Interview: New Indie Game Uses Local AI Model Trained Entirely on Studio's Own Artists

Studio Atelico is developing two closely connected projects: an on-device ethical AI engine for game developers and Bobium Brawlers, a mobile card-and-dice battler in which players create creatures through text descriptions. The game translates each description into a portrait matching its established visual identity and a playable deck that follows rules authored by the development team.

The studio, formed by AI and game industry veterans from companies including Uber, Meta, SEGA, and Creative Assembly, is approaching generative technology as a player-facing mechanic rather than primarily a production shortcut. Its image model was trained using artwork commissioned from participating artists, while its publicly available Artist Rights Contract proposes explicit consent, project-specific usage rights, upfront compensation, and continued participation in a project’s success.

That approach has still required considerable experimentation. After producing more than ten prototypes, Studio Atelico found that giving AI greater control did not necessarily make a game more creative or enjoyable. Excessive variability could make its rules feel arbitrary and prevent players from developing mastery. We spoke with Studio Atelico CEO and Co-Founder Piero Molino to learn more.

Studio Atelico is building both an AI engine and a game that uses it. What made you decide to develop the technology and the game in parallel, rather than building one first and proving the other later?

Piero Molino, CEO of Studio Atelico: From the beginning, we believed the engine and the game would make each other better. Building Bobium Brawlers constantly exposes new gameplay ideas and technical challenges, while every improvement to the engine expands what's possible in the game.

We never saw the engine as something that would be "finished" first and then applied later. The game is where we discover which mechanics are actually compelling, while the engine gives us the tools to make those mechanics practical. Developing them in parallel creates a continuous feedback loop that's difficult to achieve if you build either one in isolation.

At the same time, we're already working with other studios using the engine, which gives us a larger perspective and helps ensure we're building something broadly useful, not just something tailored to one title.

A lot of AI conversations in games focus on making existing workflows faster. Your pitch is more about creating experiences that were not previously possible. How do you define that difference in practice?

Piero Molino: Most conversations around AI focus on whether it can make games cheaper or faster to build. We're much more interested in asking what kinds of games AI makes possible that couldn't exist before.

That's a fundamentally different design philosophy. Rather than replacing existing workflows, we use AI to create mechanics centered on player creativity and personalization. In Bobium Brawlers, players aren't choosing from a predefined list of creatures; they're inventing their own, and the game immediately turns those ideas into both artwork and gameplay within a designer-authored framework.

On top of that, Ai provides new input modalities that were not possible before (text, voice, drawing) and more unscripted, emergent gameplay. That's the opportunity we find exciting: expand what games can be.

When you started building AI-native game technology, what assumptions did the team have that turned out to be wrong, incomplete, or more complicated than expected?

Piero Molino: One assumption we had early on was that the more freedom we gave AI, the more interesting the game would become. We discovered it was much more nuanced than that.

We built more than ten prototypes that explored different points along the spectrum. Some were essentially traditional games with AI sprinkled on top, while others gave AI much more control over the experience. Neither extreme proved to be the most fun. Too little AI didn't create experiences that felt genuinely new, while too much AI often made the game feel arbitrary and made it harder for players to develop mastery by creating a mental model of the rules of the game.

The breakthrough for us was realizing that AI works best when it's paired with strong game mechanics. Designers define the rules, creative direction, and constraints, while AI provides personalization, creativity, and surprise within those boundaries. We don't think AI should replace authored design; it should expand what's possible inside it.

That realization has shaped both the development of the Atelico AI engine and Bobium Brawlers. Today, we think the most exciting AI-native games won't be the ones that maximize AI, but the ones that use it to give players forms of creativity and expression that they simply did not have before.

AI is obviously a very touchy subject for artists, creators, and developers. Does your technology generate things? If so, what was it trained on?

Piero Molino: Yes, our technology generates content during gameplay, but we believe how those models are built is just as important as what they produce.

For Bobium Brawlers, our image generation model was trained on artwork created by our own artists, with their participation, specifically to capture the game's visual style. The goal isn't to imitate the broader internet or replace artists; it's to allow players to instantly create creatures that fit naturally within the game's world.

More broadly, we've also made ARC, our Artist Rights Contract, available to everyone, because we believe consent, transparency, and fair participation should be part of how AI is developed. We think creators should have visibility into how their work is used and share in the value when it contributes to training models.

Many developers are skeptical of AI in games, especially when it feels like a cost-cutting tool or a replacement for creative labor. How do you respond to that concern, and how do you make sure the technology remains in service of developers and players?

Piero Molino: We understand those concerns, and in many cases we share them. We are passionate gamers ourselves, and nobody wants games filled with generic, low-quality AI content.

That's why we've built the engine around a very different philosophy. We don't believe AI should replace creative teams or generate content without direction. Developers define the world's rules, tone, characters, and authored content, and AI works within those boundaries. One way to see it is just a more advanced version of procedural Content Generation.

We think "AI slop" isn't an inevitable consequence of AI, but the result of using AI without creative intent or control. If you provide developers with systems that produce dynamic, personalized experiences while remaining faithful to the vision they've authored, they’ll feel empowered.

Ultimately, we’d love developers to build games that weren't possible without AI, not make existing games cheaper to produce.

Can you share an example of something the AI engine enabled inside the game that would have been difficult or impossible to create with a more traditional pipeline?

Piero Molino: Our creature creation system in Bobium Brawlers is probably the clearest example. A player can type a short description of an original creature they've imagined, and within seconds the game generates both a portrait in the game’s art style and a gameplay deck that adheres to the description while following designer-authored rules.

Every player ends up with a completely personalized collection that reflects their own imagination, meaning no two players have exactly the same experience. Supporting that level of creativity and personalization simply wouldn't be feasible with a traditional content pipeline.

How do you think about creative control when AI systems are part of the player experience? Where do you want the AI to be surprising, and where does it need to stay firmly within designer-authored boundaries?

Piero Molino: It really depends on the type of game one is building. In the market and throughout the history of games, there have always been games where the authorial intent was more prominent and games where the toy-like qualities made the game fun.

I think a good analogy for how we are using AI systems is sandbox games. In those games, developers are providing the tools to players, and players end up combining them in the way they want to make their own fun. Designers author the rules of the experience; they provide boundaries within which the game keeps being consistent.

To achieve that, we give mechanisms for designers to provide AI with examples, structure, authored content, and clear guardrails so it can surprise players without losing the creative vision of the game. We want AI to be inventive in how it interprets player input, creates combinations, or introduces variation, while operating within boundaries intentionally designed by humans.

The sweet spot is when players feel like they're discovering something unexpected, while designers remain confident the experience is still consistent with the gameplay, tone, and world they've created.

The team is also exploring on-device AI.  What does that mean exactly? Why is that important for the kind of experiences you want to build, and what are the biggest technical or design challenges that come with it? 

Piero Molino: On-device AI means the models run directly on the player's hardware instead of sending requests to cloud servers.

That certainly brings technical advantages like lower latency, predictable performance, and better privacy. But for us, the bigger benefit is what it unlocks creatively. If every AI interaction has a cost attached to it, developers inevitably begin designing around that limitation. We think that limits the kinds of experiences you can build.

Running AI locally removes those economic constraints and allows developers to treat AI as a core gameplay system instead of something players can only occasionally interact with. The engineering challenge, of course, is making those models efficient enough to run across a wide range of hardware while still delivering high-quality results. That's one of the core problems the Atelico AI Engine is designed to solve.

AI-native systems can create a lot of unpredictability. How do you test, debug, and evaluate player-facing AI behavior when the output may not be the same every time?

Piero Molino: Rather than trying to eliminate variability, we design around it. Instead of evaluating whether the AI always produces the exact same output, we focus on whether every output stays within the game's creative and gameplay constraints. In practice this means evaluating systems in a statistical way, like you evaluate a machine learning system, or how you evaluate a clinical trial and aiming at statistical confidence because certainty may be unachievable. It’s an iterative process, where you get better at every repetition by looking for outliers, and gradually adjusting the aim.

Checklists are a great tool too: you define beforehand behavior that you would desire or prevent, and checking the AI systems against it gives you a baseline confidence that in the most obvious and egregious situations the system behaves as intended.

Finally, for preventing inappropriate behavior, red teaming is the best way to stress test the system.

Given the statistical nature of AI systems, often you need to find the best trade-off, and you won’t possibly identify all issues ahead of time, so you need to pair with best effort during development with quick reaction after shipping it.

What does the production pipeline look like when designers, writers, engineers, and artists are building around AI systems? How does that change collaboration compared to a more traditional game team structure?

Piero Molino: Let me give you an example from the development of the creature generation feature in Bobium Brawlers. We followed a very similar iterative process for the game’s deck generation system.

José Luna, the artist we are working with to define the visual style of the creatures, could not simply draw every possible creature in the game. Players can always come up with descriptions we have never seen before. Instead, he began by drawing a small set of around 20 creatures. We used those images to train our image generation model and integrated it into an early development build.

Once José could see the model’s output inside the game, he quickly identified problems. In some cases, the model overgeneralized. For example, it began generating every dragon as purple because the only dragon in the initial dataset was purple. In other cases, it failed to generalize enough. It did not consistently understand that every creature, including those described as inanimate objects, should have eyes.

Those results informed the next round of artwork. José created additional creatures specifically designed to give the model the examples it was missing, and we trained it again. The process continued when we released our first beta and playtesters uncovered new edge cases that we had not encountered internally. José has now drawn around 100 creatures, and he will probably create more after launch. The more people experiment with the system, the more unusual cases we discover.

In parallel, broader playtesting helped us evaluate content safety. We found that we could block a large portion of inappropriate requests, but not all of them. We established statistical metrics to track the balance. We wanted to minimize false negatives, meaning inappropriate content that slipped through, without creating too many false positives, where legitimate player requests were blocked because the safeguards had become too restrictive. As more people played, we collected additional data and developed a clearer understanding of where we wanted that boundary to sit. Systems like this always involve trade-offs, and exposing them to a wider range of players is often the only way to reveal their weaknesses.

More generally, developing AI-driven features requires closer and more frequent collaboration between creatives and engineers. It also requires continuous iteration, along with quantitative metrics and qualitative evaluation protocols that allow the team to measure whether the system is actually improving.

Looking back over the past year, what has worked better than expected, what has not worked, and how has your overall philosophy around AI-native game development evolved?

Piero Molino: Over the past year, our philosophy has become much more nuanced. Early on, we were focused on proving that AI could work inside games from a technical perspective. Today, we're much more interested in proving it can be used responsibly and creatively.

We've become increasingly convinced that success isn't measured by how much AI a game uses, but by whether players even think about the AI at all. The best AI-native games should simply feel like great games with mechanics that couldn't exist otherwise.

One thing that worked better than expected was how reliably and consistently we could make the AI system's output align with the designer's intent. We made the models very much steerable. That revealed one thing that did not work as we originally expected: our initial expectation that AI would provide variability in generated content, but when you impose strong steering, the AI system follows what you tell it too strictly.

We discovered the fundamental tension between control and variability, and we had to add more mechanisms to bring variability back. This is probably the hardest lesson we learned.

Piero Molino, Studio Atelico CEO and Co-Founder

Interview conducted by David Jagneaux