On 26 August 2026 the Ministry of Law and the Intellectual Property Office of Singapore opened a public consultation on artificial intelligence and Singapore's intellectual property regime, and one of its copyright questions is how responsibility for infringement is shared between the people who build AI tools, the people who deploy them and the people who use them (MinLaw, Public Consultation on AI and Singapore's IP Regime). Nine days later, in New York, the parties in the newspaper case against OpenAI and Microsoft filed cross-motions for summary judgment (AI Lawsuit Tracker, New York Times v. OpenAI). If your business publishes images, copy or audio that an AI tool produced, the exposure sits with what you publish, and this is the season the rules around it are being written.

The Answer in 60 Seconds Singapore's Copyright Act 2021 lets a business copy works for computational data analysis under conditions in section 244, and a contract cannot remove that permission (Copyright Act 2021, section 244; section 187). That protects training and analysis. It does not protect an output that reproduces someone else's expression, character, mark or likeness. IPOS says so directly: "As a user, developer or deployer of AI systems, you could be liable for copyright infringement by using AI irresponsibly" (IPOS infographic).

Three documents decide what happens when a demand letter arrives: the AI vendor's terms, which indemnify only paid tiers used as the vendor specifies; your contract with the client or platform; and your insurance wording. Published Singapore wordings name copyright and trademark infringement under media liability, and one insurer's generative AI endorsement names "usage rights infringement" as a covered event (Chubb Singapore, PremierTech; AXA XL, 21 October 2024). Whether your own policy says either is a question for your wording, read with your broker.

The Sourced Detail

What Singapore law says today

The Copyright Act 2021 permits a copy made for computational data analysis when five conditions hold: the copy is made for that purpose, it is not used for any other purpose, it is not supplied to anyone except to verify results or for collaborative research, the person making it has lawful access to the source, and either the source copy is not an infringing copy, or the person did not know it was (and, where it came from a flagrantly infringing online location, could not reasonably have known it), or the use of infringing copies is necessary for a prescribed purpose (section 244). The statute's own illustrations say lawful access is not obtained by circumventing a paywall or breaching a database's terms of use. Section 187(1)(c) makes any contract term void to the extent it excludes that permitted use (section 187).

The exception covers the copying that analysis and training need. The output is a separate question. Section 191 sets the fair use factors - purpose and character, nature of the work, amount taken, and effect on the market (section 191) - and no Singapore court has applied them to a generative AI output. IPOS's guidance states the position in one sentence: "While the Singapore courts have yet to decide on an AI output infringement case, the Copyright Act 2021 (Act) and past court rulings on non-AI works already provide valuable guidance" (IPOS infographic).

On authorship, the IPOS-CAIDG landscape report When Code Creates records that an AI system cannot be named as the author of an authorial work, because provisions of the Copyright Act 2021 imply that an author is a natural person (IPOS-CAIDG report, 2024). For a business, that means an image produced with no human creative contribution gives you weak ownership of your own marketing and no reduction in your exposure if it copies someone else's.

The consultation that is open now

The MinLaw and IPOS consultation runs from 26 August 2026 to 22 October 2026, with feedback due by 5 PM on the closing day (MinLaw consultation page). The paper sets out three copyright questions: how to foster greater certainty and accountability in AI training; copyright risk management in AI deployment and use, including how responsibility is assessed among developers, deployers and end users; and the nature of human creativity in AI-assisted works (consultation paper, 36 pages). The second question is the one that reaches an SME that publishes AI output. A business with a view on where that responsibility sits has until 22 October to put it on the record.

Where the cases stand in September 2026

The cases that set the tone for every insurer's wording are in the United States and the United Kingdom, and six of them moved this year.

  • Bartz v. Anthropic. The USD 1.5 billion class settlement over pirated books received final approval on 20 July 2026 from Judge Araceli Martinez-Olguin, who took the case over on Judge Alsup's retirement; the payments run in instalments to 25 September 2027 (Authors Guild). The training side of that settlement is set out in our guide to the Bartz settlement.
  • New York Times v. Microsoft and OpenAI. Judge Sidney Stein on 4 April 2025 let the direct and contributory copyright claims proceed (S.D.N.Y. opinion); in January 2026 he affirmed an order compelling OpenAI to produce 20 million de-identified ChatGPT logs, and cross-motions for summary judgment were filed on 4 September 2026, with replies due on 6 November 2026 and no trial date set (AI Lawsuit Tracker).
  • Andersen v. Stability AI. The visual artists' case is set for a jury trial on 5 April 2027 (joint case management statement, January 2026).
  • Disney and Warner Bros. v. Midjourney. The studios' two cases, which place Midjourney outputs of named characters beside the originals, were consolidated for all purposes on 4 November 2025 (consolidation report; Georgetown Tech Institute on the complaint).
  • Getty Images v. Stability AI. The High Court's 205-page judgment of 4 November 2025 dismissed most of Getty's claims and found trade mark infringement only for early outputs reproducing the Getty and iStock watermarks, findings the judge called "both historic and extremely limited in scope" (judgment, para 757). Getty has permission to appeal the secondary infringement ruling (Taylor Wessing, January 2026).
  • ANI v. OpenAI. The Delhi High Court on 24 July 2026 refused the news agency an interim injunction, on a prima facie view that storage of its articles for training fell within the Indian fair dealing provision; the suit continues (Mondaq).

Read together, the training question is being settled case by case, and the output question - a named character, a watermark, a recognisable lyric - is the one no defendant has won outright.

What the vendor terms promise

Vendor indemnities are the first document to read, and each one has conditions.

  • Microsoft. The Customer Copyright Commitment, announced on 7 September 2023, defends paid commercial Copilot customers against copyright claims over output, on condition that "customers must use the content filters and other safety systems built into the product and must not attempt to generate infringing materials" (Microsoft, 7 September 2023).
  • OpenAI. The Copyright Shield covers ChatGPT Enterprise and API customers, "but not users of the free versions ChatGPT and ChatGPT+" (Proskauer).
  • Google. The generative AI indemnity for Google Cloud and Workspace "only applies if you didn't try to intentionally create or use generated output to infringe the rights of others" (Google Cloud).
  • Adobe. The Generative AI Product Specific Terms effective 23 April 2026 cap the Firefly indemnity: "our total maximum aggregate liability with respect to Indemnified Firefly Outputs will in no event exceed US$10,000 per (a) any Indemnified Firefly Output or (b) Infringement Claim", and exclude claims arising from "any modification of an Output" (Adobe terms, 23 April 2026).
  • Midjourney. The Terms of Service effective 27 May 2026 carry no indemnity for users; instead, to the extent permitted by law, the user indemnifies Midjourney against third-party claims arising out of or relating to the user's use of the Services and Assets (Midjourney Terms of Service).

The pattern holds across the four that offer an indemnity: none on a free tier, conditions on how the tool was used, and in Adobe's case a per-claim figure and an exclusion for edited output. Midjourney offers none. An image your designer cleaned up after generation sits outside the Adobe promise.

What the published insurance wordings say

Singapore has 52 direct general insurers listed in the MAS directory (MAS Financial Institutions Directory), and their published product pages show where copyright sits. Chubb Singapore's PremierTech page lists "Media Liability and Infringement of Intellectual Property protection for trademark, copyright and other intellectual property infringements" (Chubb Singapore). That is a media liability section inside a technology liability product: the class of cover the output risk falls under.

Two published endorsements name AI directly. AXA XL's Generative AI Endorsement to its CyberRiskConnect policy, announced on 21 October 2024 as available "throughout the U.S and Canada, U.K. and Lloyd's market, and Europe and Asia", covers "usage rights infringement", defined as "Negligently failing to obtain appropriate permissions to use particular items or data, such as copyrighted materials, software licenses, and intellectual property in machine learning" (AXA XL release). Coalition's Affirmative AI Endorsement of 26 March 2024 applies to its US Surplus and Canadian cyber policies and responds to AI-driven security failures and funds-transfer fraud, not to copyright (Coalition release).

The movement runs both ways. In the United States, ISO "introduced three AI-related endorsements, CG 40 47, CG 40 48 and CG 35 08 for optional use in commercial general liability policies" (Policyholder Pulse) - exclusions, not grants of cover. A wording that says nothing about AI today is a wording whose next renewal is worth reading line by line.

The market backdrop favours that reading. Marsh's index for the first quarter of 2026 reported global commercial rates down 5 percent, the seventh consecutive quarterly decline, with financial and professional lines down 7 percent in both Pacific and Asia (Marsh, 22 April 2026).

Likeness and personal data

A voice or face generated to resemble a real person brings in a second body of law. The PDPC's Advisory Guidelines on the use of personal data in AI recommendation and decision systems were issued on 1 March 2024 (PDPC guidelines). The Criminal Law (Miscellaneous Amendments) Act 2025 extended the Penal Code's "intimate image" definition to wholly AI-generated images, and that amendment came into force on 17 August 2026 (MHA, commencement). A marketing clip that imitates a known Singapore voice without consent can raise a passing off question and a personal data question at the same time.

Common Mistakes

  1. Treating the section 244 exception as a licence to publish. It permits copying for analysis (section 244). It says nothing about an output that reproduces protected expression, and IPOS says a user can be liable for that output.
  2. Prompting by a named living artist or a named character. The studio cases against Midjourney turn on outputs of named characters; a prompt that names the character is the first document the other side asks for.
  3. Relying on a free tier's output for client work. Every vendor promise above excludes the free tier. The indemnity you read about in a press release belongs to the paid plan your team is not on.
  4. Editing an AI image and assuming the vendor still stands behind it. Adobe's terms exclude claims arising from any modification of an output.
  5. Assuming the business owns what the tool made. Without a human author, copyright ownership of a purely AI-generated authorial work is weak, and a competitor who copies your campaign image has less to answer for.
  6. Checking for copyright and missing the trade mark. The only part of Getty's case that succeeded in London was trade mark infringement for outputs carrying its watermark (judgment, para 757). A logo on a van in the background of a generated street scene, a sports crest on a jersey, a stock-library watermark in a corner: each is a trade mark question before it is a copyright one, and each is visible to anyone who looks at the image.
  7. Reading the policy after the letter arrives. A demand letter or takedown notice is the moment the policy's notice clause starts to run. The wording, and what it counts as a claim, belongs in the file before any of this happens.

What This Means for Your Business

If your team uses AI tools to produce anything that leaves the building, the output is the exposure, and the evidence of how it was made is your defence.

For a business that publishes its own marketing. Keep a register of the tools in use, the plan tier each one runs on, and which outputs were published where. Record the prompt, the date and the person who approved the output. Remove prompts that name living artists or characters from your templates.

For an F&B, retail or events business. Menu photography, promotion posters and social posts are where generated images reach the public fastest. A campaign image made "in the style of" a known photographer or illustrator, or one that places a recognisable character on your poster, is the pattern the studio cases are built on. Keep the approved version of every published image with its prompt, and take down any image the moment a rights holder writes to you, while the question of who is right is still open.

For an agency or a content producer. Your client contract decides who carries an infringing deliverable. Read it next to the vendor terms: if the client directs the prompt and the vendor excludes edited output, the gap between the two documents is yours unless the contract moves it.

For a software company whose product generates content. Your customers' outputs become your claims. The Chubb page above shows media liability sitting inside a technology liability product; the question for your renewal is whether your own wording carries that section and at what sub-limit.

For every business on this list. Put your view into the MinLaw and IPOS consultation by 5 PM on 22 October 2026 if the allocation of responsibility between developer, deployer and user matters to you. It is the one document in this guide you can still change.

Questions to Ask Your Adviser

  1. Does my current wording include a media liability or intellectual property infringement section, and does it name AI-generated content, exclude it, or say nothing?
  2. What is the sub-limit for intellectual property claims, and how does it compare with the value of the work my business publishes?
  3. Is a generative AI endorsement available on my cyber, technology or media policy at renewal, and what events does it name?
  4. Has any AI exclusion been added to my policies at the last renewal, and if so, what does it remove?
  5. Does the policy treat a takedown notice or a cease-and-desist letter as a claim that must be notified, and within what period?
  6. If an AI vendor's indemnity also responds, which pays first, and does the insurer pursue the vendor afterwards?
  7. For client work, does my professional indemnity policy respond to an infringement claim brought by my client's customer as well as by my client?

Related Information

Published 8 May 2026. Source verified 13 September 2026.