AI-Generated Content Copyright
Copyright of AI-Generated Content
What is copyright in AI-generated content?
- AI-only output is hard to treat as copyright-protected
- Parts with human creative contribution can be protected
- Training-data legality is still being debated country by country
What this principle really changes is less infringement liability than whether you can assert rights at all. A brand image generated from a few lines of prompt can be reused by a competitor as is, and you would struggle to claim infringement. If an outside designer delivers work made purely with AI, the company has paid for a file it has no exclusive rights to. So the practical question shifts from 'can we use it?' to 'how much human input do we build in?'
There are two issues, outputs and inputs, and the one that actually costs companies money is similarity of the output. Whether training data is lawful is a dispute between the model provider and rights holders, and countries reach different conclusions. Japan has a provision broadly permitting use for information analysis, while the EU lets rights holders opt out of text and data mining. By contrast, if an image we generate turns out to resemble an existing work, liability for infringement falls on the company that published it, regardless of the training-stage debate.
That does not mean every form of protection disappears. Even a logo made with AI can still be registered as a trademark, and a product’s appearance can still be defended under design rights or unfair competition law. In copyright registration practice, applicants are told to exclude AI-generated parts and declare only the human contribution. In the U.S., there is already a case where the images in a comic book made of AI images were denied protection, and only the text and the arrangement were recognized as copyrightable. There is no fixed percentage for 'how much you must edit'; what matters is whether a person substantively changed the expression.
The most common misreading in practice is 'no copyright means we can use it however we like.' An output having no rights attached is a completely separate matter from that output infringing someone else’s work. If you put an image resembling a famous character on an outdoor ad, the removal costs and settlement are still the company’s to bear. Relying on an AI provider’s indemnity is also risky. It usually applies only to paid plans, and comes with a condition that excludes you from coverage if you turned off safety filters or put a specific artist’s name in the prompt.
John KimInstructor John Kim, author of "B2B Marketing Design" and CEO of Nanum Consulting
Jaryong KooMarketing PhD and leading authority on AI, data analytics, and personal brandingThe debate starts from the Copyright Act’s definition of a work as a 'creative production expressing human thoughts or emotions.' In 2023, the Ministry of Culture, Sports and Tourism and the Korea Copyright Commission set out Korea’s position in the Guide to Generative AI and Copyright, and the U.S. Copyright Office likewise reaffirmed the human authorship requirement.
Consider a hypothetical case. Company A, a household goods manufacturer with 320 employees, had a nine-person marketing team generating over 200 catalog and social media images a month with generative AI. There was not a single line of internal policy, and no record was kept of which tools were used to make what. One remark from a retailer’s merchandiser kicked off a three-month overhaul.
- It started with a remark from a retailer’s merchandiser: 'I feel like I’ve seen this illustration somewhere.' That night, the marketing team lead took down all 11 images posted across three channels.
- A legal manager reviewed 412 AI-produced items from the past six months. Only 37 recorded which tool had been used, and more than half of those were files that came from team members’ personal accounts.
- The review scope was narrowed to three types: outdoor ads, packaging, and product detail pages. The team lead said, 'If we ask people to check every internal report image, nobody will follow the rule,' and made reverse image search mandatory only for externally published materials.
- For cover images that needed commercial use, designers had to adjust the composition and colors by hand and keep their PSD layer history. 'Your edit history is your claim to the rights' became a saying on the team.
- They consolidated to two company-owned accounts, checked the terms for the scope of commercial use and indemnity conditions, and added disabling safety filters to the list of prohibited actions.
- Three months later, pre-publication rejections fell from 14 a month to 3, and the share of items with production records went from 9% to 100%. There has not been a single post recall since.