(By You Yunting) Recently, Adobe, a software company that has long been vocal in enforcing its copyrights, together with its management, was sued by its own shareholders in the Northern District of California for alleged securities fraud based on false statements, due to using pirated materials to train its AI. The core allegations in the complaint include Adobe’s use of a large number of pirated books and AI-generated images sourced from its competitor Midjourney to train its AI models. I have often discussed the legal issues of training AI on pirated content, but today I would like to talk about what legal risks arise from using AI-generated images to train AI models.
Let me first introduce the case. The lawsuit against Adobe is essentially a securities-fraud misrepresentation claim. Adobe is being sued for copyright infringement for using datasets including SlimPajama and Books3, which contain pirated books, to train its Slim LM model. Separately, Bloomberg reported that Adobe’s image-generating AI, Firefly, used in its training data some AI-generated images from platforms like Midjourney, and that Midjourney’s own AI model was trained on a large number of unauthorized images. However, in its filings with the U.S. Securities and Exchange Commission and in its public promotions, Adobe consistently represented that it “trains AI only on licensed content”. Both practices are alleged to constitute misrepresentation that misled investors.
Whether using images generated by Midjourney AI constitutes infringement first depends on whether those AI-generated images themselves are protected by copyright. The fundamental principle of copyright law is that only works originating from human authorship are entitled to copyright protection. Purely AI-autogenerated content, lacking human creative input, can hardly qualify as a “work” under copyright law. If the image itself has no copyright, then subsequent use of it for training would naturally be unlikely to constitute infringement of that image’s copyright.
However, Chinese court rulings have introduced some controversy on this issue. In 2023, the Beijing Internet Court, in the “Spring Breeze Brings Tenderness” case, held that the plaintiff, through carefully designed prompts, parameter adjustments, and final image selection, had demonstrated personal creativity, and thus the AI-generated image constituted an artistic work entitled to copyright protection. If this holding stands, then using AI-generated images to train AI could also raise potential disputes over infringement of the prompt author’s copyright in that artistic work.
But the trend shifted in 2024. The Zhangjiagang Court in Jiangsu Province, in the “Fantasy Wing Transparent Art Chair” case, took the opposite view, ruling that the plaintiff had only entered simple prompts without a record of substantial intellectual input, and therefore the image did not constitute a copyrightable work. In 2025, the Huangpu District Court in Shanghai, in a case involving prompt infringement claims over AI-generated images, held that even relatively lengthy prompts still amounted to abstract ideas rather than literary works—let alone the images themselves—and dismissed the plaintiff’s claims. Overall, in my view, the Beijing Internet Court’s decision was a good exploratory effort, but the current mainstream trend still leans toward denying that AI-generated images constitute works protected by copyright law.
In practice, the main risk of using AI-generated images as training material comes from the distinctive expressive elements embedded in those materials. The Midjourney AI model was trained on a massive number of internet images, many of which were not authorized by their original creators. If those distinctive expressions in the training data are retained by the AI and later reproduced in new output images, those generated images could still create copyright conflicts with the original works.
Of course, from a compliance perspective, Adobe’s approach appears more cautious than that of many peers, including Midjourney. Adobe did not directly scrape images from Midjourney’s website in bulk to build its training library; instead, it allows users to upload images to Adobe Stock. Adobe then decides, according to its rules, whether to incorporate the relevant content into its training system. For material that is adopted, the uploader may even have an opportunity to receive corresponding rewards.
This model at least avoids two common types of risks. The first is violation of website terms of service. Even if AI-generated images are not ultimately protected by copyright law, large-scale scraping of platform content may still breach the website’s terms of service. The second is the risk of unfair competition. If a company systematically scrapes a competitor’s platform for large volumes of accumulated content to train its own commercial model, even if individual images lack copyright, such conduct could still be found to constitute misappropriation of others’ achievements and disruption of market competition order.
However, the volume of AI training material is huge. The complaint cites a media report stating: “Adobe said that images from Midjourney accounted for only 5% of its training materials. That is not a great defense. The company has 248 million licensed images, so that ‘only’ could be as many as 1.25 million images.” (The original data here is erroneous; 5% of 248 million is actually 12.4 million images). On the surface, with such a large base, the copyright risks are indeed substantial. But if Adobe actually reviewed the images that entered its training system according to its rules—though such review cannot eliminate risks 100%—at least the direct legal exposure for image infringement would be relatively low.
Adobe’s real problem lies in concealing facts. On one hand, it publicly promoted Firefly as different from Midjourney, claiming that its generated content is commercially usable by default, that it trains only on licensed content, and that it commits to bearing responsibility for copyright disputes. On the other hand, it quietly incorporated Midjourney’s copyright-controversial AI images into its training data. The stark discrepancy between its publicly claimed training data sources and the actual training data it used is the core issue in the shareholder lawsuit.
The shareholder complaint ends with a highly cautionary passage worth quoting: “For a software company like Adobe, intellectual property, including copyright protection, is the lifeblood of the company. If third parties can freely infringe Adobe’s copyrights, Adobe’s software would be worthless, and the company’s revenues would collapse. By the same token, if Adobe infringes the copyrights of others, it exposes itself to massive damages. Worse still, if Adobe’s own software development practices violate intellectual property laws … the consequences would be catastrophic. These violations would threaten the substantial recurring revenue that the company derives from its software, and would therefore not only expose the company to liability for damages, but also adversely affect its revenues, profits, and growth.”
In fact, this very language was originally used by Adobe to articulate its value proposition to investors, but now it has been turned against the company by its own shareholders.
Disclaimer of Bridge IP Law Commentary
Short Link: