Totoro in Trouble? Assessing the Legal Implications of AI-Generated Ghibli Images

Since the launch of OpenAI’s GPT-4o image generation in March 2025, AI-generated images replicating Studio Ghibli’s signature hand-drawn aesthetic have flooded social media. Critics warn that generative AI could threaten artists’ livelihoods, while others see it as a potential collaborative tool. This article will first assess the nature of copyright infringement claims against generative AI platforms. It then adds to the legal discourse by assessing potential avenues of recourse beyond copyright infringement by reviewing these claims under Singapore’s legal regime. It concludes by discussing the regulatory initiatives aimed at addressing the challenges posed by generative AI.

Introduction

Unless you have been on a digital detox, you would have probably noticed the influx of Studio Ghibli-style images of your friends and family taking over social media. Since OpenAI’s launch of its improved GPT-4o on 25 March 2025, the internet has been flooded with AI-generated images mimicking the hand-drawn aesthetic of Ghibli classics like My Neighbor Totoro and Spirited Away. While some critics have argued that AI-image generators could put artists out of work, others have posited that AI-image generators could be used as a collaborative tool to assist artists and creators.¹See the Straits Times article with views from artists against the use of generative AI, accessible at: <https://www.straitstimes.com/life/artists-fight-ai-programs-that-copy-their-styles>See the article in the Harvard Gazette for differing views from artists and creatives supporting the incorporation of generative AI in their creative process, accessible at <https://news.harvard.edu/gazette/story/2023/08/is-art-generated-by-artificial-intelligence-real-art/> Hayao Miyazaki, the co-founder of Studio Ghibli, has previously opposed incorporating AI in his creative process, adding to the broader controversy concerning AI-generated content.²See Miyazaki’s translated comments in the video from 1:14 – 1:27, where he states “I am utterly disgusted… I would never wish to incorporate this technology into my work at all. I strongly feel that this is an insult to life itself.”, accessible at <https://www.youtube.com/watch?v=ngZ0K3lWKRc>

To date, several lawsuits have been filed outside of Singapore against generative AI platforms.³See, for example, the recently reported claim for copyright infringement filed by Ziff Davis against Open AI in the Delaware federal court, accessible at <https://www.channelnewsasia.com/business/publisher-ziff-davis-sues-openai-copyright-infringement-5087931>Also see the complaint filed by The New York Times Company on 27 December 2023 for copyright infringement against OpenAI and Microsoft for their use of its articles to train their LLMs in The New York Times Company v OpenAI and Microsoft, accessible at <https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec2023.pdf>Also see the complaint filed by several publishers on 18 October 2023 for copyright infringement against Anthropic for its use of copyrighted musical compositions, including song lyrics, to train its AI models accessible at <https://www.musicbusinessworldwide.com/files/2023/10/UMG-lawsuit.pdf>Also see the amended complaint filed in the class action suit on 31 October 2024 for copyright infringement against Stability AI, Runway AI, Midjourney and DeviantArt for their use of copyrighted works to train their various machine learning models in Sarah Andersen and others v Stability AI Ltd and others, accessible at <https://www.courtlistener.com/docket/66732129/238/andersen-v-stability-ai-ltd/> As Singapore continues to establish its position as a flourishing technology hub with a strong commitment to supporting AI innovation,⁴See page 61 of the NAIS 2.0 where it reads “Singapore’s commitment to developing and deploying AI well, and we will leverage these gains to be a global pace-setter at the forefront of AI.”, accessible at <https://www.edb.gov.sg/content/dam/edb-en/business-insights/market-and-industry-reports/singapores-national-ai-strategy-ai-for-the-public-good-for-singapore-and-the-world/nais2023.pdf> understanding how Singapore’s legal regime would react to generative AI is crucial for businesses operating in the region.

This article will first address the nature of copyright infringement claims against generative AI platforms, drawing from recent judgments in the legal landscape. Second, it will venture beyond normative considerations of copyright infringement to assess other avenues of recourse against generative AI platforms. Third, it will consider the causes of action against generative AI platforms under Singapore’s legal regime. Last, it briefly discusses the regulatory initiatives aimed at addressing the novel challenges arising from the proliferation of generative AI.

Copyright Infringement Against Generative AI Platforms

The arguments for copyright infringement can be broadly categorised into two categories. The first category concerns the input data used to train these platforms (Input Argument). Most generative AI platforms rely on Large Language Models (LLMs) to generate outputs.⁵See for example OpenAI’s discussion of the reasoning models for GPT o3 and o4-mini, clarifying that both models are LLMs, accessible at <https://platform.openai.com/docs/guides/reasoning?api-mode=responses> LLMs are trained with large volumes of text and data. ⁶See more detailed discussion of how LLMs work in the Google article discussing generative AI, accessible at <https://blog.google/inside-google/googlers/ask-a-techspert/what-is-generative-ai/> The Input Argument asserts that relying on web-scraping to collect data that is used to train LLMs constitutes copyright infringement, as it often involves copying copyrighted works without obtaining prior consent from rights holders.⁷See, for example, Getty Images (US) Inc and others v Stability AI Ltd (2025) EWHC 38 (Ch) at (7) – (9) The second category relates to the outputs generated by LLMs (Output Argument). The Output Argument posits that LLMs are reproducing or copying a substantial portion of copyrighted works in the generated outputs.⁸ibid These arguments have also been raised simultaneously in proceedings.⁹ibid

Recently, the Hangzhou Internet Court (HIC) found a Chinese generative AI platform provider liable for contributory copyright infringement of information network dissemination rights by facilitating the generation of images substantially similar to the copyrighted Ultraman character for commercial purposes.¹⁰See the opinion of the Hangzhou Intermediate Court dated 10 February 2025, accessible at https://mp.weixin.qq.com/s?__biz=MzU4NzExNTkyMQ==&mid=2247507667&idx=1&sn=c524cc81dff2bf48a3469f94173fa8b7&scene=25&ascene=60&devicetype=android-35&version=28003842&nettype=h2g2&abtest_cookie=AAACAA%3D%3D&lang=en&session_us=gh_512c7f738a36&countrycode=US&exportkey=n_ChQIAhIQVqEbFU3Qel6lUFUfvTmSHRLkAQIE97dBBAEAAAAAAPUIC7ifeLMAAAAOpnltbLcz9gKNyK89dVj0szNvVpdxYa4yL2j%2Bxzxnkln8fScT3fvqoxf3bPpwPzw79qGNAcaxztKTEupPeUjnFPKU0hzhoaeIPZQ5EmzgJdwRMZ1qxTsHEjMsTZoxvBB4iy9og0l30J%2FMiYclNo6MuY1bCfScbfphulkH3Mlj0U5ZCfe6jBivxtXkRyjYRWUeEKeBH36zntaymdQYBvjtvy34gSS0T5bqHhLKJYdQrzOvm0IN7BNAw7M%2FD6CK3jDZ4TR2apnMeb1e7GJkYA%3D%3D&pass_ticket=PWGM6uRv2tJ%2BIe1AeeXpOZ1uQGTVwnfUM7JiPfCXArERXSKAJTpoxrMP9BXrRz9V&wx_header=3 The infringed information network dissemination rights referred to in this case can be likened to a rights holder’s exclusive right to communicate a work to the public in Singapore.¹¹See Article 10(12) of the Copyright Law of the People’s Republic of China, translated copy accessible at <https://wilmap.stanford.edu/node/31101>Also see Chen Qinjie, “Case Analysis and Reflections on the Determination of Infringement of information Network Dissemination Rights”, accessible at <https://ijemh.com/issue_dcp/Case%20Analysis%20and%20Reflections%20on%20the%20Determination%20of%20Infringement%20of%20Information%20Network%20Dissemination%20Rights.pdf> and Article 8 of the WIPO Copyright Treaty, as well as sections 112(1)(d) and 113(c) read with sections 146(1) and 150 of the Copyright Act 2021 In this case, the claimant was an exclusive licensee of the Ultraman Intellectual Property, while the defendant was an unnamed generative AI platform provider offering text-to-image and image-to-image generative AI services.¹²See the opinion of the Hangzhou Intermediate Court dated 10 February 2025 referred to at endnote 10 Users of the defendant’s platform created customised Ultraman AI models, by first uploading copyrighted Ultraman images to the platform and choosing the base model and processing parameters to influence how the input image would be interpreted and transformed.¹³ibid Other users could then use these uploaded images to generate Ultraman images through prompts to produce their intended output.¹⁴ibid

The HIC considered that while an AI platform can in theory be liable for direct copyright infringement if it directly engages in acts that are reserved for the copyright owner, the platform in this instance could not be liable for direct copyright infringement as the users were providing the training data to the LLMs. As for the contributory copyright infringement claim, the HIC considered various factors, including but not limited to the nature and business model of the defendant’s generative AI platform, the popularity of the Ultraman works, as well as its visibility on the defendant’s platform, and whether the defendant had adopted any effective preventive measures.¹⁵ibid Notably, the HIC found that the defendant had promoted its generative AI model by reproducing the Ultraman works on its website, and such promotion facilitated the users’ direct infringement.¹⁶ibid Further, the HIC also considered that the defendant was directly profiting from the infringement as it was charging a fee to users of its platform and had failed to take proactive measures to manage and filter infringing content.¹⁷ibid

Across the Pacific, the US District Court for the District of Delaware held by way of summary judgment that a legal research search engine, ROSS, was liable for direct copyright infingement for its copying of 2,243 headnotes found on Westlaw’s database that was used to train its own AI tool.¹⁸Thomson Reuters Enterprise Centre GmbH v Ross Intelligence Inc. No. 1:20-cv-613-SB, (D. Del. Feb 11, 2025) at page 2, accessible at: <https://storage.courtlistener.com/recap/gov.uscourts.ded.72109/gov.uscourts.ded.72109.770.0.pdf> The headnotes contained a summary of key points of law and case holdings curated by Westlaw.¹⁹ibid at page 7 Notably, this case does not involve a generative AI platform like ChatGPT, as the ROSS platform would reproduce extracts of judicial opinions that have already been written.²⁰ibid at page 17: “…when a user enters a legal question, Ross spits back relevant judicial opinions that have already been written.” AI platforms like ROSS can be distinguished from generative AI platforms, as in the latter, the generated output is not retrieved or copied from a database. Rather, the output is a product of the LLM’s inference on the most appropriate response to a prompt, which arguably creates new content.²¹See, for example, the explanation of how ChatGPT works, accessible at <https://help.openai.com/en/articles/7842364-how-chatgpt-and-our-foundation-models-are-developed> This distinction between AI platforms and generative AI platforms has a direct impact on the strength of any Output Argument. With generative AI platforms, the LLMs are not reproducing training data, but instead analysing the data to generate a new output. Proving a substantial part of a specific work has been copied may be tenuous depending on the LLMs functionality, and it may be more practical to rely on the Input Argument.

Meanwhile, the US District Court for the Northern District of California recently refused to grant a preliminary injunction sought by a group of music publishers against Anthropic, a technology company that develops general-purpose LLMs.²²Concord Music Group and others v Anthropic PBC No 24-cv-03811-EKL (N.D. Cal. Mar 25, 2025) at pages 1 – 3, accessible at <https://www.musicbusinessworldwide.com/files/2025/03/show_temp-3.pdf> The publishers argued that Anthropic had used copyrighted song lyrics to train its LLMs.²³ibid Refusing to grant the interim injunction, the Court highlighted that the publishers had failed to establish that damages (i.e., monetary compensation) would be an insufficient remedy for any purported loss.²⁴ibid at page 11 Notably, the Court highlighted that the publishers had failed to show how Anthropic’s use of the works would negatively impact the market for AI training licenses as they had not submitted any evidence showing that their abilities to negotiate such licenses with AI developers was affected.²⁵ibid at page 11

Circling back to the viral Ghibli trend, most users are able to generate their “Ghiblified” images by uploading their own images with the use of one-line prompts on the GPT-4o platform. While users are generally not uploading copyrighted images like in the case before the HIC, it could be argued that the input data used to train GPT-4o includes copyrighted Ghibli images as the generated outputs adopt the Ghibli style effortlessly. With Getty Images v Stability AI Ltd set for trial in June, the UK courts will have the opportunity to address both the Input Argument and Output Argument; namely, whether the downloading of copyrighted works on servers during the development and training stage of Stable Diffusion, and whether the resulting outputs of synthetic images constitute copyright infringement.²⁶See Getty Images (US) Inc and others v Stability AI Ltd (2025) EWHC 38 (Ch) at (9) and (123) While much of the discourse on generative AI has centred on copyright infringement, potential claimants may pursue alternative avenues to seek recourse against these platforms.

Alternative Causes of Action Against Generative AI Platforms

Training LLMs with data compiled through web-scraping may breach the terms and conditions of various websites. Web-scraping refers to the use of software to automate the search and extraction of data from various websites.²⁷See the Department of Statistics Singapore’s commentary on Web-scraping Principles, accessible at <https://www.singstat.gov.sg/find-data/concepts-methods-and-applications/stat-resource/web-scraping-principles> While some website terms and conditions expressly prohibit web-scraping (see for example RyanAir’s Terms of Use), the prohibition of web-scraping could also be more subtly expressed. To illustrate, the Ghibli website (see screenshot below) expressly states that the use of its artistic works for commercial purposes, including any other use that the rights holder deems inappropriate, is prohibited. It is arguable that the use of Ghibli images for commercial purposes (e.g. to train for-profit LLMs like GPT-4o) already breaches the terms of use of the Ghibli website and there would be no need to rely on the broader prohibition of ”any other use” in the terms of use. Separately, the ability to rely on a breach of contract claim may be hindered when the developers of the LLMs are not directly engaging in web-scraping, but are instead relying on an intermediary web-scraping service provider to collect vast amounts of data to train their LLMs. In such an instance, there would be a need to identify the proper parties to any potential claim.

Source: https://www.ghibli.jp/works/chihiro/

In the US, social media giants have tried to bring claims for breaches of their website terms and conditions against companies that engage in web-scraping.²⁸See for example X Corp v Bright Data Ltd, No. C 23-03698 WHA (N.D. Cal. May 9, 2024) (“X Corp Judgment”), accessible at <https://storage.courtlistener.com/recap/gov.uscourts.cand.415869/gov.uscourts.cand.415869.83.0.pdf> as well as Meta Platforms, Inc. v Bright Data Ltd., No. 23-cv-00077-EMC (N.D. Cal. Jan 23, 2024) (“Meta Judgment”), accessible at <https://www.courthousenews.com/wp-content/uploads/2024/01/meta-platforms-v-bright-data-ruling-motion-for-summary-judgment.pdf> For instance, Meta and X alleged that Bright Data, a data-scraping company, developed and used unauthorised automation software to scrape data from their respective social media platforms which it then advertised and sold.²⁹See the X Corp Judgment at page 5 – 6 from line 14 at page 5 to line 5 on page 6 and the Meta Judgment at page 4 lines 19 – 28 The US courts have generally dismissed these suits despite the defendants profiting from the sale of scraped data, when the data scraped involves publicly available data accessible to anyone without the need to log in to a social media account.³⁰See the Meta Judgment at page 25 lines 6 – 11 This position was also taken by the US Court of Appeals for the Ninth Circuit in hiQ Labs, Inc. v LinkedIn Corporation, where the Court highlighted that giving social media companies free rein to decide who can collect and use publicly available data risks the possible creation of information monopolies.³¹hiQ Labs, Inc. v LinkedIn Corporation No. 17-16783 (9th Cir. Apr. 18, 2022) at page 43, accessible at <https://cdn.ca9.uscourts.gov/datastore/opinions/2022/04/18/17-16783.pdf> In Computerlife Software Inc v Newman, the US Court of Appeals for the Eleventh Circuit held that while web-scraping can be perfectly legitimate, the extent and scale of web-scraping in certain instances of publicly available data may amount to a misappropriation of a trade secret.³²Compulife Software Inc v Newman, No. 21-14074 (11th Cir. Aug. 1, 2024) at page 23, accessible at <https://media.ca11.uscourts.gov/opinions/pub/files/202114071.pdf>

While there are no reported cases in Singapore addressing whether the practice of web-scraping would constitute a breach of confidence, the case of PropertyGuru Pte Ltd v 99 Pte Ltd [2018] SGHC 52 briefly discusses web-scraping. In this case, 99 Pte Ltd (99) had admitted to scraping property listings from PropertyGuru’s website, which were posted on 99’s own website.³³PropertyGuru Pte Ltd v 99 Pte Ltd (2018) SGHC 52 at (9) The parties entered into a settlement agreement thereafter.³⁴PropertyGuru Pte Ltd v 99 Pte Ltd (2018) SGHC 52 at (13) However, PropertyGuru claimed that 99 breached the terms of the settlement agreement, induced property agents to breach the terms and conditions of PropertyGuru’s website, as well as infringed the copyright in the photographs posted on the PropertyGuru website, which were reposted on 99’s website.³⁵ibid at (3) and (90) In respect of the breach of contract claim, the Court found that 99 had breached a clause in the settlement agreement by facilitating the cross-posting of nine photographs.³⁶ibid at (36) and (60) As for the inducement of breach of contract claim, PropertyGuru had failed to properly plead the causal link between the acts of inducement by 99 and the breaches of the terms and conditions by the property agents.³⁷ibid at (82) and (88) As the breach of contract claim related to a breach of the settlement agreement as opposed to the breach of the terms and conditions of PropertyGuru’s website, the judgment is not directly relevant. That said, the availability of alternative legal avenues against generative AI platforms warrants closer examination within the Singapore legal framework.

Assessing the Causes of Action Against Generative AI Platforms in Singapore

The author of an artistic work generally has the exclusive right to make a copy, publish or communicate the work to the public.³⁸Section 113 of the Copyright Act 2021 This exclusive right is subject to statutory permitted uses, fair use provisions and contractual agreements between parties. In particular, section 244 of the Copyright Act 2021 (CA) limits this exclusive right by allowing the copying of a copyrighted work for the purpose of computational data analysis (CDA). CDA refers to using works for training AI programmes for both commercial and non-commercial use.³⁹See the Second Reading Speech by Second Minister for Law, Mr Edwin Tong SC, on the Copyright Bill on 13 September 2021 at (55) – (57) , accessible at <https://www.mlaw.gov.sg/news/parliamentary-speeches/2021-09-13-second-reading-speech-by-second-minister-for-law-edwin-tong-on-copyright-bill/> as well as the illustration in section 243 of the Copyright Act 2021 Copyright holders are not entitled to exclude or restrict the application of this CDA carve out by entering into contractual agreements.⁴⁰See sections 187, 186 and 188 of the Copyright Act 2021 The CDA carve out applies when data is obtained through lawful means, such as when a work is already available in the public domain, and accessed without circumventing a pay wall.⁴¹See the illustrations in section 244(2)(d) of the Copyright Act 2021

While ChatGPT’s use of scraped data to train its models has been heavily critiqued, such use seems to fall within the permitted use exceptions under Singapore’s copyright regime and any Input Argument raised involving lawfully obtained copyrighted material to train LLMs would prima facie be defeated.⁴²See the critique from a former employee of OpenAI, accessible at <https://suchir.net/fair_use.html> Further, the rights holder would not be able to bring a claim for breach of contract by way of web-scraping given that section 187(1)(c) of the CA renders any clause in a contract that has the effect of limiting or restricting the CDA carve out void. This section applies retrospectively to contracts entered into before the CA came into force.⁴³Section 187(3) of the Copyright Act However, if it can be proven that the data used to train these LLMs was extracted by breaching a pay wall, or perhaps by way of unauthorised access of a rights holder’s servers, the copyright holder may still have a claim for copyright infringement.

As for the Output Argument, there would be a need to show that the generated output copied a substantial part of the artistic work. In Singapore, it has been accepted that copyright protects artistic expression, not facts or ideas.⁴⁴Global Yellow Pages Ltd v Promedia Directories Pte Ltd and another matter (2017) SGCA 28 at (15) This focus on artistic expression limits the scope for protecting an artist’s style.⁴⁵See for example Norowzian v Arks Ltd (No. 2) (2000) EMLR 67 at (228) that concerned a dramatic work and whether the alleged similarity between the filming and editing styles in two films could amount to copyright infringement. The UK Court of Appeal held that copyright could not subsist in mere style or technique; confer this position with the comments in Sobel, “Elements of Style: Copyright Similarity and Generative AI”, Harvard Journal of Law & Technology Vol 38 Number 1 Fall 2024 at (74), accessible at <Click Here> To determine what constitutes copying of a substantial part of a work is contingent on the facts of the case, but a focus is placed on the quality of the copying as opposed to the quantity.⁴⁶Flamelite (S) Pte Ltd and others v Lam Heng Chung and others (2001) 3 SLR(R) 610 at (35) The concept that copyright should not be extended to protect an artist’s style has been endorsed in other Commonwealth jurisdictions even though it has not been directly addressed by the Singapore courts.⁴⁷See for example IPC Media v Highbury (2004) EWHC 2985 (Ch) at (14) and Cummins v Vella (2002) FCAFC 218 at (41) – (44) In essence, it appears that unless it can be proven that the data used to train LLMs was obtained through unlawful means, the avenues for bringing a claim against developers of LLMs are limited in Singapore.

The Secret World of Gen AI: Future Considerations

As generative AI continues to evolve at a staggering pace, regulatory responses have begun to reflect the complexity of governing such transformative technology. Singapore’s Model AI Governance Framework for Generative AI (Gen AI Framework) acknowledges the tension between balancing the interests of copyright holders and ensuring data accessibility to facilitate innovation. The Gen AI Framework proposes facilitating access to quality data used to train LLMs, through the use of data analysis tools to debias and remove inappropriate content.⁴⁸See the Model AI Governance Framework for Generative AI published on 30 May 2024 at page 11, accessible at <https://aiverifyfoundation.sg/wp-content/uploads/2024/05/Model-AI-Governance-Framework-for-Generative-AI-May-2024-1-1.pdf> For example, the team that collated LAION-5B, an open-source dataset containing over 5 billion image-text pairs, has developed a watermark detection model. This model allows developers to filter watermarked images from the dataset used to train LLMs. However, LAION’s watermark detection model does not address the concern of web-scraping copyrighted content at the outset and is not always effective in filtering watermarked images.⁴⁹See comments on watermark and safety inference, accessible at <https://laion.ai/blog/laion-5b/> Further, copyrighted content is not always watermarked. Common Crawl, one of the largest publicly available repositories of web data, allows website owners to opt out from their Common Crawl bot collecting their data by adding a special instruction called a robots.txt file to their website.⁵⁰See the FAQs on Common Crawl, accessible at <https://commoncrawl.org/faq> Having opt-out mechanisms in place provides copyright holders with some comfort, while ensuring access to sufficient data to enhance innovation.

Notably, the UK has also taken steps to develop common principles for guiding the use of generative AI in government organisations.⁵¹Artificial Intelligence Playbook for the UK Government, at page 9 accessible at: <https://assets.publishing.service.gov.uk/media/67aca2f7e400ae62338324bd/AI_Playbook_for_the_UK_Government__12_02_.pdf> Conversely, the EU has departed from relying on soft law mechanisms, embracing a more proactive approach towards regulating generative AI. The European Union’s Regulation 2024/1689, commonly referred to as the Artificial Intelligence Act (AI Act) is the first attempt worldwide to harmonise the regulation of AI.⁵²See the overview of the AI Act, accessible at <https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai> The AI Act adopts a risk-based approach towards regulating AI, including generative AI, which is referred to as general-purpose AI models (GPAIMs) in the AI Act.⁵³See Recitals 26 – 27 of the AI Act The AI Act requires providers of GPAIMs (i.e., companies that develop such models with a view to placing them on the EU market for payment of a fee or for free) to comply with several obligations.⁵⁴See Articles 51 – 55 of the AI Act All providers of GPAIMs, including GPAIMs that are released under a free and open-source licence, are required to prepare and publicly disclose a summary of the content used for training GPAIMs, as well as have policies in place to ensure compliance with EU copyright and related laws.⁵⁵See Article 53(1)(c) and 53(1)(d) of the AI Act Providers of GPAIMs with systemic risk (e.g., GPAIMs that have high impact capabilities due to the number of registered end-users) have additional obligations, which include but are not limited to structured reporting obligations to the AI Office, assessing and mitigating possible systemic risks and ensuring an adequate level of cybersecurity protection for GPAIMs and the physical infrastructure of the model.⁵⁶See Recital 111, Articles 51(1) and 55(1) as well as Annex XIII of the AI Act The fines meted out for non-compliance with the obligations for providers of GPAIMs are significant, ranging from 3% of their annual total worldwide turnover in the preceding financial year, or EUR 15, 000, 000 if higher.⁵⁷See Article 101(1) of the AI Act As with the famed General Data Protection Regulation, the AI Act has extraterritorial reach and companies based outside the EU which offer their services to the EU market would also need to comply with these obligations from 2 August 2025.⁵⁸See Recital 179 and Article 2(1) of the AI Act That said, these obligations will not apply to GPAIMs developed for the sole purpose of scientific research and development.⁵⁹See Article 2(6) of the AI Act While efforts to regulate LLMs differ in scope and approach, they share a common thread in balancing innovation with accountability.

The viral trend arguably served as a good marketing tool for Studio Ghibli, reigniting global interest for Studio Ghibli’s art and movies. However, LLMs’ ability to generate artistic works that simulate the style of an artist within minutes raises profound questions for the future of the creative landscape. The secret world of Gen AI is beginning to unfold, and the next chapter will depend on how regulatory authorities balance the interests of copyright holders with the agenda of facilitating technological advancement. With a nuanced approach towards regulation, generative AI could become a tool that enhances, rather than one that undermines the work of artists.

The post Totoro in Trouble? Assessing the Legal Implications of AI-Generated Ghibli Images appeared first on The Singapore Law Gazette.