From Predictable NPCs to Characters that Improvise
For decades, artificial intelligence (AI) in video games has controlled the behavior of enemies or allies by following rigid scripts or highly defined programs. However, with generative language models similar to ChatGPT, non-playable characters (NPCs) can now converse and interact much more freely and naturally.
A recent example is “Retail Mage” (2024), a role-playing game (RPG) where players manage a magical shop with NPCs driven by AI-generated improvisation. Instead of sticking to pre-defined phrases, players can type any query or response, and the NPC will generate coherent dialogue in real-time. This allows for unexpected conversations, humorous or creative situations: developers report that possibilities and experiences multiply when there are no exhausted dialogue lines or easily repeatable ones.
“Mecha BREAK” (2025), a multiplayer shooter game, showcased a pilot version at Gamescom 2024 with an NPC that spoke and interacted using AI. Players could freely ask the character about upcoming missions, receiving tactical advice generated on the spot.
Worlds Created on the Fly
AI generative capabilities not only give improvised voice to characters but can also invent stories, missions, and entire worlds as players progress.
“AI Roguelite” (2023) is a notable case: it’s presented as “the first text-based role-playing game where AI determines 100% of each location, enemy, object, and mechanic.” In every playthrough, the game generates descriptions, scenarios, events, and even unique images and music using AI models. Players praise the almost infinite flexibility of this approach: one can experience classic epic tales of knights versus dragons or surreal adventures, depending on user inputs and the model’s creativity.
There are also hybrid projects that combine handcrafted content with AI generation. For example, “Nyric” allows creating 3D sandbox worlds (open-world virtual environments where players have freedom to explore, build, and modify without a predefined storyline) from simple text descriptions, using AI to fill them in and adapt them on the fly. Imagine writing “an enchanted forest under a starry night” and watching the game construct an immersive landscape to explore.
“inZOI” uses AI to generate unique textures and objects based on user instructions, employing small language models to give NPCs more “psychological depth.” In other words, these characters not only have different dialogues in each game but also “think” and act with some autonomy, simulating more complex motivations.
Games that Feel
Another revolutionary aspect is affective AI, which measures and responds to players’ emotions. The premise is simple: if gaming is an emotional experience, why not make the game react to how we feel?
“Nevermind” (2015), a psychological horror game, is a pioneer in this field. Using biofeedback, the system detects a player’s stress and fear through a heart rate sensor or camera. If it notices the player is too calm, it increases difficulty and scares to induce pressure. If it detects panic, the game can ease intensity. Essentially, the game observes and adapts the experience to maintain players in that optimal zone between boredom and anxiety, known as “flow.”
The Player at the Center: Customized Difficulty and Experiences
Beyond emotions, AI is enabling games to adapt to each player’s preferences and playstyle in unprecedented ways. For years, we’ve seen basic difficulty adjustment systems: from choosing easy/hard modes to “AI Director” in “Left 4 Dead,” which modulated enemy intensity based on player performance.
The new generation of AI takes this much further. For example, “MIR5,” an upcoming action RPG, announced that it will feature “final bosses” controlled by AI. Final bosses are the primary enemies of a level or the game’s narrative. These characters are designed to be significant challenges and often mark key moments in the story. Thanks to AI, they will learn and adapt to a player’s tactics, ensuring each encounter is personalized and always a challenge.
Personalized content generation is another growing area. If a player tends to explore extensively and converse with all characters, AI could detect this pattern and generate more narrative missions; if another prefers quick action, the game could offer additional combats or skip lengthy dialogues.
Various academic studies have long worked on player modeling (identifying these preferences and optimal engagement states) to tailor experiences. The difference now is that the game engine can dynamically reconfigure itself thanks to AI generative tools.
In the future, player models can be adjusted and not remain static or focused on a limited number of profiles.
