# Generative AI's copyright problem > Generative AI platforms trained on billions of images and texts without permission face mounting legal challenges over intellectual property rights. - Title: Generative AI's copyright problem · Harvard Business Review - Summary: Generative AI platforms trained on billions of images and texts without permission face mounting legal challenges over intellectual property rights… - Keywords: generative ai, copyright infringement, intellectual property, fair use doctrine, stable diffusion, midjourney, dall-e 2, andersen v. stability ai, getty images, andy warhol foundation, content licensing, ai training data - Source: Harvard Business Review — https://hbr.org/2023/04/generative-ai-has-an-intellectual-property-problem - Published: 2023-04-07T12:05:51+00:00 - Read time: 4 min - Topics: generative ai, copyright infringement, intellectual property, fair use doctrine, stable diffusion, midjourney, dall-e 2, andersen v. stability ai, getty images, andy warhol foundation, content licensing, ai training data ## AI's remarkable capabilities Image generators like Stable Diffusion, Midjourney and DALL-E 2 produce visuals in styles from photographs to watercolors. The Museum of Modern Art in New York hosted an AI-generated installation from its collection. ## Text generators mimic style Text generators write essays, poems and summaries, proving adept at mimicking style and form, though they can take creative license with facts. ## How generative AI actually works Generative AI platforms are trained on data lakes and billions of parameters constructed by processing huge archives of images and text. They recover patterns to create rules and make predictions when responding to prompts. ## The legal risks emerge This training process comes with legal risks including intellectual property infringement. Courts must now decide whether copyright, patent and trademark infringement apply to AI creations and who owns AI-generated content. ## Artists sue over unlicensed training In Andersen v. Stability AI et al., filed in late two thousand twenty-two, three artists sued generative AI platforms for using their original works without license to train AI in their styles. ## The derivative works question The lawsuit claims AI-generated works may be insufficiently transformative from existing protected works, making them unauthorized derivative works. If courts agree, substantial infringement penalties can apply. ## Getty Images sues Stable Diffusion Getty, an image licensing service, filed a lawsuit against the creators of Stable Diffusion alleging improper use of its photos, violating copyright and trademark rights in its watermarked photograph collection. ## Fair use doctrine at stake These cases hinge on interpreting the fair use doctrine, which allows copyrighted work for criticism, comment, news reporting, teaching, scholarship or research, and for transformative use not intended by the original. > for purposes such as criticism (including satire), comment, news reporting, teaching (including multiple copies for classroom use), scholarship, or research ## Google's precedent on transformation Google successfully defended itself against a lawsuit by arguing that transformative use allowed scraping text from books to create its search engine. This decision remains precedential for now. ## Warhol case could reshape AI law A U.S. Supreme Court case against the Andy Warhol Foundation, brought by photographer Lynn Goldsmith over a Prince image, could refine copyright law on when art is sufficiently transformative. ## Willful infringement penalties If a business knows training data includes unlicensed works or AI generates unauthorized derivative works, it could face willful infringement charges. Damages can reach one hundred fifty thousand dollars per instance. ## Trade secrets at risk Companies also risk accidentally sharing confidential trade secrets or business information by inputting data into generative AI tools without proper safeguards. ## Developers must license training data AI developers should ensure compliance by licensing and compensating those who own intellectual property they add to training data, whether through licensing or revenue sharing from the AI tool. ## The LAION-5B problem Stable Diffusion, Midjourney and others built models on the LAION-five B dataset, which contains almost six billion tagged images from indiscriminate web scraping and includes substantial copyrighted creations. ## Opt-out versus opt-in Stability AI announced artists can opt out of the next generation of Stable Diffusion. But this puts the burden on creators to protect their IP rather than requiring developers to secure it first. ## Audit trails for transparency Developers should maintain provenance of AI-generated content by recording the platform used, settings employed, seed-data metadata and tags. This allows verification and demonstrates the user's intent. ## Creators must monitor datasets Content creators should proactively search compiled datasets for their work, including visual elements like logos and textual elements like image tags. Automation tools can handle terabytes of data cost-effectively. ## Trademark monitoring evolves Brands must monitor for derivative works trained on their images, not just exact logos or colors. Stylistic elements may suggest salient brand content was used to produce unauthorized AI-generated works. ## Businesses need contract protections Businesses should demand terms confirming proper licensure of training data, broad indemnification for infringement, and AI-related language in confidentiality provisions barring input of confidential information into AI tools. ## Creators can build their own AI Content creators with sufficient intellectual property libraries may build their own datasets to train AI platforms using lawfully sourced content, then license these tools to interested parties with cleared title. ## Key takeaway Generative AI platforms risk copyright infringement lawsuits because they train on unlicensed content, and courts are still deciding what counts as fair use.