We Tested JSON Prompting in Veo 3: Useful Hack or Just Influencer Hype?

The world of AI-assisted filmmaking and content creation is evolving at a breakneck pace, and with every new tool or technique, creators seek ways to enhance their workflows and outputs. One of the hottest topics recently has been the rise of JSON prompting in Veo 3, a structured prompting technique that promises better, more consistent AI-generated video outputs. But is it truly a game-changer or just another influencer fad? Alongside this, the AI community is buzzing about Higgsfield Steel's controversial marketing approach and a groundbreaking AI system called AlphaGo, which can autonomously design new AI architectures. This article dives deep into these subjects, unpacking their implications, testing claims, and exploring what the future might hold.

JSON prompting has recently become a hot topic on social media, especially within the Veo 3 AI video generation community. At its core, JSON (JavaScript Object Notation) is a human-readable text format used to structure information, combining elements of code and natural language. The idea behind JSON prompting is to impose a strict structure on the AI prompts by dividing them into fields such as description, style, camera, lighting, and environment.

This structured approach is meant to improve output consistency across multiple shots, especially when working on short films or sequences where characters, scenes, and lighting need to stay constant. By modularizing the prompt into reusable components, creators can easily tweak individual elements—like switching from a wide shot to a medium shot by only adjusting the camera field—without rewriting the entire prompt.

One exciting possibility is integrating JSON prompting with cloud code or visual language models (VLMs) that can generate these JSON prompts from sketches, frames, or descriptions. For instance, AI models like ChatGPT can take a rough shot description and output a fully structured JSON prompt, which can then be used to generate the video.

Creators like Dave Clark have experimented with this concept, developing agentic workflows that convert entire scripts into shot-by-shot JSON prompt sequences. This automation could drastically speed up the process of creating AI-generated short films, allowing for more efficient iteration and consistency.

However, this approach is still in its infancy. While JSON prompting offers a more organized way to approach AI video generation, it remains to be seen how scalable and effective it will be for high-end productions that require granular control down to individual frames rather than shots.

Despite the enthusiasm, JSON prompting has its critics. Some argue that the hype around JSON is largely influencer-driven and that a well-crafted natural language prompt can perform just as well, if not better. Jason Zada, a prominent figure in the AI space, has voiced skepticism, suggesting that highly structured prompts might add unnecessary complexity without significantly improving results.

The core of this debate revolves around consistency and granularity. Movie making is fundamentally shot-by-shot, and while JSON prompting encapsulates this structure, it still operates on a shot level. High-end productions often require control at even finer levels, such as frame-by-frame adjustments, which JSON prompts may not accommodate effectively.

Moreover, JSON formatting introduces additional characters—parentheses, labels, and whitespace—that consume more tokens in AI models, potentially increasing cost without adding semantic value. Alternatives like YAML, another markup language, have been proposed as more token-efficient ways to structure prompts, but ultimately, the AI models are trained on natural language, not JSON or YAML scripts.

To evaluate these claims, tests were conducted using Gemini and Veo 3. A simple scene was set up: a close-up of a coffee cup at a diner, tilting up to a detective looking out the window. The same shot was prompted twice—once using a structured JSON prompt and once using a traditional paragraph-style text prompt.

The JSON prompt produced a shot that matched the envisioned scene quite closely, including specific time markers and a handheld camera feel. The paragraph prompt generated a different interpretation with a dolly-like effect, showing that even with the same input, outputs can vary due to randomness in seed generation.

Further tests with dialogue lines and reverse shots revealed that neither approach perfectly captured every nuance without precise prompting, such as screen direction. Both methods have their strengths and weaknesses, and the results suggest that JSON prompting offers better structure but not necessarily superior creative control or output quality.

One of the most contentious topics in AI-generated media is copyright. Interestingly, Veo 3 has shown the ability to generate complex copyrighted elements, such as logos and character likenesses, without triggering copyright errors. For example, a JSON prompt depicting Superman was successfully generated without being blocked.

This raises questions about the legal gray areas surrounding AI-generated fan content. Some argue that such creations fall under parody or fanfiction laws, which provide some leeway. There are popular AI-generated fan projects, like a Star Wars fan series featuring characters and scenes never included in the original films, which remain available and monetized on platforms like YouTube without takedowns.

However, the ethics of using real people's likenesses or copyrighted material without permission remain a hot-button issue. While some companies like Adobe are cautious due to past mistakes, others like Higgsfield Steel take a more cavalier approach, leading to controversy.

Alongside prompting techniques, creators are discovering innovative hacks to extend Veo 3’s capabilities. Google recently shared a neat trick using the “first frame” feature. By uploading a marked-up image with text boxes indicating desired actions and then instructing the AI to “follow the instructions on the image,” users can gain greater control over the resulting video.

This method mimics how directors traditionally give notes using frame annotations. It allows for more precise storytelling, albeit at the cost of a few unusable seconds at the start as the AI dissolves the text. This hack exemplifies how AI tools are still being explored and adapted in creative ways beyond their initial design.

Higgsfield Steel is a new AI product that has stirred debate due to its marketing and ethical stance. The tool claims to “steal” camera angles, composition, color, and lighting from source images to generate new content. Its demos include recreations of copyrighted characters and real people—such as Daniel Craig as a Nightwalker from Game of Thrones—without clear permissions.

What makes this more troubling is the marketing strategy embraced by several top AI creators who appear to be paid ambassadors. The messaging encourages users to “forget about copyright” and “just grab it,” a stance that many find problematic and “icky.”

This approach contrasts sharply with companies like Adobe, which emphasize respecting copyright and artist rights. Higgsfield Steel’s founder is a former Snapchat executive with funding from Kazakhstan, adding an international dimension to the legal and ethical questions.

While such aggressive marketing may gain short-term attention, it risks alienating professional studios and enterprises that prioritize ethical AI use. The AI community is watching closely to see how this plays out as the industry matures.

Perhaps the most mind-bending development discussed is AlphaGo, an AI research system designed to autonomously invent new AI architectures. Traditionally, AI model design has been a human-driven process, requiring deep expertise and creativity to assemble novel combinations of model components.

AlphaGo breaks this task into four agentic behaviors:

This system can iterate thousands of times faster than human researchers, exploring vast design spaces and potentially discovering architectures beyond human imagination. It represents a critical step toward artificial superintelligence (ASI), where AI systems surpass human intelligence and can self-improve exponentially.

AlphaGo’s development signals a future where AI not only aids human creators but also autonomously advances AI technology itself. This could accelerate innovation but also raises concerns about control and oversight.

In this context, Nvidia CEO Jensen Huang’s recent comments add perspective. He highlighted that just 150 people with the right resources can change the world, referencing teams like OpenAI and DeepMind. His vision anticipates rapid advancements fueled by concentrated talent and funding, underscoring the importance of strategic leadership in this transformative era.

The landscape of AI in media creation is dynamic and complex. JSON prompting in Veo 3 offers an intriguing method to impose structure and consistency, particularly useful for short-form content, but it’s not a silver bullet for all scenarios. The debate between structured and natural language prompts highlights the evolving nature of AI-human collaboration.

Meanwhile, tools like Higgsfield Steel challenge ethical boundaries with provocative marketing, reminding the community of the ongoing tensions between innovation and responsibility. On the frontier of AI research, systems like AlphaGo hint at a future where AI designs itself, potentially revolutionizing technology but also demanding careful stewardship.

For creators, the key takeaway is to stay informed, experiment with new tools, and maintain a balance between leveraging AI’s power and respecting the creative and legal frameworks that safeguard art and culture.

As AI continues to reshape storytelling and content creation, the journey is as exciting as it is uncertain. Embracing this future with curiosity, caution, and creativity will be essential for all who navigate the new digital frontier.