Building a Game with GenAI

vs. Building a Game Using GenAI

ℹ️ How far can generative AI go in games?

One of my roles is to identify and share practical AI use cases that truly help game studios. and organizing conferences where real-world applications are shown.

But every time I prepare for these events, I run into the same challenge:

“How can we find use cases that are publicly shareable and genuinely valuable from a game company’s perspective?”

This is particularly difficult with generative AI, where many projects remain confidential, or only a few well-known examples are repeatedly shared. A more cautious and strategic approach is required.

Recently, I came across an intriguing experiment involving generative AI:

A small team of just eight developers at Jam & Tea Studios launched a game called Retail Mage in only six months.

ℹ️ Can AI become a Game Master?

Retail Mage isn’t just about generating content automatically. The team asked a more ambitious question:

“Can AI understand context like a human and respond to players in real time?”

For example, if a player says, “I’m bored,” an NPC might suggest, “Wanna play hide-and-seek?” and other nearby NPCs will voluntarily join the event. This all happens without a pre-written script—AI interprets the context and generates real-time interactions.

This led the team to a bold conclusion:

“Generative AI can go beyond being a tool—it can become a true Game Master.”

ℹ️ Experimenting with AI-Driven Game Design

The core of Retail Mage lies in improvisation. No matter what the player does, the AI understands their intent and dynamically adjusts the game world.

Instead of following fixed rules, the AI interprets the situation and decides how to guide the game flow. For example, if a player says, “I’ll give you an IOU,” the AI evaluates whether that’s a reasonable solution based on the game’s setting and character states.

This isn’t just a trigger-based interaction—it’s true context-driven design.

Even character creation is handled by AI. With just a few keywords, it generates a full personality, background, and even retains memory. A character tagged “brave” may be approached for help during in-game crises, making the interaction feel more natural.

Dialogues are also free form. There are no preset options. Whether the player types or speaks, the AI understands tone and context and responds naturally. The only thing the developer sets is the character’s personality and scenario, everything else is handled by AI.

Set character traits and context as AI generates dialogue and triggers APIs in real time

The team wasn’t building a game that followed a fixed script, but one that reacts to the player’s imagination.

ℹ️ The Dilemma of AI-Driven Games

Although the game gained attention for its innovative AI-driven approach and launched on Steam in just six months, the team reflected on a critical gap—it felt like something in-between a demo and a full-fledged game.

They pointed out three major challenges:

1. Lack of Familiar UI and Game Systems

Because of its AI-first design, many standard UI elements were omitted. The game relied heavily on text input rather than buttons, creating friction for players familiar with conventional systems. This even became a barrier to entry for some users.

2. Too “Human-like” AI

Unexpected problems arose when NPCs behaved too independently. Some resolved issues on their own and left the scene. Others ignored the player entirely, making the player feel like an observer rather than the protagonist.

3. Technical Challenge: Real-time Inference

“Inference” in generative AI refers to deriving appropriate actions or dialogue based on given input. This happens every time a player interacts with the game. The AI breaks down the input into “tokens” and generates a suitable response.

Each sentence is broken into dozens of tokens, and each token contributes to the processing load.

GenAI processes language as tokens not words, often turning a single sentence into many tokens

The first technical challenge was speed:

Game engines run at 60 frames per second. But AI processes each input like a human—hearing, thinking, then responding. This created a lag, where the game progressed faster than the AI could keep up. The team had to re-tune the game rhythm to match AI capabilities.

The bigger challenge was cost:

Every interaction requires a server call to a large language model (LLM). The more tokens are used and the larger the model, the higher the cost.

While the team optimized performance using open-source tools and custom models, the issue wasn’t fully resolved.

These challenges are not unique to Retail Mage. They are fundamental to making a game with generative AI at its core.

ℹ️ A Meaningful Experiment—But…

The team proved their belief that “Games are creative spaces, and AI is a tool to expand that space.”

But whether that experience translated into fun for the player is another story.

While AI can be highly creative, not every player seeks creativity all the time.

An overly open system may lead to uncertainty or fatigue. In some cases, it can even feel like the game is demanding players to “learn” how to play, instead of just enjoying it.

Those risks undermining the basic premise of games as a service.

In the business of games, we often go back to two foundational truths:

“Games are a service.”

“The user is always right.”

This case reminded us why those truths still matter.

♻️ Building a Game With AI, or Using AI In Games?

Creating a game built around generative AI is bold and ambitious.

But this case raises a more important question:

“Are we building games around AI, or are we using AI to make better games?”

Innovation in technology is essential in the gaming industry. But what matters more is whether that innovation improves the player’s experience.

Because at the heart of any successful game—no matter the tech—there must always be the player.

The next challenge isn’t just about adopting AI. It’s about designing better services using AI.

※ Disclaimer: This content reflects the author’s personal views and includes only publicly available examples. It does not represent the official position of any company mentioned