Reflections on AI, generative tools, and the false dichotomy in game development
A conflict that may never have existed
Code, as a form of language, is in itself a form of art. It has syntax, rhythm, structure, and abstraction. It is an expressive medium capable of translating intention into behavior, vision into systems. Drawing a strict line between art and technology in the digital world means ignoring that the very language used to build interactive worlds is already a creative form.
We are used to describing Art and Science as opposites: inspiration versus calculation, intuition versus logic, soul versus algorithm. Yet video games have never been purely artistic. Rendering is applied mathematics, animation is physics translated into movement, gameplay loops are behavioral psychology, level design is spatial engineering, and optimization is systems architecture. The idea that artificial intelligence suddenly introduces “science” into an artistic territory is a convenient simplification, but historically inaccurate.
Digital art has always been mediated through technology.
The current discomfort is not about technology.
It is about identity.
The real fear: authorship and data origins
When part of the artistic world opposes generative AI, the central issue is not merely the tool itself.
It is the origin.
The recurring questions are always the same: who is the author, who signs the work, who is responsible for the creative gesture, and on which works the model was trained.
Many generative models have been trained on enormous quantities of images, texts, and online content. It is not always clear whether the original author gave consent, whether the material was protected by copyright, whether the usage falls under fair use, or whether the model is reproducing recognizable fragments rather than learning statistical patterns.
The suspicion is that the work of thousands of professionals has been absorbed without direct compensation.
This is not just a perception. In recent years, several concrete legal disputes have emerged.
The most well-known example is the case of Getty Images versus Stability AI, initiated in both the United Kingdom and the United States. Getty accused Stability AI of using millions of copyrighted images from its archive to train Stable Diffusion without licensing agreements. The central issue is not merely the duplication of individual images, but the large-scale use of proprietary datasets as training material.
At the same time, a group of artists in the United States filed lawsuits against Stability AI, Midjourney, and DeviantArt, arguing that training models on copyrighted works constitutes a systemic violation of intellectual property rights.
At the center of the legal debate lies both a technical and juridical question: does training a model constitute copying, or is it a transformative statistical process that falls under lawful use?
These cases are not marginal. They represent some of the first attempts to define a legal framework for generative AI.
But another question emerges, one discussed far less frequently.
What about code?
Software, including video games and graphics engines, is built daily on MIT libraries, GPL components, proprietary frameworks, open-source snippets, shared shaders, algorithms published in academic papers, and closed-source toolchains.
Modern projects integrate code written by hundreds of people who are not part of the final development team.
An engine is a mosaic of middleware, plugins, public repositories, mathematical models, and previous implementations.
Every shader is built upon published physical models. Every algorithm implements academic theories. Every pipeline inherits structures created elsewhere.
Digital creation has always been cumulative, what changes with AI is the scale and the automation.
No longer conscious citation, but massive statistical learning.
In software development, legal frameworks have gradually evolved to regulate this cumulative structure through licenses, attribution requirements, distribution clauses, compatibility rules between GPL and MIT, and commercial agreements.
In generative AI, that framework is still being defined.
The real point of friction is not the existence of training itself.
It is transparency.
If training datasets were declared, traceable, and based on consensual or proprietary sources, the debate would likely take a different shape.
The real problem is not the tool, it is the absence of shared rules.
Placeholder today, controversy tomorrow
In recent years we have seen games and media projects criticized for using AI tools even during preliminary stages of production.
One emblematic case involved Clair Obscur: Expedition 33, developed by Sandfall Interactive. In December 2025, the title was stripped of major honors at the Indie Game Awards (IGA), including Game of the Year and Best Debut Game, for violating an internal policy prohibiting generative AI tools.
The studio confirmed that AI had been used only in a limited way during early development phases to generate placeholder textures, which were later replaced in the final release.
The decision reignited a debate extending far beyond a single title: is it legitimate to penalize a work for temporary use of generative tools, even when those assets are absent from the final product? Where exactly is the line between prototyping and the final product, between technical support and ethical violation?
Not every controversy, however, results in formal sanctions. In several cases, the use of generative tools has triggered strong public backlash without institutional consequences. Promotional concept art withdrawn after users identified typical AI-generated artifacts, studios forced to clarify limited AI usage in marketing material, and communities divided between defense and condemnation are all examples of situations that affected public perception even without official penalties.
In these cases, the issue is not regulatory, but cultural. AI becomes a matter of identity and perception before it becomes a technical discussion.
In many situations, these tools were simply used for temporary concepts, placeholder visuals, synthetic prototype voices, or non-final promotional material.
Yet public reactions were often immediate and polarized.
Within game development, placeholders are standard practice: greyboxing levels, temporary assets, UI mockups, stock audio, and procedural test terrain are all part of normal production workflows.
The difference is not technical.
It is perceptual.
AI has become a cultural symbol before becoming merely a production tool.
Production reality: development is friction
Development is constant iteration. Mechanics change, art direction evolves, code is rewritten, and pipelines adapt continuously.
Generative tools can accelerate ideation, create rapid references, generate draft dialogue, produce shader variations, support prototyping, and reduce production bottlenecks.
But they do not guarantee cohesion.
The “Village People” effect
The real risk is not the replacement of humans by machines, but the fragmentation of the project itself. When generative tools are used without a clear direction and without professional oversight, the result is rarely a creative revolution. More often, it becomes a collection of elements that may function individually but fail to communicate with one another.
Visual language may shift from one scene to another, lighting may follow conflicting logic, interfaces may feel as though they belong to entirely different products, and systems written in incompatible ways may survive in isolation only to collapse when scaling becomes necessary. Each component can appear valid on its own. The problem emerges when everything is observed as part of a unified system.
Cohesion is a silent quality: when it works, nobody notices it; when it is absent, the fracture becomes immediately visible. It is the invisible work of those who integrate, refine, align, rewrite, and decide what must ultimately be discarded. In that sense, AI does not eliminate the professional. On the contrary, it makes the need for clear direction even more evident.
This is a dynamic already visible in other technological fields as well, as explored in my article about DLSS 5. Even there, the real issue was not the technology itself, but the way it was integrated within a coherent direction. The same principle applies to generative tools.
Art and Science have always been the same thing
Mature projects integrate multiple disciplines: technical artists, gameplay engineers, narrative designers, systems architects, and strong creative direction. This plurality is not a compromise between fields, but the very condition that allows digital works to exist. A video game is not a collage of isolated contributions. It is a coherent system where every part exists in relation to the balance of the whole.
AI enters this environment as another tool, not as an alternative principle. In the right hands, it becomes leverage. Without guidance, it becomes noise. The difference lies not in the tool itself, but in the direction behind it.
A constructive position
There is little value in choosing between blind enthusiasm and ideological rejection. The strongest position is recognizing that AI is support rather than replacement, that vision remains human, and that cohesion is a professional responsibility.
In an era where access to tools becomes increasingly democratic, what distinguishes one project from another is not the technology being used, but taste, direction, architecture, and integration capability.
In other words, the craft.
Personal considerations
As both a developer and an author, I do not place subjective emphasis on the goodness or justice of a tool. Instead, I evaluate the possibilities and means available for realizing a project.
We can debate datasets, copyright, toolchains, generative shaders, and language models endlessly. But no tool writes a good game on its own.
What ultimately makes the difference is still the same thing.
Hours spent refining mechanics. Endless debugging sessions. Iterations on a color palette. Refactors nobody will ever see, but which quietly hold everything together.
AI can accelerate. It can suggest. It can reduce friction.
But without vision, it remains little more than disconnected fragments digested poorly, like leftover pizza at midnight.
And perhaps that is precisely the point.
What truly makes the difference is always the same thing:
Sweat.
Cover kindly offered half by Chiara Roscini, half by ChatGPT
https://chiararoscini.com/gow/
