TL;DR

  • For Singapore SMEs, AI-generated content is now a live IP litigation risk: training data is broadly defensible after Bartz v. Anthropic and Kadrey v. Meta (June 2025), but outputs that reproduce protectable expression, named characters, watermarks or celebrity likenesses are clearly exposed — and Anthropic's USD 1.5 billion settlement on 5 September 2025 shows what the downside looks like.
  • Vendor "copyright shields" from Microsoft, OpenAI, Google, Adobe, Anthropic and AWS are real but narrow: they are conditional on paid commercial tiers, on guardrails being kept on, and on the user not modifying outputs or knowingly infringing — and trademark and right-of-publicity claims are usually carved out.
  • The right place to put residual risk is a properly worded Media Liability or Multimedia Liability policy (with a Tech E&O or Professional Indemnity stack for SaaS firms and agencies), supplemented where available by an affirmative AI endorsement such as AXA XL's CyberRiskConnect Gen AI Endorsement; the Marsh Q1 2026 soft market gives Singapore SMEs leverage to negotiate this at renewal.

When OpenAI rolled out the GPT-4o native image generator on 25 March 2025, Singapore feeds filled overnight with "Ghibli-style" portraits — including one posted by Sam Altman as his own profile picture. Nine days later, on 3 April 2025, OpenAI COO Brad Lightcap posted on X: "very crazy first week for images in chatgpt – over 130M users have generated 700M+ (!) images since last tuesday." Studio Ghibli has not publicly commented. Hayao Miyazaki's 2016 line — "I am utterly disgusted… I strongly feel that this is an insult to life itself" — was recirculated millions of times. None of the heat translated into a Singapore court filing. But the underlying risk is now sitting on every SME's marketing rota: an image, a paragraph, a voice clip generated by AI that turns out to be a copy, a trademark, or a person's likeness — and a demand letter that arrives by email a week later.

This article is about that demand letter. It is article 415 in the COVA Emerging Risk / AI series. It covers AI-generated content copyright and IP infringement risk for Singapore SMEs, framed as a Media Liability, Intellectual Property Liability, Professional Indemnity (PI) and Technology Errors & Omissions (Tech E&O) insurance issue. The companion articles in this same-day cluster cover regulatory anchors (article 411), chatbot misrepresentation (412), deepfake funds-transfer fraud (414), AI bias in hiring (416) and AI-generated code security (417). All of them sit at /emerging-risk/ai/ on covarage.com.

Two threshold disclosures up front. COVA (Covarage Pte. Ltd., UEN 202531227H) is registered as an introducer under MAS Notice FAA-N02 — "Notice on Appointment and Use of Introducers by Financial Advisers" issued under section 58 of the Financial Advisers Act (Cap. 110). We do not advise on, recommend, rank or arrange insurance. We provide factual information sourced from primary regulators and route SMEs to licensed Independent Financial Advisers (IFAs) and brokers. Second, every regulatory claim, statutory section, court case and product detail below is linked to a primary source. Where Singapore-specific data does not yet exist, we say so and pivot to the closest comparable jurisdiction.

Key Findings

  • The AI training fair-use question is partly settled and trends pro-developer in the US (Bartz, Kadrey), but training on pirated source data is not protected and produced the largest copyright settlement in US history (Anthropic, USD 1.5 billion).
  • Outputs that reproduce identifiable expression — Disney/Universal/Warner characters, Getty watermarks, music lyrics — are not safe. Trade mark and right-of-publicity exposure is at least as serious as copyright exposure.
  • Singapore copyright law is friendlier to AI training than most jurisdictions because of the section 243–244 Computational Data Analysis exception, but it does not protect AI users from infringement claims based on outputs.
  • Vendor indemnities are useful first lines of defence but cap quickly, exclude trademarks, exclude free tiers, and require strict guardrail compliance.
  • Singapore-distributed Media Liability, IP, PI and Tech E&O capacity is sufficient for SMEs but rarely names "AI-generated content" affirmatively; the Marsh GIMI Q1 2026 soft market is the moment to negotiate clearer wording.

Details

What changed between 2023 and 2026

Three years ago the question of whether AI training and AI output could infringe copyright was an academic argument. As of May 2026 it is litigated, partly settled, and at least partly priced into insurance contracts. The empirical record now looks like this:

Anthropic paid USD 1.5 billion. In Bartz v. Anthropic PBC, No. 3:24-cv-05417 (N.D. Cal.), Judge William Alsup on 23 June 2025 ruled that training Claude on copyrighted books was "exceedingly transformative" fair use, but that Anthropic's downloading and retention of more than seven million pirated books from Library Genesis and Pirate Library Mirror was not fair use. Statutory damages of up to USD 150,000 per work pushed theoretical exposure above USD 70 billion. On 5 September 2025 the parties announced a USD 1.5 billion class settlement covering approximately 500,000 works at roughly USD 3,000 per book — the largest publicly reported copyright recovery in US history. Judge Alsup gave preliminary approval on 25 September 2025 and the final fairness hearing has been rescheduled to 14 May 2026 (Anthropic has paid the first USD 300 million tranche into escrow).

The visual artists' class action survived. In Andersen v. Stability AI Ltd., No. 3:23-cv-00201-WHO (N.D. Cal.), Judge William H. Orrick on 12 August 2024 denied most of Stability AI, Midjourney, Runway and DeviantArt's motions to dismiss the First Amended Complaint, allowing direct copyright, induced infringement, Lanham Act false-endorsement and trade-dress claims to proceed. Trial is scheduled for 8 September 2026. The case is currently in heavy discovery — the joint status report dated 16 October 2025 details the search-term protocol and custodian disputes still being negotiated.

Hollywood entered the fight. Disney Enterprises, Universal City Studios, DreamWorks, Marvel, Lucasfilm and Twentieth Century Fox sued Midjourney on 11 June 2025 in Disney Enterprises Inc. v. Midjourney Inc., No. 2:25-cv-05275 (C.D. Cal.), calling Midjourney a "bottomless pit of plagiarism" with revenue "more than USD 300 million in 2024." On 4 September 2025 Warner Bros. Discovery filed its own complaint in the Central District of California, attaching dozens of side-by-side comparisons of Midjourney outputs to copyrighted stills of Batman, Superman, Bugs Bunny, Scooby-Doo and Rick and Morty. Both cases are at the responsive-pleading stage.

Getty largely lost in London. In Getty Images (US) Inc. v. Stability AI Ltd. [2025] EWHC 2863 (Ch), handed down on 4 November 2025, Mrs Justice Joanna Smith DBE rejected almost all of Getty's claims after Getty abandoned its primary copyright infringement and database right claims mid-trial — there was no evidence the training had occurred in the UK. Stability won the secondary copyright infringement claim because the model weights of Stable Diffusion were not held to be "infringing copies." Getty did secure "extremely limited" trade mark infringement findings under sections 10(1) and 10(2) of the UK Trade Marks Act 1994 for early Stable Diffusion outputs that reproduced the Getty/iStock watermark. The 205-page judgment is, as Bird & Bird put it, "highly fact-specific" — and the territoriality lesson is the most important one for Singapore SMEs, who use models trained almost entirely outside Singapore.

The New York Times case is in late discovery. In New York Times Co. v. Microsoft Corp. & OpenAI, No. 1:23-cv-11195 (S.D.N.Y.), filed 27 December 2023, Judge Sidney Stein in March 2024 allowed the central copyright claims to proceed; the case has been consolidated with parallel actions from the Daily News, Center for Investigative Reporting and others. By late 2025 OpenAI was publicly fighting court orders to retain ChatGPT user logs (the company published a series of statements on its OpenAI v. NYT page in November 2025). No trial date is set as of May 2026.

On training data, two judges split. In Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417 (N.D. Cal.), Judge Vince Chhabria on 25 June 2025 — two days after Bartz — also held LLM training fair use, even on shadow-library sources, but reasoned that the plaintiffs had simply put up bad evidence on market harm. Two judges in the same building reached different views on whether piracy of source materials matters. Authors and music publishers are still pressing.

Music publishers extracted guardrails, not damages — yet. In Concord Music Group, Inc. v. Anthropic PBC, No. 5:24-cv-03811 (N.D. Cal.), Universal Music, Concord and ABKCO sued in 2023 over Claude reproducing copyrighted lyrics. Judge Eumi K. Lee's 30 December 2024 stipulation locked in technical guardrails preventing Claude from outputting lyrics. On 6 October 2025 the judge reinstated the publishers' contributory and vicarious infringement claims, ruling that Anthropic's own guardrails were evidence it knew about the infringement.

Outside the US: Hamburg and Delhi. In Robert Kneschke v. LAION e.V., the Hamburg Regional Court on 27 September 2024 (Case 310 O 227/23) and the Hanseatic Higher Regional Court on 10 December 2025 (5 U 104/24) held that LAION's downloading of a copyrighted photograph for inclusion in the LAION-5B training dataset (used by Stable Diffusion) was covered by the German text-and-data-mining exception for scientific research under section 60d of the Urheberrechtsgesetz, implementing Article 3 of EU Directive 2019/790. In India, ANI Media Pvt Ltd v. OpenAI OpCo LLC, CS(COMM) 1028/2024 in the Delhi High Court before Justice Amit Bansal, was argued through 2025 and judgment was reserved in early 2026.

The picture for SMEs is therefore not "fair use will save AI" or "AI is dead." It is: training is broadly defensible if you obtained the source data legally; outputs that reproduce identifiable expression are not safe; trademarks, watermarks and named characters are particularly exposed; and procedural posture and territoriality matter enormously.

Singapore copyright law applied to AI output

The Copyright Act 2021 came into force on 21 November 2021, replacing the Copyright Act 1987. Two features matter most for AI.

Section 243 — the Computational Data Analysis (CDA) exception. Section 243 defines computational data analysis to include "(a) using a computer program to identify, extract and analyse information or data from the work or recording; and (b) using the work or recording as an example of a type of information or data to improve the functioning of a computer program in relation to that type of information or data." An illustration in the statute itself names the use of images to train a computer program to recognise images. Section 244 makes the actual copying a permitted use, subject to five conditions: the copy is made for CDA purposes; it is not used for any other purpose; it is not supplied to anyone except for verification or collaborative research; the user has lawful access to the source; and where the source is from a "flagrantly infringing online location" the user must not have known that. Lawful access excludes circumvention of paywalls or breach of database terms of use. The CDA exception cannot be contracted out of, and it permits commercial use — making Singapore one of the most permissive AI training jurisdictions in the world. The Intellectual Property Office of Singapore (IPOS) confirmed the scope in its November 2022 Factsheet on the Copyright Act 2021. The Singapore Academy of Law published a 2025 reform paper on the CDA exception that is worth reading if you are a developer.

Authorship. The Copyright Act 2021 implies that an author of an authorial work must be a natural person. The IPOS-SMU CAIDG Joint Landscape Report, When Code Creates: A Landscape Report on Issues at the Intersection of Artificial Intelligence and Intellectual Property Law, published 28 February 2024, sets out the working position: in relation to authorial works, under Singapore law, an AI system cannot be named as an author because certain provisions of the Copyright Act 2021 imply that the author must be a natural person. For non-authorial works (sound recordings, films) the Act refers to a "maker," which can be a company, so copyright may still subsist. IPOS has also published a public-facing infographic — "How does Singapore law treat AI-generated content that may infringe copyright?" — that says directly: "As a user, developer or deployer of AI systems, you could be liable for copyright infringement by using AI irresponsibly. 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."

Fair use. Section 191 of the Copyright Act 2021 sets out the multifactor fair use test (purpose, nature, amount, market effect). It is similar in shape to 17 USC §107 but the Singapore courts have not applied it to a generative AI output dispute. Foundation cases on originality and authorship — Asia Pacific Publishing Pte Ltd v Pioneers & Leaders (Publishers) Pte Ltd [2011] SGCA 37 and Global Yellow Pages Ltd v Promedia Directories Pte Ltd [2017] SGCA 28 — would do the heavy lifting if such a case were filed.

Other rights. The Trade Marks Act 1998 covers AI outputs that reproduce or imitate registered marks (a recognisable car badge, a sports league logo, a watermark). The Registered Designs Act 2000 covers virtual designs after the 2017 amendments. Common law passing off covers unregistered trade dress and celebrity likeness misuse. The Patents Act 1994 — confirmed by IPOS in line with the UK and US in the DABUS line of cases — does not allow an AI to be named as inventor; IPOS's Supplemental Guidance for the Examination of AI-Related Patent Applications, released 11 October 2024, sets out how AI-related claims will be assessed.

Personal data. The Personal Data Protection Act 2012 (PDPA) applies to AI-generated content of named individuals. The PDPC's Advisory Guidelines on Use of Personal Data in AI Recommendation and Decision Systems, issued 1 March 2024, is the clearest official statement on AI and personal data in Singapore. There is no PDPC enforcement decision yet specifically about AI-generated deepfakes of named individuals; the deepfakes of senior political figures in 2023–2024 were addressed by takedowns and CSA advisories rather than PDPA enforcement. Parliament passed the Elections (Integrity of Online Advertising) (Amendment) Act in October 2024, prohibiting deepfake election content. The Criminal Law (Miscellaneous Amendments) Bill 2025, introduced 14 October 2025, proposes to expand the "intimate image" definition under Penal Code section 377BE to include AI-generated synthetic images.

There are no Singapore High Court or Court of Appeal decisions yet on AI-generated content copyright. The B2C2 Ltd v Quoine Pte Ltd [2019] 4 SLR 17 line on attribution of acts to autonomous algorithms is the closest analogue, and it is about trading bots, not generative AI.

The Studio Ghibli moment and why "style" is a trap

Style itself is not protectable under copyright in Singapore, the US, the UK or the EU — the idea/expression dichotomy means a visual aesthetic, a tone of voice or a colour palette cannot be owned. That is the standard answer. The problem is that "style" prompts often produce specific compositions, characters and trade dress that are protectable. On 26 March 2025 — the day after the GPT-4o image-generator launch — an OpenAI spokesperson told TechCrunch and NBC News: "We continue to prevent generations in the style of individual living artists, but we do permit broader studio styles—which people have used to generate and share some truly delightful and inspired original fan creations." That carve-out does not work cleanly in law: Studio Ghibli's identity is inseparable from Hayao Miyazaki's individual hand-drawn vocabulary, and a Ghibli-style poster pushed into commercial use can run into both passing off and trademark territory if it includes trade dress elements.

For a Singapore SME, the practical rules are:

  • Do not prompt by named living artist. The Andersen plaintiffs' theory in N.D. Cal. is that prompting "in the style of [artist name]" is itself probative of induced infringement.
  • Do not prompt by named copyrighted character. Disney v. Midjourney and Warner Bros. v. Midjourney are essentially "user types Yoda or Bugs Bunny, model returns Yoda or Bugs Bunny" cases.
  • Do not assume "studio style" prompts are safe. The Ghibli wave produced specific, recognisable composition elements that map back to specific films.
  • Watch for embedded watermarks. The one piece of Getty's case that survived was Stable Diffusion outputs reproducing the "GETTY IMAGES" and "ISTOCK" watermarks.

Will the AI vendor defend you? The reality check on indemnities

Vendor indemnities have been heavily marketed since September 2023. They are real, but they are narrower than the press releases suggest, and Singapore SMEs need to read them carefully.

Microsoft Customer Copyright Commitment (CCC). Announced 7 September 2023 by Microsoft Vice Chair Brad Smith, effective 1 October 2023. Microsoft will defend paid commercial Copilot and Azure OpenAI Service customers against third-party copyright claims arising from output content, and will pay any adverse judgments or settlements, but only if the customer (i) keeps the content filters and other safety systems enabled, (ii) does not attempt to generate infringing material, (iii) has rights in the input it used, and (iv) the claim is not about trademark or related rights. The CCC was extended to Copilot Studio on 1 June 2025. The required mitigations for Azure OpenAI are listed in the official Microsoft Learn page at "Customer Copyright Commitment Required Mitigations" — and as of 3 April 2026 the GitHub Copilot Duplicate Detection filter is no longer required for CCC coverage. The CCC does not extend to free products, custom-built Copilot services or consumer products.

OpenAI Copyright Shield. Announced by Sam Altman at OpenAI DevDay on 6 November 2023. Applies to ChatGPT Enterprise, ChatGPT Team and the OpenAI API; does not apply to free ChatGPT or ChatGPT Plus. The contractual indemnity is in OpenAI's Business Terms (section 10), and is subject to standard exclusions: the user "knew or should have known" the output was infringing, the user disabled filtering or safety features, the output was modified or used with non-OpenAI products, or the user did not have rights in the input.

Google Cloud / Gemini for Workspace. In October 2023 Google announced a "two-pronged" indemnity covering training data and generated output for Duet AI / Gemini for Workspace customers, with carve-outs for intentional infringement and for outputs the user knows or should know are infringing.

Adobe Firefly. Adobe's indemnity is the narrowest in scope but the most defensible: only outputs from "Indemnified Firefly Outputs" produced by Eligible Firefly Features, and only on qualifying enterprise plans. Adobe's "Adobe Generative AI Product Specific Terms" effective 17 June 2025 set the cap explicitly: "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." The indemnity covers copyright, trademark, publicity rights and privacy rights claims.

Anthropic, AWS, IBM watsonx.ai. Each of these has IP indemnity language for paid commercial customers, with the standard guardrail conditions and the standard carve-outs for user-supplied input and intentional misuse.

The reality check Singapore SMEs need to apply:

  1. Free tier = no indemnity. Almost every vendor indemnity is conditional on a paid commercial subscription.
  2. Liability caps are usually 12 months of fees paid. Adobe is the rare exception with an explicit per-claim figure; most enterprise software indemnities cap at fees paid in the prior 12 months. For an SME paying SGD 50 per seat per month, that is not large.
  3. Trademark and right-of-publicity carve-outs are common. Microsoft's CCC explicitly carves out trademark claims. Most of the Hollywood and music-publisher cases are partly trademark or right-of-publicity cases.
  4. Duty to defend vs duty to reimburse. Most vendor indemnities are duty-to-defend with vendor's choice of counsel. SME founders sometimes prefer to hire their own lawyers, which can mean reimbursement only.
  5. "Approved guardrails" preconditions. If you turned off a safety filter to get a sharper image, you have likely broken the indemnity. The Microsoft Learn page on required mitigations is updated regularly and you must follow it.
  6. Modified outputs. If your designer cleaned up an AI output in Photoshop, the indemnity may no longer apply. Adobe's terms tie the indemnity to the unaltered Firefly output.

In short: vendor indemnity is a useful first layer but it does not replace insurance.

How Singapore-distributed insurance actually responds

Five lines of cover are in play. The order matters.

Media Liability / Multimedia Liability Insurance. This is the line designed for the exposure. It covers content published on the company's own channels — copyright infringement, trademark infringement, defamation, invasion of privacy, right of publicity. In Singapore the most explicit named product is Beazley MediaTech, distributed via the myBeazley platform and the Lloyd's Asia syndicates 623 and 2623; Beazley's product page describes the cover as including "Defamation, invasion of privacy and plagiarism" and "Unfair competition alleged with copyright or trademark infringement." Chubb Singapore's PremierTech combines Tech Liability and cyber, and explicitly includes "Media Liability and Infringement of Intellectual Property protection for trademark, copyright and other intellectual property infringements." AIG Singapore's MultiMedia Professional Liability is offered as a separate proposal form. Hiscox writes media liability through its London market. Markel International Singapore Pte Ltd writes via Lloyd's Asia.

The unanswered question on Media Liability is whether AI-generated content is treated as "your content" for coverage purposes. Most current wordings do not exclude AI outputs, but most also do not affirmatively name them. Reading the wording with your broker is therefore essential.

Intellectual Property Liability Insurance. A standalone IP defence cover for an SME's products and content. In the Singapore market this is typically bundled inside Media Liability or Tech E&O wordings rather than sold as a separate primary product. CFC Underwriting and certain Lloyd's specialty syndicates sell standalone IP infringement defence; placement is via a Singapore-licensed broker into London or Lloyd's Asia.

Professional Indemnity (PI). For agencies, consultancies, content creators and marketing firms producing content for clients. PI responds to claims that the agency's professional services caused the client's loss — including IP infringement of the client's deliverables. AIG Singapore's Technology Professional Liability and QBE Singapore's Professional Indemnity are typical SG-issued products.

Tech Errors & Omissions (Tech E&O). For SaaS and tech firms whose product output may infringe — for example, an AI-generated content tool that ships infringing output to customers. Tokio Marine HCC's TechGuard combines Tech E&O and cyber. AXA XL's CyberRiskConnect APAC product, with the Generative AI Endorsement announced in a London/New York press release dated 21 October 2024, is the most explicit affirmative AI cover currently distributed into Singapore — it extends coverage to "data poisoning," regulatory violations, and "usage rights infringement," defined to include negligent use of copyrighted materials or software licences in machine-learning applications. Michael Colao, AXA XL's Global Chief Underwriting Officer, Cyber, said in the launch statement that "more companies are investing in developing their own Generative AI to unlock new possibilities. They must also be mindful of the risks that come with it, such as potential misuse, compliance issues, and the need for robust safeguards." The endorsement is available across the U.S., Canada, the U.K. and Lloyd's market, Europe and Asia per AXA XL's product sheet.

Cyber Insurance. Cyber policies generally do not cover IP infringement of content. Coalition's Affirmative AI Endorsement, launched 26 March 2024 in the US Surplus and Canadian markets, expands coverage for AI-driven security failures and AI-enabled funds-transfer fraud — but it is not a copyright cover, and Coalition does not currently distribute in Singapore. Coalition also added a Deepfake Response Endorsement globally in 2025. Munich Re's aiSure is performance-failure cover, not IP cover. Armilla Insurance Services (Lloyd's coverholder) launched an AI Liability Insurance policy with Chaucer in April 2025, and a follow-up combined cyber + AI liability product with Chaucer (USD 25 million AI aggregate, USD 10 million cyber) — but these are US-led with global territorial limits and require Singapore broker placement.

D&O. Directors' liability cover responds where there is a board-level governance failure on AI use — IP exposures uncovered after disclosure failures, for example. It does not cover the underlying IP claim itself.

The key 2026 market-context point: the Marsh Global Insurance Market Index for Q1 2026, published on 22 April 2026 by corporate.marsh.com, reports that global commercial insurance rates declined by 5% — the seventh consecutive quarter of decline — with financial and professional lines down 5% globally and Asia composite rates also down 5%; financial and professional lines specifically fell 7% in both Pacific and Asia. John Donnelly, President, Global Placement, Marsh Risk, said in the release that "the current competitive environment is expected to persist as insurer profitability remains strong." That gives Singapore SMEs negotiating leverage on coverage extensions and affirmative AI endorsements at renewal.

Concrete scenarios for Singapore SMEs

Scenario 1 — F&B chain, Midjourney campaign image. A Singapore restaurant group prompts Midjourney for a marketing image "in the style of a famous food photographer." The output is used on Instagram and the chain's homepage. The photographer's law firm sends a cease-and-desist citing copyright infringement and passing off. Likely insurance response: Media Liability picks up defence costs and any settlement, subject to the policy's IP infringement sub-limits and a "deliberate infringement" exclusion (which the photographer's lawyers will argue applies because the prompt named the artist). Vendor indemnity is unlikely to apply because Midjourney does not offer one and even paid Midjourney subscribers receive no IP indemnification under Midjourney's February 2026 Terms of Service.

Scenario 2 — SaaS company, ChatGPT-drafted help articles. A Singapore B2B SaaS firm uses ChatGPT (Enterprise tier) to draft product help articles. Several lengthy passages turn out to be near-verbatim copies of a competitor's documentation. The competitor sues for copyright infringement and passing off. Likely response: Tech E&O or PI picks up the third-party claim. OpenAI Copyright Shield may apply because the customer is on the Enterprise tier — but only if the customer did not modify the output and did not "know or should have known" it was infringing.

Scenario 3 — Digital agency, Adobe Firefly trademark mishap. A Singapore agency uses Adobe Firefly to generate stock-style imagery for a client. One image inadvertently contains a recognisable third-party trademark (a logo on a vehicle in the background). The trademark holder demands damages. Likely response: Adobe's Firefly Output Indemnity may apply, capped at USD 10,000 per claim per Adobe's 17 June 2025 Generative AI Product Specific Terms; agency PI picks up the balance. The Microsoft CCC and OpenAI Copyright Shield would not apply because the output came from Firefly.

Scenario 4 — D2C brand, ElevenLabs voice clone of a celebrity. A Singapore direct-to-consumer brand uses ElevenLabs to produce ad voiceovers that closely match a Singapore celebrity's voice without consent. The celebrity sues for passing off and PDPA breach (the voice is personal data). Likely response: Media Liability may respond on passing off; Cyber and PI may not. No vendor indemnity applies — ElevenLabs is a content-generation tool, not an indemnifying enterprise vendor for celebrity-likeness use.

Scenario 5 — Startup, Studio Ghibli-style product art. A Singapore startup uses an AI image generator to produce Ghibli-style product marketing. Studio Ghibli's IP rights group sends a takedown. Likely response: takedown compliance is the cheapest answer; Media Liability picks up any damages claim that follows. OpenAI's "broader studio styles" carve-out is a content policy, not a legal safe harbour.

Singapore Insurance Market Context

Singapore's Media Liability, IP, PI and Tech E&O capacity is concentrated in three places:

  1. MAS-licensed direct general insurers writing local policies — including AIG, Chubb, AXA XL Insurance Singapore, Tokio Marine, MSIG, Sompo, Liberty Specialty Markets, QBE, Allianz Commercial. The MAS Financial Institutions Directory at eservices.mas.gov.sg/fid lists 52 direct insurers (general) and 19 Lloyd's Asia Scheme entities as current figures.
  2. Lloyd's Asia Scheme service companies — Beazley Pte Ltd (Syndicate 623 and Syndicate 2623), Markel International Singapore, Catlin Singapore (Syndicate 2003 / AXA XL), Antares, Canopius, Allied World, Chaucer and others. The Lloyd's Singapore broker scheme was repealed on 1 November 2022; risks must be placed via a Lloyd's Asia service company or through an authorised Singapore intermediary.
  3. London market and overseas wholesale — placement via a Singapore-licensed broker for risks that local insurers will not write or for affirmative AI extensions not yet in local wordings.

Standard exclusions across Media Liability, IP, PI and Tech E&O wordings include deliberate infringement, prior knowledge, contractual liability assumed, fraud, criminal acts, and IP claims where the matter was disclosed before policy inception. Most policies are claims-made, meaning the policy that responds is the one in force when the claim is first made against the insured — not when the AI output was published. For a 2025-published Ghibli-style poster that gets sued in 2027, the 2027 policy responds, subject to retroactive date provisions.

The Marsh GIMI Q1 2026 numbers (Asia composite -5%; Asia financial and professional lines -7%) mean SMEs at renewal have leverage to ask for explicit AI affirmative coverage, broader retro dates, and removal of "silent AI" exclusions that some insurers have started inserting.

What this means for your business

Three operational shifts are now non-negotiable for Singapore SMEs using AI to generate content.

Inventory and document everything. Keep records of which AI tools your team uses; what they are used for (marketing copy, product imagery, customer comms, internal training, code); which outputs went where; the prompts used, the model and version, the date, and the human reviewer. This evidence trail is what insurers and defence lawyers need to defend a Bartz-pattern or Andersen-pattern claim.

Vendor-indemnity-aware procurement. When choosing between Midjourney (no indemnity), Adobe Firefly (USD 10,000 per claim cap, narrow scope), OpenAI ChatGPT Enterprise (Copyright Shield), Microsoft Copilot (Customer Copyright Commitment), or Google Gemini for Workspace, you are making a risk-allocation decision, not a feature decision. For commercial output that will be published or sold, the indemnified-by-vendor tools have real legal value over the un-indemnified ones.

Insurance gap audit. Look at your Media Liability, PI, Tech E&O, Cyber and D&O stack against your AI content exposure. Specific items to audit: Does the Media Liability wording exclude AI-generated content? Does the PI wording cover IP infringement claims by your client's customers (not just by your client)? Does the Tech E&O wording include affirmative AI language? Is there a sub-limit on IP claims, and how does it compare to potential US-style statutory damages? Does the Cyber policy carry a "silent AI" exclusion that would knock out AI-related defamation or copyright? What is the retroactive date on each claims-made policy?

A policy of "human review before publication" combined with "do not name living artists or copyrighted characters in prompts" eliminates a meaningful share of the litigation surface. A policy of "use Microsoft, OpenAI Enterprise, Google or Adobe paid tools only for client-facing content" further narrows it. Insurance is the third line, not the first.

Recommendations

Stage the work in three blocks. The benchmarks below are the triggers that should change your posture.

Within 30 days — hygiene.

  • Build an AI tools inventory (tool, vendor, plan tier, who uses it, what outputs, where outputs go).
  • Document a written rule against prompting by named living artist or named copyrighted character.
  • Establish a human-review checkpoint before any AI-generated content is published or shipped to a client.

Within 90 days — procurement and contracts.

  • Move client-facing creative work onto vendors with named indemnities (Microsoft CCC, OpenAI Copyright Shield, Google Gemini for Workspace, Adobe Firefly enterprise tier).
  • Read each vendor indemnity for the guardrail conditions, the cap (Adobe is USD 10,000/claim; most others cap at fees paid), the trademark and right-of-publicity carve-outs, and the input/modification carve-outs.
  • Update client master service agreements to push back IP indemnities and to limit the agency's liability for outputs the client directs.

At your next renewal — insurance.

  • Ask your broker or IFA to confirm in writing whether AI-generated content is covered under your existing Media Liability, PI and Tech E&O policies.
  • Request an affirmative AI endorsement (e.g., AXA XL CyberRiskConnect Generative AI endorsement) where capacity allows.
  • Use the Marsh GIMI Q1 2026 -5% Asia signal to ask for higher IP sub-limits, broader retro dates, and removal of any "silent AI" exclusions inserted in the last 12 months.

Triggers that should change your posture:

  • A US class settlement or judgment against an image-generation vendor on output (not training) — this would push insurers to insert AI-output exclusions and would compress the soft-market window.
  • A Singapore High Court decision applying section 191 fair use or section 243 CDA exception to a generative AI dispute.
  • Any change in vendor indemnity terms (caps, guardrail requirements, jurisdictional carve-outs) — these are unilaterally amendable and you need to be reading the change-log.
  • A claim or cease-and-desist landing in your inbox: notify your broker the same day. Most claims-made policies have strict notice clauses and a "demand letter" can constitute a claim.

Caveats

  • The Marsh GIMI numbers are averages across Marsh's client portfolio; individual SME experience will vary by sector, claims history, and broker.
  • Vendor indemnity terms are amended frequently. The figures in this article are accurate as of May 2026; review the live terms before relying on them.
  • No Singapore court has yet decided an AI-generated content copyright case. Where this article applies US, UK or German precedent to Singapore SMEs, the application is by analogy only.
  • The IPOS-SMU Landscape Report and IPOS infographics are guidance, not statute. The Copyright Act 2021 and the courts' interpretation of it remain the authoritative sources.
  • Coalition's Affirmative AI Endorsement, Munich Re aiSure and Armilla AI Liability Insurance are not currently distributed as Singapore-issued products and require placement via a Singapore-licensed broker into US, UK or Lloyd's markets.
  • COVA does not advise on, recommend, rank or arrange any insurance product. SMEs requiring advice on coverage selection should consult a MAS-licensed financial adviser or insurance broker.

Questions to Ask Your Adviser

  1. Does our current Media Liability or Multimedia Liability wording explicitly cover AI-generated content, or is the AI exposure left to be argued under "your content"?
  2. Does our Professional Indemnity policy respond to a third-party copyright or trademark claim arising from AI-generated deliverables we produce for our clients — and what is the IP sub-limit?
  3. Does our Tech E&O wording include an affirmative AI endorsement (for example along the lines of AXA XL's CyberRiskConnect Generative AI Endorsement covering "usage rights infringement"), and if not, is one available at our next renewal?
  4. What is the retroactive date on each of our claims-made policies, and does it cover AI-generated content we have already published?
  5. Have any of our policies been amended in the last 12 months to insert a "silent AI" exclusion, an "AI-generated content" exclusion, or an AI-related sub-limit — and if so, can it be removed or rewritten in the current soft market?
  6. How do the indemnities from our AI vendors (Microsoft CCC, OpenAI Copyright Shield, Adobe Firefly, Google) interact with our insurance — does our insurer subrogate against the vendor, or treat the vendor indemnity as primary?
  7. If we receive a cease-and-desist or DMCA-style takedown notice on AI-generated content, what is the policy notice trigger — is the takedown letter itself a "claim" or does the policy require a formal demand for damages?
  8. Given the Marsh Q1 2026 Asia financial-and-professional-lines decline of 7%, are there broader coverage extensions or higher limits available at our next renewal that we should be exploring?

[Match with a licensed IFA →]

Related Information

Published 8 May 2026. Source verified 8 May 2026. COVA is an introducer under MAS Notice FAA-N02. We do not recommend insurance products. We provide factual information sourced from primary regulators and route you to a licensed IFA who can match a policy to your specific situation.