AI superhero video generation is easier than entertainment lawyers hoped. Google Gemini Omni can produce videos featuring recognizable superheroes if users craft prompts carefully, exposing a growing gap between what generative AI systems can do and what rights holders can actually control.
Key Takeaways
- Gemini Omni generates recognizable superhero videos through prompt engineering, bypassing intended safeguards.
- DC, Marvel, and Disney face enforcement challenges as AI video tools outpace takedown infrastructure.
- Disney and OpenAI signed a licensing deal covering 200+ animated characters for ChatGPT and Sora, suggesting rights holders are negotiating rather than relying on takedowns alone.
- The deal excludes real actor likenesses but raises unanswered questions about artist compensation and credit.
- Open-prompt generation systems create a cat-and-mouse dynamic between brand protection and user capability.
Why Prompt Engineering Defeats Current Safeguards
The core problem is architectural: generative video systems filter outputs for trust and safety, but determined users can work around those filters through careful language. Gemini Omni’s filtering is not granular enough to catch every variation of a superhero description, especially when prompts avoid direct character names and instead describe visual attributes, powers, costumes, and narrative contexts. A user might not ask for Batman directly but instead request a video of a masked vigilante in a black suit operating in a crime-ridden city—and the system generates something unmistakably Batman-like.
This is not a novel problem, but AI video generation makes it urgent. Image generators have faced similar issues for years, but video output carries higher commercial value and greater brand damage. A single viral video of a recognizable Spider-Man scene generated in seconds undermines years of licensing negotiations and theatrical release windows.
The Licensing Counterattack: Disney’s OpenAI Deal
Rather than wait for enforcement to fail, Disney has taken a different approach. Disney and OpenAI signed an agreement that licenses over 200 animated and masked characters from Disney’s multiverse for use in OpenAI’s platforms, including ChatGPT and Sora. This deal represents a fundamental shift: instead of fighting AI generation, major rights holders are now negotiating what AI systems can legally generate.
The agreement explicitly excludes the likenesses or voices of real actors, meaning deepfake-style celebrity content remains off-limits. But the scope—200+ characters—signals that Disney sees licensing as the future rather than litigation. The question is whether other studios follow, or whether Google, Meta, and other AI developers face a patchwork of licensing agreements that fragments the ecosystem.
What remains unresolved is compensation and credit. The deal raises unanswered questions about how artists whose work trained these models are acknowledged or compensated. If an animator’s character designs from a 1990s Disney film become part of the training data that powers Sora, does that animator receive royalties or credit when users generate new scenes with that character? The licensing deal does not address this.
Why This Matters for Users and Creators
For casual users, AI superhero video generation sounds like fun—a way to see favorite characters in new scenarios without waiting for studios to produce them. But the legal and ethical consequences ripple outward. If you generate and share a recognizable superhero video, you are now potentially liable for copyright infringement, even if the tool made it easy. Disney and DC have shown willingness to pursue takedowns aggressively, and user-generated content platforms are increasingly cooperative with rights holders.
For creators, the risk is even sharper. Independent artists who build fan communities around AI-generated content face sudden platform removals, account suspensions, and cease-and-desist letters. The studios are not primarily interested in individual users; they are interested in stopping large-scale monetization and brand dilution. But the enforcement net is wide and often indiscriminate.
The deeper issue is that generative AI systems are trained on decades of copyrighted creative work—films, comics, concept art, storyboards—without explicit permission or compensation. Gemini Omni’s ability to generate superhero videos is partly a reflection of how thoroughly Marvel, DC, and Disney content saturates the training data. The model learned to generate recognizable superheroes because it ingested thousands of hours of superhero media. Licensing deals like Disney’s are an attempt to legalize what was already happening in the training phase.
The Enforcement Paradox
Google, OpenAI, and other developers face a paradox. Stricter filtering makes their systems less useful and more frustrating—users will simply switch to competitors or find workarounds. Looser filtering invites legal liability and brand-safety backlash from studios. The Disney-OpenAI deal suggests that licensing is the middle path: allow generation of specific approved characters while maintaining enough filtering to prevent unlicensed output.
But this approach only works if every major studio negotiates a deal. If Marvel refuses to license to Google while Disney licenses to OpenAI, users will notice the inconsistency and the system will feel arbitrary. And smaller studios without the negotiating power of Disney will be left behind, their characters generated freely while they have no recourse.
Will Gemini Omni Tighten Its Filters?
Google has not publicly committed to stricter superhero filtering in response to this article’s findings. The company typically adjusts safety policies incrementally rather than through dramatic overhauls. Expect Gemini Omni to gradually become harder to prompt into generating recognizable DC and Marvel content, but not impossible—the arms race between users and filters is ongoing.
The real question is whether Google will follow Disney’s lead and negotiate licensing agreements with Marvel, DC, and other studios. Such deals would be more expensive than relying on filtering alone, but they would also provide legal cover and a competitive advantage. A system that can openly generate Spider-Man content under license is more valuable than one that can secretly do so under a veil of plausible deniability.
Can studios actually enforce against AI-generated content?
Yes, but imperfectly. Studios use automated detection tools to find and remove infringing content from platforms like YouTube, TikTok, and Reddit. However, detection systems struggle with AI-generated variations that do not match known copyrighted footage frame-for-frame. A Gemini Omni-generated Batman video will not match any existing Batman film, making it harder to flag automatically. Studios then rely on manual reports and cease-and-desist letters, which are reactive rather than preventive.
Why doesn’t Google just block all superhero generation?
Blocking all superhero generation would be both technically difficult and commercially undesirable. Technically, the model cannot distinguish between original superhero concepts and copyrighted ones—if you ask it to generate a caped vigilante, it will produce something that looks like Superman or Batman because those archetypes dominate its training data. Commercially, superhero content is popular and useful for entertainment, creative writing, and visual brainstorming. Removing it entirely would make Gemini Omni less competitive against other video generators.
Closing
The real story is not that Gemini Omni can generate superhero videos—it is that the entertainment industry’s approach to AI copyright is shifting from enforcement to negotiation. Disney’s deal with OpenAI signals that studios believe licensing is cheaper and more reliable than litigation. But this only works if studios coordinate and if users understand that generating unlicensed superhero content carries real legal risk. For now, the gap between what is technically possible and what is legally permitted remains dangerously wide.
Edited by the All Things Geek team.
Source: TechRadar
