.2 Barbados: The Legislative Vanguard

Barbados distinguished itself in 2025 by passing significant copyright reform, positioning itself as a leader in protecting creator rights in the digital age.

3.2.1 The Copyright Bill, 2025

Passed in October 2025, the Copyright Bill, 2025 replaced the outdated 1998 legislation. The bill was explicitly designed to address "digital content and streaming" and to "strengthen penalties for infringement".⁶²¹

Key provisions impacting AI and data rights include:

• Empowerment of CMOs: The Act formally empowers Collective Management Organizations (like COSCAP) to take direct legal action against infringers on behalf of their members. Previously, CMOs lacked the locus standi to sue, leaving individual creators to fight costly legal battles. This change significantly increases the litigation risk for AI companies that might ingest Barbadian music or literature without a license.⁶²¹
• Extension of Term: The Act extends copyright protection by 20 years, aligning with international trends and ensuring a longer period of exclusivity for rights holders.²¹
• No TDM Exception: Notably, despite global trends, the Bill did not introduce a broad fair use or TDM exception for AI training. By strengthening enforcement without creating an AI loophole, Barbados has effectively signaled a "permission-first" regime. AI developers wishing to train on Barbadian works (e.g., the extensive catalog of soca and calypso) must negotiate licenses.²⁰

3.3 Jamaica: Judicial Regulation and "Originality"

Jamaica has focused its 2025 regulatory efforts on the application of AI in the justice system rather than amending copyright law to facilitate AI development.

3.3.1 Practice Directions and "Hallucinations"

Chief Justice Bryan Sykes issued the Practice Direction on the Use of Artificial Intelligence in 2025. This regulation strictly prohibits the use of AI to draft affidavits, witness statements, or expert evidence to prevent the submission of "hallucinations" (fabricated case law or facts) to the court. While this regulates the legal profession, it reflects a broader judicial skepticism toward generative AI's reliability.⁸

3.3.2 Copyright Law and Originality

Jamaica continues to operate under the Copyright Act with traditional fair dealing exceptions (Sections 52-55) limited to "research or private study".²²

A critical jurisprudential barrier to AI in Jamaica is the "originality" requirement. The Jamaican Supreme Court, in Cabel Stephenson v. Doreen Hibbert, reiterated that originality requires "skill and labour" from a human author.²³ This precedent implies that:

1. Inputs: Training on Jamaican works for commercial purposes is not "private study" and is likely infringement.
2. Outputs: AI-generated works are likely not copyrightable in Jamaica, as they lack the requisite human "skill and labour".²³

Additionally, Jamaica amended its Copyright Act to accede to the Marrakesh Treaty (facilitating access for the blind), showing a willingness to create humanitarian exceptions, but has notably refrained from creating commercial AI exceptions.²⁴

3.4 Trinidad and Tobago: The "Human Author" Stance

Trinidad and Tobago remains the most conservative of the three jurisdictions regarding AI copyright reform.

• Legislative Status: The Copyright Act contains no express exception for text and data mining or AI training. Section 24 lists specific exceptions, none of which broadly cover machine learning.²⁵
• Judicial Rulings: In May 2025, Justice Westmin James sanctioned attorneys for submitting AI-generated fictitious case law, reinforcing the judiciary's view of AI as a tool requiring strict human oversight.⁹
• Policy Stance: The prevailing legal consensus, supported by the UNESCO/CARICOM AI Policy Roadmap, is that the region must "safeguard intellectual property rights" rather than dilute them. Legal experts in Port of Spain emphasize that a "human hand" is required for copyright, and that the mass scraping of data for commercial gain infringes on the moral and economic rights of creators.²⁶

3.5 Regional Oversight: The Caribbean Court of Justice (CCJ)

The CCJ, the final appellate court for Barbados and other CARICOM states, issued Practice Direction No. 1 of 2025 regarding Generative AI. This direction mandates that any use of AI in court proceedings must be disclosed and verified. This high-level judicial oversight reinforces a regional norm: AI is a high-risk tool that must be subordinated to human verification and legal accountability, a stance that aligns with a restrictive view of AI's copyright privileges.²⁷

---

4. Comparative Analysis: The "Compliance Fracture"

The divergence in legal frameworks across these three regions creates a complex global compliance landscape for AI developers. The following table illustrates the "Compliance Fracture" as of December 2025.

Table 1: Comparative Liability for AI Training (Dec 2025)

 

Feature	United States	European Union	Caribbean (Barbados/Jamaica/T&T)
Primary Legal Doctrine	Fair Use (17 U.S.C. § 107) - Common Law	TDM Exception (DSM Arts 3 & 4) - Statutory	Fair Dealing - Statutory (Closed List)
Legality of Training (Open Web)	Likely Fair Use, provided the source is lawful and use is transformative (Bartz).10	Permitted by default, unless rightsholder opts out via machine-readable means.5	Likely Infringing for commercial use. No general TDM exception exists.20
Legality of Training (Shadow Libraries)	Infringement. Fair use does not apply to illicitly acquired data (Bartz).2	Infringement. Requirement for "lawfully accessible" source (DSM Art 4).16	Infringement.
Opt-Out Requirement	Voluntary. No legal obligation to respect opt-outs (though often respected for PR).	Mandatory. Must respect machine-readable protocols (robots.txt, C2PA).17	N/A. Permission is required by default (Opt-In regime).
Liability Focus	Provenance & Output (Did you steal the data? Does output substitute the original?)	Process & Compliance (Did you check for opt-outs? Did you record the training data?)	Authorization (Did you obtain a license from the CMO?)
2025 Key Development	Bartz & Kadrey bifurcated fair use; NYT discovery on regurgitation.12	Full application of AI Act; German courts define "memorization" as infringement.19	Barbados Copyright Bill 2025 empowers CMOs; Jamaica/T&T regulate AI in courts.69

4.2 Second-Order Insight: The "Data Laundering" Risk and Regulatory Arbitrage

The stark difference between the EU's "Opt-Out" regime and the US "Conditional Fair Use" creates a significant risk of regulatory arbitrage.

• The "Havana" Harbor: AI developers might be tempted to move training operations to jurisdictions with broad TDM exceptions (like Japan, Singapore, or potentially Israel) to avoid the EU's strict transparency requirements or the US's litigation risks.²⁰
• The "Shadow Library" Trap: However, the US ruling in Bartz acts as a global restraint. Since the court ruled that the possession of an illicit dataset (like Books3) negates fair use, a company cannot simply "wash" pirated data in a permissive jurisdiction and then deploy the model in the US. The "fruit of the poisonous tree" doctrine effectively extends US copyright standards to the global supply chain of any AI model intended for the American market.²

4.3 The "North-South" Divide in AI Copyright

The Caribbean's resistance to adopting US-style fair use or EU-style TDM exceptions highlights a growing "North-South" divide.

• The Global North (US/EU): Generally favors exceptions that allow their massive tech industries to ingest data at scale, viewing data as a "raw material" for innovation.
• The Global South (Caribbean): Views their cultural output (music, literature, art) as a primary economic asset. By rejecting TDM exceptions and strengthening CMOs (as Barbados did), these nations are asserting that their data is a "finished product" requiring compensation.⁶²¹
• Economic Implication: This creates a barrier to entry for local Caribbean AI startups, who cannot rely on "fair use" to train models on local content. They face high transaction costs (licensing) that their Silicon Valley competitors (who might ignore Caribbean jurisdiction or claim fair use in the US) do not. This could ironically entrench the dominance of foreign AI models in the Caribbean region.

---

5. Future Outlook (2026 and Beyond)

As we look toward 2026, the legal conflict is evolving from "Can we train?" to "What can we output?"

• The "Substitution" War: The filing of NYT v. Perplexity in late 2025 signals that the next great legal battle will be over Retrieval-Augmented Generation (RAG). If courts rule that summarizing real-time news is "substitution" and not "fair use," the entire business model of AI search could collapse or be forced into a licensing model.¹³
• Technical Standardization: In the EU, the focus will be on the adoption of the TDM Reservation Protocol and C2PA. We expect 2026 to be the year where "machine-readable rights" become a standard part of file metadata, forcing a technological overhaul of how content is published on the web.⁴¹⁷
• Caribbean Licensing Markets: With Barbados empowering CMOs, we expect to see the first test cases of Caribbean collective management organizations demanding royalties from global AI platforms for the use of Caribbean music and literature. This could lead to new cross-border licensing agreements or, alternatively, the "geoblocking" of Caribbean content from AI training sets.⁶

6. Conclusion

In 2025, the answer to "Does AI training constitute Fair Use?" depends entirely on where the server—and the lawyer—is located.

• In the US: Yes, it is Fair Use, if you didn't steal the data and your AI doesn't plagiarize the output. The courts have carved out a space for innovation but fenced it off from piracy and direct market competition.¹⁰¹¹
• In the EU: Yes, it is a TDM Exception, if you respect the technical "No Trespassing" signs (opt-outs) and follow the bureaucratic map of the AI Act.⁵
• In the Caribbean: No, it is likely Infringement. The region has chosen to prioritize the economic rights of its creators over the unbridled expansion of AI training, betting that strong copyright enforcement will ultimately yield better economic returns than permissive open-ended exceptions.⁶

The "Fair Use" consensus of the early 2020s has shattered. In its place is a complex, fragmented global map of compliance, where data provenance, technical protocols, and jurisdictional boundaries define the legality of artificial intelligence.

7. References

 

Ref ID	Title / Jurisdiction	Description / Source
1	US Copyright Office Part 3 Report	Copyright and Artificial Intelligence: Part 3 – Generative AI Training (May 9, 2025). Defined US policy on fair use for AI training. 11
2	Bartz v. Anthropic (Ruling)	Bartz v. Anthropic PBC, No. 24-cv-05417 (N.D. Cal. June 23, 2025). Ruled that training on pirated "shadow library" data (Books3) is not fair use. 3
3	Bartz v. Anthropic (Settlement)	Settlement details ($1.5 billion) following the class certification and summary judgment ruling.
4	EUIPO GenAI Study 2025	Study on Generative Artificial Intelligence and Copyright (2025). Identifies protocols like TDM Rep and C2PA for EU opt-outs. 2828
5	EU TDM Protocols	European Commission consultation on TDM opt-out protocols under the AI Act and GPAI Code of Practice. 2925
6	Barbados Copyright Bill 2025	Copyright Bill, 2025 (Passed Oct 2025). Updates copyright law, empowers CMOs, does not include broad AI/TDM exceptions. 87
7	Barbados Senate Passage	Confirmation of the Copyright Bill's passage in the Barbados Senate (Oct 2025). 30
8	Jamaica AI Practice Direction	Practice Direction on the Use of Artificial Intelligence (2025) issued by Chief Justice Bryan Sykes. 9
9	Trinidad AI Court Ruling	Justice Westmin James sanctions attorneys for submitting AI-generated fake cases (Nexgen Pathology Services Ltd v Darceuil Duncan).31
10	Bartz v. Anthropic (Training)	Ruling that the act of intermediate copying for pattern recognition is transformative and fair use. 313
11	Kadrey v. Meta	Kadrey v. Meta Platforms, Inc. (N.D. Cal. June 25, 2025). Denied summary judgment for Meta based on "market substitution" theory. 1432
12	NYT v. OpenAI (Discovery)	Order compelling OpenAI to produce 20 million chat logs to test for regurgitation (Judge Ona Wang, Dec 2025). 16
13	NYT v. Perplexity	The New York Times v. Perplexity AI (Filed Dec 5, 2025). Lawsuit targeting RAG and AI search as market substitution. 1817
14	Andersen v. Stability AI	Andersen v. Stability AI (2025 Rulings). Allowed claims regarding compressed copies and model weights to proceed. 3319
15	EU AI Act Timeline	Implementation dates for GPAI models (August 2, 2025). 21
16	EU DSM Directive Arts 3 & 4	Analysis of the Text and Data Mining exceptions and the machine-readable opt-out requirement. 2523
17	GPAI Code of Practice	EU Code establishing robots.txt and other protocols as standards for AI Act compliance (July 2025). 2434
18	Kneschke v. LAION	Kneschke v. LAION e.V. (Hamburg Regional Court, Sep 2025). Ruled creation of datasets for AI training is TDM for scientific research. 355
19	GEMA v. OpenAI	GEMA v. OpenAI (Munich Regional Court, Nov 2025). Ruled memorization/reproduction of lyrics exceeds TDM exception. 265
20	Caribbean Fair Dealing Analysis	Analysis of "Fair Dealing" vs "Fair Use" in Commonwealth Caribbean jurisdictions and lack of TDM exceptions. 3637
21	Barbados CMO Powers	Legislative details on empowering COSCAP (Barbados CMO) to sue for infringement. 7
22	Jamaica Copyright Act	Sections 52-55 of the Copyright Act regarding fair dealing for research/private study. 38
23	Jamaica Originality Case	Cabel Stephenson v. Doreen Hibbert (2022). Supreme Court ruling on "skill and labour" and human authorship. 39
24	Jamaica Marrakesh Treaty	Amendments to Jamaica's Copyright Act for the visually impaired (Marrakesh Treaty) without broader AI exceptions. 40
25	Trinidad & Tobago Copyright	Copyright Act lack of TDM exception and Section 24 limitations. 41
26	Caribbean AI Policy	UNESCO/CARICOM AI Policy Roadmap emphasizing protection of IP rights. 42
27	CCJ Practice Direction	Practice Direction No. 1 of 2025 on the Use of Generative AI Tools in Court Proceedings (Feb 2025). 1010

 

Works cited

1. Copyright and Artificial Intelligence | U.S. Copyright Office, accessed December 5, 2025, https://www.copyright.gov/ai/
2. Copyright Office Issues Key Guidance on Fair Use in Generative AI Training - Wiley Rein, accessed December 5, 2025, https://www.wiley.law/alert-Copyright-Office-Issues-Key-Guidance-on-Fair-Use-in-Generative-AI-Training
3. Two California District Judges Rule That Using Books to Train AI is Fair Use, accessed December 5, 2025, https://www.whitecase.com/insight-alert/two-california-district-judges-rule-using-books-train-ai-fair-use
4. Fair Use and AI Training: Two Recent Decisions Highlight the Complexity of This Issue, accessed December 5, 2025, https://www.skadden.com/insights/publications/2025/07/fair-use-and-ai-training
5. GEMA, Getty and beyond: AI and copyright litigation in the EU and the UK - Taylor Wessing, accessed December 5, 2025, https://www.taylorwessing.com/en/interface/2025/online-and-ai-generated-content/ai-and-copyright-litigation-in-the-eu-and-the-uk
6. EXECUTIVE BRIEFING The Development of Generative Artificial Intelligence from a Copyright Perspective - EUIPO, accessed December 5, 2025, https://euipo.europa.eu/tunnel-web/secure/webdav/guest/document_library/observatory/documents/reports/2025_GenAI_from_copyright_perspective/2025_GenAI_from_copyright_perspective_executive_brief_en.pdf
7. New copyright law gives creators legal bite in protecting works - Barbados Today, accessed December 5, 2025, https://barbadostoday.bb/2025/10/22/new-copyright-law-gives-creators-legal-bite-in-protecting-works/
8. Copyright Bill hailed as milestone for creative future - Barbados Today, accessed December 5, 2025, https://barbadostoday.bb/2025/10/22/copyright-bill-hailed-as-milestone-for-creative-future/
9. Jamaica becomes regional leader with new rules for AI use in courts - CNW Network, accessed December 5, 2025, https://www.caribbeannationalweekly.com/news/jamaica-becomes-regional-leader-with-new-rules-for-ai-use-in-courts/
10. PRACTICE DIRECTION NO. 1 OF 2025 THE USE OF GENERATIVE ARTIFICIAL INTELLIGENCE TOOLS IN COURT PROCEEDINGS This Practice Directio - Caribbean Court of Justice, accessed December 5, 2025, https://ccj.org/wp-content/uploads/2025/02/PRACTICE-DIRECTION-NO.-1-OF-2025-THE-USE-OF-GENERATIVE-ARTIFICIAL-INTELLIGENCE-TOOLS.pdf
11. Copyright and Artificial Intelligence, Part 3: Generative AI Training Pre-Publication Version, accessed December 5, 2025, https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf
12. Bartz v. Anthropic - BakerHostetler, accessed December 5, 2025, https://www.bakerlaw.com/bartz-v-anthropic/
13. From Books to Bots: Key Takeaways from the Anthropic Fair Use Decision for AI Developers and Copyright Holders | Insights | Ropes & Gray LLP, accessed December 5, 2025, https://www.ropesgray.com/en/insights/alerts/2025/06/from-books-to-bots-key-takeaways-from-the-anthropic-fair-use-decision-for-ai-developers
14. Kadrey et al v. Meta Platforms, Inc., No. 3:2023cv03417 - Document 598 (N.D. Cal. 2025), accessed December 5, 2025, https://law.justia.com/cases/federal/district-courts/california/candce/3:2023cv03417/415175/598/
15. What Business Lawyers Can Learn from the First AI Copyright Fair Use Rulings, accessed December 5, 2025, https://www.americanbar.org/groups/business_law/resources/business-law-today/2025-september/ai-copyright-fair-use-rulings-business-lawyers/
16. OpenAI forced to release 20 million chat logs in NYT lawsuit - Windows Central, accessed December 5, 2025, https://www.windowscentral.com/artificial-intelligence/openai-chatgpt/judge-forces-openai-to-produce-20-million-chat-logs-in-copyright-lawsuit
17. The New York Times sues Perplexity AI for ‘unfair usage’ of its content: What the lawsuit says, accessed December 5, 2025, https://timesofindia.indiatimes.com/technology/tech-news/the-new-york-times-sues-perplexity-ai-for-unfair-usage-of-its-content-what-the-lawsuit-says/articleshow/125792040.cms
18. New York Times sues AI startup for 'illegal' copying of millions of articles - The Guardian, accessed December 5, 2025, https://www.theguardian.com/technology/2025/dec/05/new-york-times-perplexity-ai-lawsuit
19. Andersen v. Stability AI: The Landmark Case Unpacking the Copyright Risks of AI Image Generators - NYU Journal of Intellectual Property & Entertainment Law, accessed December 5, 2025, https://jipel.law.nyu.edu/andersen-v-stability-ai-the-landmark-case-unpacking-the-copyright-risks-of-ai-image-generators/
20. Top Takeaways from Order in the Andersen v. Stability AI Copyright Case, accessed December 5, 2025, https://copyrightalliance.org/andersen-v-stability-ai-copyright-case/
21. Implementation Timeline | EU Artificial Intelligence Act, accessed December 5, 2025, https://artificialintelligenceact.eu/implementation-timeline/
22. New EU Code of Practice for General-Purpose AI Models: What impact will it have on copyright? - Dentons, accessed December 5, 2025, https://www.dentons.com/ru/insights/articles/2025/september/19/new-eu-code-of-practice-for-general-purpose-ai-models
23. The New Copyright Directive: Text and Data Mining (Articles 3 and 4) - Wolters Kluwer, accessed December 5, 2025, https://legalblogs.wolterskluwer.com/copyright-blog/the-new-copyright-directive-text-and-data-mining-articles-3-and-4/
24. The General-Purpose AI Code of Practice | Shaping Europe's digital future, accessed December 5, 2025, https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai
25. EUIPO study on generative AI and copyright: Technical solutions for opt-out declarations and transparency measures - Taylor Wessing, accessed December 5, 2025, https://www.taylorwessing.com/en/insights-and-events/insights/2025/05/euipo-study-on-generative-ai-and-copyright
26. EUIPO study highlights legal gaps and policy needs in AI and copyright governance - CADE, accessed December 5, 2025, https://cadeproject.org/updates/euipo-study-highlights-legal-gaps-and-policy-needs-in-ai-and-copyright-governance/
27. text and data mining opt-out in Article 4(3) CDSMD: Adequate veto right for rightholders or a suffocating blanket for European artificial intelligence innovations? | Journal of Intellectual Property Law & Practice | Oxford Academic, accessed December 5, 2025, https://academic.oup.com/jiplp/article/19/5/453/7614898
28. Development of Generative Artificial Intelligence from a Copyright Perspective - EUIPO, accessed December 5, 2025, https://www.euipo.europa.eu/en/publications/genai-from-a-copyright-perspective-2025
29. Commission launches consultation on protocols for reserving rights from text and data mining under the AI Act and the GPAI Code of Practice | Shaping Europe's digital future, accessed December 5, 2025, https://digital-strategy.ec.europa.eu/en/consultations/commission-launches-consultation-protocols-reserving-rights-text-and-data-mining-under-ai-act-and
30. Senate passes new copyright law - Nation News, accessed December 5, 2025, https://nationnews-brb.newsmemory.com/?publink=2482e3d5b_134fbac
31. Agentic AI and the Law: Preparing Trinidad and Tobago for the Next Digital Leap, accessed December 5, 2025, https://polanc.com/agentic-ai-and-the-law-preparing-trinidad-and-tobago-for-the-next-digital-leap/
32. Meta Wins on Fair Use for Now, but Court Leaves Door Open for “Market Dilution”, accessed December 5, 2025, https://www.authorsalliance.org/2025/06/26/meta-wins-on-fair-use-for-now-but-court-leaves-door-open-for-market-dilution/
33. Copyright and AI: the Cases and the Consequences | Electronic Frontier Foundation, accessed December 5, 2025, https://www.eff.org/deeplinks/2025/02/copyright-and-ai-cases-and-consequences
34. European Commission receives final version of General-Purpose AI Code of Practice | IAPP, accessed December 5, 2025, https://iapp.org/news/a/european-commission-receives-final-version-of-general-purpose-ai-code-of-practice
35. Germany: landmark court decision deals with AI training and copyright | Global law firm, accessed December 5, 2025, https://www.nortonrosefulbright.com/en/knowledge/publications/218a3079/germany-landmark-court-decision-deals-with-ai-training-and-copyright
36. Copyright and AI in the UK: Opting-In or Opting-Out? | GRUR International | Oxford Academic, accessed December 5, 2025, https://academic.oup.com/grurint/article/74/11/1055/8209765
37. compliance of the fair use clause in the South African copyright amendment bill with the three-step test and the constitution of South Africa | Journal of Intellectual Property Law & Practice | Oxford Academic, accessed December 5, 2025, https://academic.oup.com/jiplp/article/20/7/447/8128929
38. Jamaica Copyright Act, ss. 52-55 (fair dealing) | H2O - Open Casebooks, accessed December 5, 2025, https://opencasebook.org/casebooks/405-copyrightx-jamaica/resources/9.3-jamaica-copyright-act-ss-52-55-fair-dealing/
39. AI: sufficiently recognized by the laws in the music industry? - Myers, Fletcher & Gordon, accessed December 5, 2025, https://myersfletcher.com/ai-sufficiently-recognized-by-the-laws-in-the-music-industry/
40. Copyright Act Amended to Facilitate Individuals Who are Blind or Visually Impaired - Ministry of Industry, Investment & Commerce, accessed December 5, 2025, https://www.miic.gov.jm/copyright-act-amended-facilitate-individuals-who-are-blind-or-visually-impaired/
41. Research Exceptions in Comparative Copyright Law, accessed December 5, 2025, https://digitalcommons.wcl.american.edu/cgi/viewcontent.cgi?article=1074&context=research
42. Caribbean Artificial Intelligence Policy Roadmap - UNESCO, accessed December 5, 2025, https://www.unesco.org/en/articles/caribbean-artificial-intelligence-policy-roadmap