The advancement of generative artificial intelligence (“GenAI”) technology has brought tremendous benefits and convenience across multiple industries. The speed at which content is generated through text, images, audio, and video enables individual to exponentially increases productivity.
However, the extensive usage of GenAI brings up questions on copyright protection. One recent dispute is the application of GPT-4o model in generating images that resemble the iconic style of Studio Ghibli. Despite popularity on social media platforms, the viral usage raises serious legal and ethical concerns. It is unclear whether the model was trained on Hayao Miyazaki or Studio Ghibli’s work, and if so, whether prior permission or license was obtained. In addition, replacing artistic craftsmanship with a simple prompt can significantly curb originality and future creations. This article discusses copyright implications in the GenAI era by addressing (1) use of copyrighted data in AI training and (2) authorship of AI-generated content.
What is generative artificial intelligence?
Generative artificial intelligence is a subset of artificial intelligence that applies AI to produce content based on learned patterns. Leading models include GPT (OpenAI), Gemini (Google), Llama (MetaAI), and Claude (Anthropic). Several emerging players are challenging and reshaping the industry. Earlier this year, China-based company Deepseek introduced R1 model, which attracted global attention due to its significantly lower computation-related costs.
Many of you may have also heard the term Large Language Model (LLM), a subset of GenAI specializing in language-related tasks. LLMs understand text inputs through natural language processing (NLP) and generate human-like text based on given input. Let’s use OpenAI as an example. GPT-3 is a unimodal LLM that processes and generates only text, whereas the latest GPT-4o is a multimodal GenAI model that processes and generates various data types such as text, image, and audio.
Copyrighted Work in AI Training
Training data for GenAI models and LLMs are often collected from extensive and diverse sources, such as web scraping and crawling, public datasets and libraries, and licensed domain-specific data. Platforms such as ChatGPT also collect user-generated data to train and refine the next models.
In 2023, the New York Times sued OpenAI and Microsoft, accusing them of using its newspaper articles to train ChatGPT and Copilot. According to the NY Times, these models produced near-verbatim excerpts from their website. Such replica discouraged customers from maintaining the newspaper subscriptions, reducing the company’s revenues and funding resources to support journalism. OpenAI and Microsoft responded that the articles used in AI training qualified under “fair use”, a legal principle under the U.S. Copyright Act that permits the unlicensed use of copyrighted work. Similarly, Anthropic was sued by a group of writers who claimed that the company used their books in training Claude. Applying the fair use doctrine, Anthropic argued that the usage was “transformative” and promoted creativity by extracting uncopyrightable information through learning the books.
In Andersen v. Stability AI, a group of artists filed a class-action lawsuit against Stability AI, Midjourney, DeviantArt, and Runway AI for the unlicensed use of plaintiffs’ copyrighted images in training the image-generation platforms. While some claims were dismissed, U.S. District Judge William Orrick denied Stability AI’s motion to dismiss artists’ copyright infringement claim, allowing the case to move towards the discovery phase, scheduled to take place in September 2026. The result of this case will likely serve as a legal precedent in the U.S. regarding the usage of copyrighted work in AI training.
In June 2024, the European Union adopted the world’s first regulations on AI, the EU AI Act. The transparency requirements under the AI Act state that GenAI models must publish a summary of copyrighted data used in training. This requirement is expected to apply after 12 months of the act’s entry into force, which leaves sufficient time for AI companies to comply accordingly. This requirement protects artists and copyright holders and is likely to encourage major media players to join forces to preserve originality and creativity.
AI-Generated Content & Authorship
One major point of contention in Andersen v. Stability AI was whether the image-generation platforms directly infringed upon the copyrighted work or induced users to copyright infringement. It could also be a combination of both. This brings us to the next discussion regarding the end use, namely whether one can claim authorship over AI-generated content.
US
In March 2025, a federal appeals court in Washington, D.C., ruled against the copyright entitlement of AI-generated art. This decision signifies the latest attempt of U.S. officials in tackling the challenges posed by the GenAI models in copyright law practice. Back in 2018, Dr. Stephen Thaler applied copyright for a piece of visual art created by his AI system. The application was rejected by the office 2022, due to absence of human input under the US Copyright Act of 1976.
The decision was upheld, citing the irreplaceability of human involvement. U.S. Circuit Judge Patricia Millett wrote that U.S. copyright law “requires all work to be authored in the first instance by a human being”. The ruling marks a pivotal development in the intersection between copyright law and Generative AI, highlighting human authorship as a prerequisite to claim copyright. Purely AI-generated content falls into the public domain.
China
On the contrary, China does not adopt a black-and-white approach when it comes to authorship of AI-generated content. In fact, authorship was granted to an AI-generated image for the first time in 2023.
Plaintiff generated the image via Stable Diffusion through prompts, which was then used by the defendant on a social media site without permission. According to the Copyright Law of the People’s Republic of China, the disputed image constitutes a “work”. The court emphasized “originality” and “intellectual achievement” in its ruling. The prompting life cycle from concept design, continuous adaptation, aesthetic fine-tuning to final output justifies both the intellect and originality requirements. Therefore, a right of authorship was granted.
It is unclear whether the ruling will be established as a precedent on the appellate level in China, but the landmark decision opened up the possibilities when it comes to the otherwise mutually exclusive relationship between authorship and AI-generated content.
EU
Similar to the US, the Czech Copyright Act states that author may only be a natural person. In May 2024, the Prague Municipal Court denied copyright protection of an AI-generated image. The claimant created an image using AI and sent it to a law firm, who then uploaded the image to their website without acquiring consent. The claimant aimed to confirm the authorship of the image and prohibit the law firm from using the image. Authorship was ultimately rejected because the claimant cannot provide sufficient evidence regarding how the image was created. Beyond the prompt, the claimant needed to show the additional procedures on how the image was created. The Prague Municipal Court further commented that AI-generated content is not protected by copyright as a rule of thumb. In addition, Section 2(2) of the German Copyright Act (UrhG) requires a “personal intellectual creation” to qualify for protected works. This implies that purely AI-generated content without human input does not constitute “work”, thus ineligible for copyright protection.
Under the newly adopted EU AI Act, companies and end users need to comply with the transparency requirements. AI-generated or AI-modified content – text, image, audio and video – need to have a clear label indicating its AI-generated nature so that others are aware of the information when they access such content.
Takeaway messages
Human v. Machine:
Content creation has been drastically shaped by the fast-developing generative AI tools. With one prompt, users can generate content in various formats instantly. This change harms artists’ interests in multiple ways, such as the unauthorized usage of their copyrighted work in AI training, the erosion of artistic identity, and depreciation of market value. It has become increasingly difficult for artists to stand out and profit from their work when the machine can produce a replica in seconds.
Legal Framework
The legal landscape is rapidly changing and adapting to the fast development of GenAI technology. For example, with the EU AI Act expected to be fully applicable in 2026, organizations across the EU and globally should implement compliance measures accordingly on the development, deployment, and oversight of AI systems. It is likely that other countries and regions will follow suit and adopt AI-related regulations soon.
Authorship & AI
Most court rulings deny copyright protection to AI-generated content, as reflected by the U.S. and Czech court decisions. Authorship can only be granted to a natural person, and the content generated by a model involves minimal, if any human creativity. However, China’s decision to grant authorship to an AI-generated image challenged this principle. The Beijing Internet Court justified its decision based on the sophisticated prompt design and continuous fine-tuning that led to the final output.
Best Practices
As artists and content creators, it is more important than ever to adapt to the GenAI era. Knowing the fundamentals of copyright law and stay up to date with latest regulations and court decisions equip you with solid understanding of your rights and responsibility. Including watermark or metadata in your artworks is crucial to assert authorship in copyright dispute. Software solutions also come in handy in safeguarding your artistic pursuit, such as automatic detection of copyright infringements and unauthorized usage of your work in AI training.
