Elon Musk’s AI company xAI has reportedly used clips from Hollywood films, including Universal Pictures’ Hellboy II: The Golden Army, to train its Grok AI models through an internal video annotation project called "Vision." Employees at xAI spent months labeling short video segments with detailed metadata such as shot composition, camera angles, lighting, cinematography style, and scene elements. The project, described by insiders as resembling a film school exercise, aimed to help Grok better understand visual storytelling and generate more realistic AI videos. Workers also contributed to another initiative known as "Moongazer," which focused on identifying specific visual components like transitions, captions, and infographics across a range of content—including news segments, tutorials, amateur videos, and foreign films. These efforts are part of xAI’s broader push to develop advanced video generation capabilities, with Musk announcing plans to release a full-length, watchable film powered by Grok by the end of 2026, followed by "really good movies" in 2027. The use of copyrighted material for AI training remains legally murky. While tech companies argue that training on existing media is essential for building sophisticated AI, studios and creators contend it may infringe on intellectual property rights. In August, Universal Pictures began adding warnings to its films stating that content may not be used to train AI—highlighting growing tensions between Hollywood and AI developers. xAI responded to inquiries about the project with a single phrase: “Legacy Media Lies,” offering no further explanation. Spokespeople for Universal Pictures did not comment. Legal experts say the core issue lies in whether AI training constitutes fair use. Matt Blaszczyk of the University of Michigan Law School noted that infringement risks exist at every stage—from data collection to output generation. Mark Lemley of Stanford University emphasized that high-quality training data is crucial for building capable AI systems, warning that restricting models to only licensed or amateur content could degrade performance. Other AI companies have faced similar legal challenges. OpenAI, Google, and Midjourney have all been sued over the use of copyrighted content, with Midjourney and Disney, Universal filing a joint lawsuit in June. Anthropic recently settled a $1.5 billion copyright case over the use of pirated books. Meanwhile, news organizations, including Business Insider, have sued Cohere, claiming unauthorized use of journalistic content. Despite these legal pressures, some AI tools still generate content inspired by or resembling copyrighted characters. When asked to create an image of Hellboy, ChatGPT initially refused, offering instead a character it called "Heckboy" as a "lookalike homage." Later, it allowed the inclusion of the name "Hellboy" in the image. In contrast, xAI’s Grok Imagine feature generated multiple AI depictions of Hellboy without restriction. OpenAI has since added guardrails, limiting the use of copyrighted characters in its Sora video tool and working with actors like Bryan Cranston to prevent deepfakes. However, experts like Yelena Ambartsumian, an AI and IP lawyer, warn that many companies are betting on future success to avoid paying for content now, banking on legal gray areas and the possibility of later settlement. As AI video generation advances, the legal and ethical boundaries of training data will continue to be tested, shaping not only how AI tools evolve but also who benefits from the creative work that powers them.