Elon Musk appears determined to see Grok’s name illuminated among the most recognized technologies of the modern AI age. His ambitious vision extends beyond simple conversational models into the broader realm of multimedia generation. In July, his team unveiled Grok Imagine, a powerful and experimental image and video generation platform designed to expand the capabilities of the xAI ecosystem. Earlier this month, Musk publicly shared additional details of his creative roadmap, explaining that the company intends to produce a fully watchable feature-length film before the end of 2026. By 2027, he emphasized, their aim is to deliver what he described as genuinely high-caliber cinematic productions. Musk has consistently showcased the chatbot’s advancing skill in visual generation—demonstrating everything from elaborate reinterpretations of classic scenes such as the climactic moment of “King Kong” to imaginative re-creations of the superhero saga “Iron Man,” in which he humorously inserted himself in the role of Tony Stark.
Meanwhile, behind the scenes, employees at xAI have been deeply engaged in a series of intensive internal research projects focused on video annotation, each intended to strengthen the underlying architecture that supports visual understanding in their systems. According to individuals familiar with these operations, August marked the beginning of one particularly large-scale initiative internally codenamed “Vision.” In this project, dozens of AI specialists—referred to internally as AI tutors—painstakingly labeled and annotated thousands of short clips. Insider reports indicate that the onboarding process exposed workers to footage drawn from Universal Pictures’ “Hellboy II: The Golden Army.” The annotation tasks were far from simplistic. Each worker was asked to analyze five- to ten-second segments with meticulous care, recording details about shot composition, camera angles, lens depth, cinematographic style, lighting intensity, and scene framing. Annotators also wrote exhaustive descriptive breakdowns identifying every distinguishable object or element present in the frame, along with its relationship to the broader spatial context of the setting.
As is true throughout the rapidly expanding AI sector, xAI’s internal data practices highlight an unresolved tension between technological advancement and copyright regulation. The debate centers on whether the use of copyrighted media as training material constitutes fair use or infringement. Hollywood studios and numerous rights holders maintain that training AI systems on protected creative content without permission threatens their intellectual property. Technology companies, on the other hand, argue that such material is essential for producing systems capable of nuanced understanding and lifelike output. The outcome of this ongoing legal and ethical standoff will profoundly influence not only how AI models like Grok evolve but also determine which entities stand to profit from the creative works on which such models rely.
Legal scholars have been watching closely. Matt Blaszczyk, a research fellow at the University of Michigan Law School, explained to Business Insider that potential legal exposure exists at virtually every step of the data pipeline—from initial downloads and storage to filtration and eventual output generation. The core issue, he noted, lies in whether the copyrighted content is being used purely for the machine’s internal learning process or directly to create expressive outputs that reproduce the protected work. When Business Insider sent a detailed list of follow-up questions to xAI regarding these issues, the company issued a terse response: “Legacy Media Lies,” a phrase it repeated in multiple subsequent email exchanges instead of providing clarification. Universal Pictures, whose footage reportedly appeared in annotation sessions, declined to comment publicly. Interestingly, the studio began attaching onscreen warnings in August alerting that its content may not be used to train AI models.
Two xAI workers remembered annotating clips from a broader array of sources, including different Hollywood productions, television series, and even foreign and independently produced videos. Several participants described Vision as comparable to highly specialized coursework found in advanced film programs, remarking that the analytical rigor demanded by the project exceeded that of most prior assignments within xAI. In parallel, employees were also involved in another internal undertaking titled “Moongazer,” centered on identifying and classifying specific video elements such as transition sequences, subtitles, overlays, and graphical infographics. The media used for Moongazer extended beyond Hollywood content to include news reports, tutorial videos, and a variety of other online materials.
This practice places xAI within a rapidly growing number of AI companies that are pushing to develop video-generation technologies while navigating an intricate legal boundary between fair use and infringement. Mark Lemley, who directs Stanford University’s Program in Law, Science, and Technology, emphasized the necessity for studios to balance two priorities: protecting the economic value of their creative property and supporting innovation that could ultimately enhance filmmaking itself. Lemley further argued that robust AI training depends on exposure to high-quality, professionally produced material. Restricting data solely to amateur or freely licensed content, he warned, would yield inferior, less sophisticated models. OpenAI echoed similar reasoning in its submission to a UK parliamentary committee, pointing out that modern copyright law extends to virtually every form of human-created expression—ranging from simple blog posts and photographs to software fragments and government records—making it practically impossible to train leading AI models without incorporating some copyrighted material.
The unresolved nature of these copyright debates has already sparked numerous high-profile lawsuits. In June, Disney and Universal filed a joint complaint accusing the text-to-image company Midjourney of infringing intellectual property rights by training on protected film material. Midjourney defended its actions in court by asserting that AI training constitutes a legitimate form of fair use under U.S. law. Around the same time, Anthropic reached a settlement worth $1.5 billion to resolve allegations that it had used pirated literary works to train its large language models. Various news organizations, including Business Insider, have launched similar lawsuits; in February, Insider joined a consortium of media outlets in suing the AI company Cohere, alleging that its model ingested copyrighted journalism. Cohere has moved to dismiss those claims, maintaining that its data use also falls under fair use.
Intellectual property researchers, such as Hayleigh Bosher of Brunel University, have noted that the legal field is scrambling to adapt to the unprecedented pace of generative AI innovation. Bosher explains that courts tend to focus heavily on whether AI-generated outputs pose direct commercial competition to the original works and how those outputs could affect their underlying markets. Some AI firms have started deploying measures to mitigate such risks. For instance, when OpenAI released its newest version of Sora—the company’s experimental AI-driven video generator—it initially permitted users to produce scenes featuring recognizable characters from popular entertainment franchises. After just a few days, those permissions were withdrawn as OpenAI enforced tighter restrictions to minimize unlicensed reproduction. CEO Sam Altman later published a statement pledging to give rights holders a more precise level of control over how their characters and likenesses are generated, suggesting an opt-in strategy augmented by further controls. The studio also revealed it was working directly with actor Bryan Cranston to reduce the risk of deepfake misuse on its platforms.
Nonetheless, tests conducted across different generative tools reveal inconsistent enforcement. Some systems, such as ChatGPT, Midjourney, and Google’s Gemini, occasionally reject prompts that explicitly reference copyrighted material but may still produce derivative reimaginings. For example, when Business Insider asked ChatGPT to render an image of Hellboy, it explicitly refused to recreate the copyrighted character, instead offering to produce an “inspired” alternative—a red, horned demon humorously dubbed “Heckboy.” Later tests revealed shifting behavior: the chatbot proposed creating “lookalike homages” and eventually even labeled an image “Hellboy,” illustrating the ambiguity around compliance. In contrast, xAI’s Grok Imagine tool reportedly provided numerous AI-generated images and short clips representing Hellboy when prompted, highlighting differences in companies’ ethical filters and technical safeguards.
When contacted for comment, representatives from OpenAI, Google, Midjourney, and Anthropic did not respond. Lemley reiterated that prompting a model to recreate specific copyrighted characters carries significant legal risk. Suggesting or facilitating near-identical imitations of protected characters like Hellboy, he said, could constitute particularly troublesome territory.
In a related observation, AI governance and intellectual property attorney Yelena Ambartsumian told Business Insider that many AI startups are effectively betting on the transformative nature of their outputs as a legal defense. She explained that the industry’s prevailing calculation seems to be that companies intend to push forward using as much high-quality data as possible—contending that if their interpretations qualify as transformative, they can postpone licensing costs until success or avoid them altogether if the enterprise fails. In essence, these companies are wagering that future courts will interpret their massive data consumption as a necessary condition for innovation rather than as infringement.
Individuals wishing to share insights about xAI’s internal operations are encouraged to contact the Business Insider reporting team. Tips can be sent securely to reporter G. Kay via personal email or through the encrypted messaging service Signal at the provided phone number, with additional guidance available for maintaining confidentiality while communicating sensitive information.
Sourse: https://www.businessinsider.com/grok-workers-hellboy-hollywood-movies-elon-musk-xai-2025-10
