The Answer in 60 Seconds

On 5 September 2025, Anthropic agreed to pay approximately USD 1.5 billion to settle the class action Bartz et al. v Anthropic PBC (N.D. Cal.) brought by authors alleging that Anthropic had copied their books without licence to train the Claude family of large language models. The fairness hearing was held on 14 May 2026, and the court granted final approval on 20 July 2026. The settlement is the largest publicly reported AI training-data settlement to date and crystallises a question that affects Singapore SMEs in two distinct positions: SMEs that produce content (text, images, code, data) that may have been used to train a model without licence, and SMEs that use generative AI tools whose model providers face training-data litigation. The Singapore copyright framework - the Copyright Act 2021 - contains two specific provisions relevant to AI training: the computational data analysis exception at sections 243-244, and the fair use framework at sections 190-191. This article sets out the international training-data landscape, the Singapore Copyright Act framework, the contractual position major AI vendors take with their customers, and the insurance covers that respond.

The Sourced Detail

The litigation around generative AI training data has intensified through 2024-2026 across multiple jurisdictions. The Bartz settlement is one of several headline matters - and it is meaningful for Singapore SMEs because the contractual and insurance positions that AI vendors take with their customers respond to the litigation pressure on the vendor side.

The two sides of the Singapore SME exposure are distinct: content producers whose work may have been ingested, and AI users whose tools depend on the model providers' training-data practices.

The Bartz settlement in context

In Andrea Bartz, Charles Graeber and Kirk Wallace Johnson v Anthropic PBC, filed in the Northern District of California, three authors brought a putative class action alleging that Anthropic had downloaded copies of their books from shadow-library sources to train the Claude models without licence. After class certification in 2025 and intensive motion practice, Anthropic agreed to settle for approximately USD 1.5 billion, with a fairness hearing held on 14 May 2026; the court granted final approval and entered judgment on 20 July 2026.

The settlement does not resolve the underlying legal question - whether large-scale training-data ingestion is fair use under US copyright law - but it does set the financial-impact reference point for the industry. Other matters proceeding in parallel include the New York Times' case against OpenAI and Microsoft, and various music publishers' cases against AI music-generation providers.

The Singapore relevance is two-fold. First, AI vendors face direct commercial pressure to license training data prospectively, reshaping the cost basis of foundation models. Second, AI vendors' contractual terms with customers are being revised to address training-data IP risk - both upstream (the licensing of training data) and downstream (the IP indemnity given to customers for outputs).

The Singapore Copyright Act 2021 framework

The Copyright Act 2021, in force from 21 November 2021, contains two provisions directly relevant to AI training in Singapore.

Computational data analysis (CDA) exception - sections 243 and 244. Section 243 of the Copyright Act defines computational data analysis, which includes using a work as an example of a type of information or data to improve the functioning of a computer program (the Act's illustration is using images to train a program to recognise images). Section 244 makes it a permitted use to copy a work for computational data analysis, or to prepare it for that analysis, if its conditions are met: the copy is used for no other purpose; it is not supplied to anyone except to verify the results or for collaborative research or study; the user has lawful access to the source; and the source is not an infringing copy, subject to limited exceptions. The section contains no non-commercial condition.

The CDA exception was deliberately drafted to facilitate text- and data-mining and machine-learning research in Singapore. Its application to commercial-scale AI training is the subject of evolving interpretation; the conditions (lawful access; non-supply of the work) are the contested edges.

Fair use - sections 190 and 191. Section 190 of the Copyright Act makes fair use a permitted use, and section 191 sets out the matters to be considered, including four factors: the purpose and character of the use (including whether it is of a commercial nature), the nature of the work, the amount and substantiality of the portion used, and the effect on the potential market for or value of the work. Sections 192 to 194 add rules for news reporting, criticism or review, and research or study.

Fair use is the more general framework; CDA is the specific. The interaction between them in the context of AI training is a live legal question.

The position for content-producer SMEs

Singapore SMEs that produce content - publishers, photographers, software developers, design studios, content agencies, data businesses - face the question of whether and on what terms their content may be used to train AI models.

Three operational positions:

Position 1: Explicit licensing. The SME enters into a licensing agreement with the AI vendor, defining the scope of training-data use, the duration, the consideration, and any output-attribution requirements. This is the position of the licensing deals concluded by major publishers with major AI vendors in 2024-2025.

Position 2: Robots-style opt-out. The SME's website, repositories or content stores publish machine-readable opt-out signals (robots.txt, ai.txt, terms-of-service prohibitions on AI training). The legal enforceability of such opt-outs against unauthorised scraping is the subject of ongoing dispute; the Singapore Copyright Act's CDA exception conditions intersect here.

Position 3: Detection and enforcement. The SME proactively monitors for use of its content in AI outputs and enforces through the standard copyright remedies. The detection technology is improving but is not yet mature for general use.

For SMEs in any of these positions, the question of whether the use of their content has been authorised is a question for case-by-case analysis under the Copyright Act framework. The general position cannot be assumed.

The position for AI-user SMEs

Singapore SMEs that use generative AI tools to produce content for the business (marketing copy, design assets, code, internal documents) face a different question: whether the AI's output infringes a third party's copyright, and what happens if it does.

The contractual position adopted by major AI vendors has evolved through 2024-2025 toward providing IP indemnities to enterprise customers, subject to conditions. The conditions typically include: using the AI's safety features; not modifying outputs in specified ways; not using the AI for prohibited use cases; and reporting any IP claim promptly. Consumer-tier and free-tier offerings typically do not carry the indemnity.

For Singapore SMEs, three operational positions:

Position 1: Enterprise contract with explicit IP indemnity. The vendor indemnifies the SME against IP claims arising from outputs, subject to the standard conditions. The IP indemnity caps and scope are the negotiated terms.

Position 2: Consumer/free tier without indemnity. The SME uses AI outputs without contractual IP cover. Any IP claim falls on the SME and, depending on the policy wording, on the SME's PI, Tech E&O or general liability insurance.

Position 3: Defensive verification. The SME uses similarity-detection tools and manual review to identify AI outputs that may infringe before publication. The defensive position reduces but does not eliminate the exposure.

How the Singapore insurance market responds

Three covers are operationally relevant.

Professional indemnity (PI). Where the SME provides professional services and uses AI in the deliverable, PI responds to third-party IP claims subject to policy wording. The same AI-exclusion check applies as for AI hallucinations in professional services.

Technology errors and omissions (Tech E&O). Where the SME develops or integrates AI tools as part of a technology offering, Tech E&O is the primary cover. The IP infringement extension is the relevant policy element.

Media liability / Intellectual property infringement cover. Where the SME's commercial activity is content-driven (publishing, advertising, marketing services), media liability or specific IP infringement cover is the primary route. The market for these covers in Singapore for SMEs is developing.

Cyber liability responds to an IP claim only where its wording includes it: one Singapore insurer's SME cyber cover names intellectual property rights under its media and social media liability, while a Singapore broker lists intellectual property infringement among cyber exclusions. Cyber may also engage if the SME's training-data use involves a data-protection failure.

The future direction

Three near-term developments will reshape the Singapore SME landscape.

Vendor IP indemnities will broaden as the major AI vendors compete on this dimension. Enterprise customers will increasingly receive better indemnification; the gap between consumer and enterprise tiers will widen.

Training-data licensing will scale, with publishers, photographers, and other content producers receiving direct payments from AI vendors. The market mechanics will resemble music licensing more than book-publishing royalty schemes.

Copyright Act jurisprudence will develop, with the first Singapore court decisions on the CDA exception's scope and on AI-output infringement expected through 2027-2028.

Common Mistakes / What Goes Wrong

  1. Using AI outputs without IP-indemnity tier. Consumer-tier convenience, enterprise-tier exposure.

  2. Assuming the AI vendor's training data is licensed. Most providers do not warrant this comprehensively.

  3. No defensive verification of AI outputs before publication.

  4. For content producers, no documented opt-out of AI training use.

  5. No PI or Tech E&O wording check for AI-output-IP exclusions.

  6. Misreading the Copyright Act CDA exception. The conditions are specific.

  7. Mixing licensed and unlicensed training data in proprietary AI fine-tuning.

  8. No registered IP for the SME's own content. Enforcement options are limited.

  9. Vendor contract liability cap too low to support a meaningful IP indemnity recovery.

  10. No insurance check on changes to AI vendor stack.

What This Means for Your Business

  1. If you produce content: document your IP position, publish machine-readable opt-out signals where applicable, consider licensing arrangements where there is commercial interest.

  2. If you use AI: confirm you are on an enterprise tier with an IP indemnity, run defensive verification, and check your PI / Tech E&O wording.

  3. For both: track Copyright Act developments and the major training-data litigation outcomes.

  4. Confirm insurance interaction with the licensed adviser before deploying any new AI tool or content-licensing arrangement.

  5. Maintain documented audit trails of AI tool usage, output sources, and verification steps.

  6. Review vendor terms annually as the indemnity landscape evolves.

  7. Consider media liability or IP infringement cover if content production is core to the business.

  8. Connect AI procurement to insurance audit (the 60-minute audit framework).

Questions to Ask Your Adviser

  1. For our PI / Tech E&O / media liability wordings, what is the position on AI-output IP infringement claims?
  2. If our AI tool's output is alleged to infringe third-party copyright, what is your support model at notification?
  3. Have any Singapore-market insurers introduced affirmative AI-IP endorsements that we should consider?
  4. For our content licensing position, what is the insurance interaction (representation and warranty cover, IP infringement cover)?
  5. As Copyright Act jurisprudence develops, what is your protocol for reviewing our cover's continued fit?

Related Information

Published 22 May 2026. Source verified 22 May 2026.