Translated by AI Translation Assistant

AI for Gamedev is essentially a software engineering problem—game development is perhaps one of the most complex fields in software engineering—this year, the rapid pace of Agent development has already made people worry about what will happen to jobs. As for art, it has already been hit by a wave of AI, and planners are now basically inseparable from AI. QA, due to the relatively high cost of vision models, still appears relatively safe for now.

And what exactly is AI for Gameplay—or AI Native games? The industry is still exploring and laying out plans. This article mainly discusses what AI Native can potentially do in Gameplay.

AI in Gameplay can be divided into three directions:

  • Generative: Generative AI that runs during gameplay. For example, the approach of using AI to enhance UGC experiences by the development team of the sequel to F.I.S.T.: Forged In Shadow Torch is essentially AI for DCC Tools, just running at runtime.
  • Automation: Automating some parts of the game, which could be the construction of UGC content or using LLMs to improve user experience. For instance, the AI NPCs in Honkai: Star Rail follow a path quite similar to AI Agent products on the internet.
  • Interactive: AI-driven narratives, AI-driven combat, AI real-time adaptation to player behavior… one can envision many scenarios, but execution is extremely difficult, and there are currently no successful products.

The technical paths for the first two directions are very clear—they are essentially extensions of AI functionality for DCC tools, or applications accumulated from LLMs. The Interactive direction, however, is not yet truly mastered by any company and is a key area of exploration for large studios in the future.

Interactive is a very broad concept. Almost all game content is interactive: quest lines are interactions, combat is interactions, and even customizing the game experience based on player preferences—such as the arachnophobia mode or accessibility features found in many games—are interactions.

Introducing AI into these interactive segments seems hugely promising: we could pursue AI-driven, personalized storylines for each player; we could have AI dynamically adjust character actions and poses at runtime to match player operations; we could even create non-preset equipment types at runtime based on player preferences.

However, real-world implementation faces many difficulties. Even in the text generation field where LLMs excel, there is currently no commercially successful game that truly integrates LLM capabilities into Gameplay. Something like AI Dungeon, with its pure text interaction, is indeed simple, but how is that different from talking directly to an LLM? Within a AAA narrative framework, how AI-generated text can be integrated into level design while preserving world consistency, character personality continuity, and quest logic coherence remains an unsolved puzzle.

Nevertheless, AI-driven changes to game content are just around the corner. It is foreseeable that within the next three years, games with built-in AI functionality will emerge in large numbers, and even “AI-driven Gameplay”—especially using AI to enhance storyline freedom—will be marketed as a core selling point.

Can AI Native really make games more fun?

Among the three paths of Generative, Automation, and Interactive, compared to their respective implementation difficulties, the truly fundamental question is: introducing AI, in the vast majority of cases, does not make games more fun.

Game mechanics and values are balanced against each other; no module can be too strong or too weak. Take Vampire Survivors-like games as an example: the numerical system and the mechanics built around it must be confined within a delicate range—players should neither be overly frustrated by high difficulty nor lose the desire for challenge due to it being too easy.

This means that game developers can hardly allow AI to make substantive modifications to these carefully tuned systems at runtime. What AI can do is basically perform some configurative work within an existing framework—which is not essentially different from a deterministic algorithm. We can imagine that if AI were given full control over game values, the difficulty curve would be directly shattered, and players would start engaging in a battle of wits with LLM prompts, which is not what developers want.

Presentation faces another problem: the visual effects and animations of successful games are mostly meticulously designed by artists. It is hard to imagine AI dynamically generating content at runtime that meets quality requirements—after all, even offline generation cannot achieve this now. However, in some rough indie or hyper-casual games, it might be acceptable, but this is far from bringing a revolution to the gaming industry.

On a deeper level, game design is not only about balancing numbers but also about “intentional experience design”: developers convey specific emotions and rhythms through carefully constructed rules and content. AI’s random generation may precisely disrupt this intentionality.

Unless you believe that in the future every pixel in a game will be generated by a model, but game worlds are often far more complex than reality. If that were achievable… then autonomous driving and brain-computer interfaces would probably no longer be problems.

The fun that AI-generated text can bring may not necessarily surpass that of a well-designed template plus a random event system. The dynamic content generated by simple templates and rules in a friend’s life is much more interesting than what an LLM produces. Randomly generated levels have existed for decades, and what truly makes them classics has never been the “randomness” itself, but the finely tuned random rules—essentially, the designer’s intent is still at play.

Where is the future of games?

The future of games probably does not lie in AI. After all, nearly 30 years have passed, and we are still playing The Legend of Zelda: Ocarina of Time. But the future of the gaming industry likely does not rest on single-player games. In 30 years, game prices have only increased by 10–20%, but the cost of games has at least multiplied tenfold.

On one hand, as players, we hope AI can speed up the entire production process, liberate game developers from chronic overwork, and reduce game development costs so that large single-player game studios can survive. On the other hand, as developers, we fear that AI’s development will ultimately lead to ourselves being optimized away, worsening an already cold industry.

AI makes the times change too fast; many things that seem distant for 2026 may arrive by 2028. But the classic works that have been passed down in the gaming industry will probably never go out of style.