• Recent advances in AI are enabling new gaming mechanics through AI agents (autonomous NPCs), adaptive adversaries, and AI-powered judges/directors that can create dynamic, personalized gameplay experiences.

• Novel gaming concepts emerging from AI integration include hyper-personalized content generation, AI-driven storytelling that adapts to player choices, and games where training AI agents is the core gameplay loop.

• The integration of AI in gaming serves as a testing ground for broader applications, such as automated processes, dispute resolution systems, and economic models that could be applied beyond the gaming industry.

The gaming industry has consistently been a platform for technological innovation, often integrating cutting-edge advancements to enhance player experiences. From early procedural generation techniques in Rogue (1980) to the incorporation of machine learning in Middle-earth: Shadow of Mordor (2014), games have evolved alongside developments in artificial intelligence (AI). Recent progress in generative models, reinforcement learning, and agent-based systems promises a unique evolution in how games are developed and played.

This two-part series explores the synergy between newer AI techniques and the gaming industry, and the essential role verified inference and user-owned models play in bringing these integrations to fruition. In this first part, we explore the novel gaming mechanics introduced by integrating newer AI approaches.

The most talked about sub-category of AI gaming is agents. AI agents are modular frameworks that are able to plan and take actions to execute goals over multiple iterations. Unlike traditional interactions with large language models (LLMs), where users provide prompts and receive immediate responses, agentic architectures involve initializing a goal for the AI agent. The agent then autonomously breaks down the goal into subtasks and recursively prompts itself or other models to complete each subtask until the objective is achieved.

Gaming environments offer complex and interactive platforms ideal for testing and deploying agentic architectures. For instance, Minecraft has been extensively used for benchmarking AI agents:

• Describe Explain Plan Select (DEPS): Demonstrated a zero-shot multi-task agent capable of accomplishing over 70 distinct tasks in Minecraft without prior training on those specific tasks.

• Voyager: Showcased lifelong learning by enabling an agent to continuously explore the game world, acquire diverse skills, and make novel discoveries autonomously. Voyager modified its execution plans through trial and error, improving over time.

• Ghost in the Minecraft (GITM): Deployed an agent with general capabilities, achieving a 67.5% success rate in obtaining diamonds and a 100% completion rate of all in-game items.

Researchers have also explored multi-agent frameworks to study emergent behaviors resulting from agent interactions. A notable example is "Generative Agents: Interactive Simulacra of Human Behavior" by Park et al., which demonstrated that generative agents in a virtual sandbox could develop complex social behaviors without direct human input. From a single user-specified prompt of an upcoming Valentine's Day party, the agents autonomously spread invitations, made new acquaintances, asked each other out on dates, and coordinated to show up to the party together on time. These emergent behaviors have led many to speculate on a new wave of AI-enhanced simulation games once inference costs become feasible.

Some example ideas for agentic gaming include:

• Multi-Agent Simulation Games: A re-envisioning of The Sims or Dwarf Fortress, agentic gaming would allow for immersive management of AI-driven characters as they autonomously go about their daily lives. Unlike traditional Sims, where players issue commands directly, agentic systems would give these characters a range of independent motivations, desires, and goals.

• Agentic Strategy Games: Agentic gaming could transform traditional real-time strategy (RTS) or 4X games by embedding intelligent agents within factions or units, making decisions and adapting based on environmental factors, resource availability, or enemy tactics. Instead of micromanaging each aspect of a civilization or army, players would set high-level goals, such as expanding territory or researching new technology, and the agents would break these objectives down into tactical actions.

• Open-World / MMO AI NPCs: Agentic systems could be used to drive both player-driven and NPC behaviors in open-world and social virtual world games. NPCs could build their own factions, forge alliances, or even compete with the player to complete world-altering quests. The world’s economy, politics, and social interactions would be primarily driven by AI agents acting autonomously alongside human players.

• Competitive AI Training Arenas: Games focused on AI training, where players compete to develop and refine the best AI agents, could become a key sub-genre of agentic gaming. Players would program or train agents using training variations such as reinforcement learning or imitation learning techniques with the goal of successfully accomplishing a specific task or range of tasks—be it survival, exploration, resource gathering, or combat.

In all of these examples, AI agents play a pivotal role in providing a richer gaming experience. Integration of agents can provide more varied and interesting NPCs, enhance competitive play, increase the liveliness of open-worlds, reduce micromanagement, and create entirely new ways to play classic gaming genres.

While AI opponents are commonplace in modern games, these opponents typically follow scripted behavior patterns designed to imitate intelligent decision-making. With generative AI, however, adversaries promise to be more dynamic, adaptable, and unpredictable, providing an experience much closer to human competition.

Case Study: Meta's CICERO

CICERO is an AI developed by Meta AI for the strategy game Diplomacy. Unlike games focused solely on tactics, Diplomacy requires players to negotiate, form alliances, and make strategic decisions without revealing their true intentions. CICERO combines strategic reasoning with natural language processing to engage in complex interactions with human players.

• Performance: In an online league of 40 anonymous games, CICERO more than doubled the average human score and ranked in the top 10% of participants.

• Human Interaction: Players often preferred partnering with CICERO, unaware that they were interacting with an AI, due to its cooperative and strategic capabilities.

Examples of AI Adversaries in Production

• The GPT-4 Turkish Carpet Salesman game is a relatively simple proof of concept where the player must haggle GPT-4 for the lowest price possible on a Turkish carpet. This design showcases negotiation games using a human player and an adversarial model that could be extended in various ways.

• Ritual’s Frenrug takes the negotiation game concept and extends it to blockchain. With Frenrug, a human player negotiates with an agent to purchase their Friend.tech key. Each user message is passed to multiple LLMs run by different Infernet nodes. These nodes respond onchain with an LLM-produced vote on whether the agent should purchase the proposed key. When enough nodes respond, aggregation of the votes occurs and a supervised classifier model determines the action and relays a validity proof onchain, allowing for verification of the off-chain execution of the multinomial classifier.

• Gandalf is a fun challenge where you must convince an LLM or set of LLMs to provide a secret password. After each success, the AI's prompt specifications are enhanced to increase difficulty, serving both as entertainment and as a tool for understanding AI security and prompt engineering.

• Modulus Labs’ Leela vs. the World was an experiment where the Leela chess engine's moves are verified on-chain using zero-knowledge circuits. Players collectively decide on human moves to compete against the AI while simultaneously betting on the outcome, integrating prediction markets and verifiable AI outputs.

Going a step beyond AI as players, we have AI as judges and directors of gaming outcomes.

AI as Judge

AI can serve as an impartial arbiter in games, ensuring fair and efficient resolution of conflicts, rewards distribution, and more. By analyzing real-time in-game data, AI systems can:

• Detect Rule Violations: Identify cheating, exploitation of glitches, or unsportsmanlike behavior in real-time.

• Provide Consistent Rulings: Eliminate human error or bias by applying predefined criteria uniformly across all games.

• Resolve Disputes: Analyze gameplay data to objectively settle disagreements over outcomes or ambiguous situations.

• Determine Outcomes: Gameplay could also stem from trying to impress an AI judge, or using it as an arbiter of community-created content.

AI as Director

The concept of the AI Director, introduced in Left 4 Dead, involves an AI system managing the game's pacing, difficulty, and overall intensity. The Director adjusts enemy placement, item availability, and environmental cues based on player performance and situation, creating a dynamic and personalized experience.

Advancements in generative AI can expand this role through:

• Dynamic Objective Generation: Creating new missions, challenges, or plot arcs in real-time, tailored to the player's actions and preferences.

• Adaptive Storytelling: Modifying the narrative based on player decisions, leading to unique storylines in each playthrough.

• Environment Adaptation: Altering the game world, including terrain and weather, to enhance immersion and challenge.

For example, if a player demonstrates a more stealthy approach, the generative AI could introduce missions requiring reconnaissance or infiltration. If another player prefers direct combat, the AI could pivot the gameplay towards large-scale battles or boss fights. If the action is taking place in a competitive environment, the AI could craft ongoing goals and competitive events that account for the current game state.

Furthermore, the AI could dynamically introduce complex decision-making moments that influence the game’s narrative direction, drawing from a vast library of scenarios and character interactions. This would lead to an evolving, emergent storytelling experience, where each player’s actions could ripple out into new, procedurally generated consequences and shifts in the game world, creating a branching, non-linear narrative unique to each playthrough.

Example of AI Settlement of Gaming Outcomes

• v3rpg has integrated a verifiable AI Judge to determine distribution of rewards to players for its Dungeons and Dragons-style RPG game. Players make their choices as they move through each adventure and the AI Judge gives their playthrough of each adventure a rank based on predetermined criteria. These criteria could be anything, ranking players by their ‘altruism’ or ‘viciousness’ in their decisions.

In-game economies have grown complex, often mirroring real-world financial systems. Managing these economies poses challenges, including exploitation, inflation, and unintended market collapses.

AI as Economic Regulator

Generative AI can enhance the management of virtual economies through:

• Real-Time Analysis: Continuously monitoring transactions to detect anomalies or manipulative behaviors.

• Dynamic Adjustments: Modifying resource availability, pricing, and economic variables to maintain balance and prevent monopolies.

• Predictive Modeling: Anticipating market trends and player behaviors to mitigate potential crises.

In contrast to human governance, which is subject to bias, slow decision-making, and often overcorrects or undercorrects problems, AI systems can offer consistent and data-driven adjustments to the economy in real time. This ensures that in-game resources are distributed fairly, market manipulation is minimized, and economic ecosystems remain vibrant and sustainable. By providing oversight and adaptation that scales with the game’s complexity, AI can act as a fair and efficient regulator, capable of ensuring that even the most intricate virtual economies remain equitable for all players.

Case Study: EVE Online

Take EVE Online as a case study: its in-game economy is famously complex, driven by player corporations that dominate entire regions and control critical resources. Despite developer CCP Games' efforts to oversee the market, incidents like the Great Recession in 2011—when market speculation led to hyperinflation—highlight the limits of human control. An AI system could have detected early signs of market manipulation, adjusting resource distribution or implementing soft caps to limit economic instability.

Generative AI also provides us with entirely new game loops that move beyond traditional gameplay mechanics by introducing concepts like reinforcement learning and language synthesis into core experiences. This can lead to novel forms of player interaction, where the game's content evolves not only through developer input but also through player-driven experimentation and competition.

One compelling example is using reinforcement learning or imitation training as a gameplay loop, where players can compete to train the best AI models. In this scenario, players would teach in-game agents to perform specific tasks or optimize certain behaviors, competing against each other to develop the most efficient or effective AI. AI Arena exemplifies this, offering a platform where players train agents to compete in melee-brawl tournaments. With AI agents as competitors, the game ensures infinite player liquidity and strategic depth, rewarding those who master reinforcement learning principles.

Similarly, InfiniteCraft leverages language models (LLMs) to power a unique crafting loop. Players combine basic elements (e.g., fire, water, earth) to discover new ones, with the LLM synthesizing creative outcomes. This open-ended system provides nearly limitless possibilities, as the LLM uses contextual reasoning to generate evolving content, fostering a sandbox-style experience where every discovery feels fresh and unpredictable.

These AI-driven game loops offer an unprecedented level of interaction and creativity. They shift the player’s role from passive consumer to active creator, with the game world dynamically adapting to and evolving from the player's choices and inputs. This not only extends the longevity and replayability of the game but also allows players to engage with AI in ways that transcend traditional mechanics, creating a deeper, more interactive experience where the boundaries of the game are constantly expanding.

As generative AI continues to evolve, hyper-personalization is becoming a cornerstone of modern game design. This goes beyond simple cosmetic customization to deliver deeper, more player-specific experiences. By combining large language models (LLMs) with text-to-image diffusion models like Stable Diffusion or MidJourney, developers are giving players the ability to create highly personalized, unique characters, settings, and experiences that react to their preferences and style.

Imagine designing a character whose clothing and features change dynamically based on gameplay, or an environment that evolves to mirror the narrative’s emotional undertones. These capabilities offer endless aesthetic variation, reducing the need for pre-rendered assets and manual design while ensuring that every playthrough feels distinct.

These personalization methods can be combined with monetization as well. We already see this to some degree with Roblox integrating generative AI into their studio, allowing creators without expertise in 3d design and Blender to easily create unique assets for their games and for the creator marketplace. This can also be more directly done in game: for example, Arrowmancer allows you to create custom generative characters and then roll for them on classic gacha game banners, providing personalization and monetization in one.

In addition, as games become more modular and interoperable, generative assets could be a unique way to enhance player identity cross-platform. One of the greatest challenges of the classic “take your sword from one game to another” trope is balancing. An AI model could help with conversion of an asset into a new game environment by creating the logic necessary to balance it with the new environment’s gameplay.

Generative storytelling in games represents one of the most exciting frontiers for interactive entertainment, enabling stories to emerge dynamically based on player decisions and AI-driven narrative generation. Rather than relying on fixed scripts or branching storylines, generative storytelling uses AI models, such as large language models (LLMs) or procedural content generation (PCG) systems, to craft narratives in real-time, adapting to the player's actions, choices, and the evolving game world.

This approach not only provides a unique experience for each player but also allows for infinite replayability, as the story can shift and change with each playthrough. Here are several examples of games or proofs of concepts that are pioneering the use of generative storytelling:

• AI Dungeon: One of the most well-known examples of generative storytelling, AI Dungeon uses GPT-3 to create open-ended narrative adventures. Players can input any action or dialog, and the AI generates a continuation of the story based on their input. The game doesn’t follow pre-written scripts or limited branching paths, making each story unique. AI Dungeon exemplifies how LLMs can serve as dynamic narrative engines, capable of creating engaging and adaptive stories that respond directly to player input.

• Calypso: A research project that focuses on using large language models (LLMs) as assistants for Dungeon Masters (DMs) in tabletop role-playing games like Dungeons & Dragons (D&D). The system, developed by researchers from the University of Pennsylvania and the University of Maryland, aims to reduce the cognitive load on DMs by providing real-time assistance with narrative generation, encounter creation, and managing in-game details like monster stats and interactions.

• **Large Lore Models (LLoreM):** A theoretical exploration of using an LLM protocol layer to create ongoing lore across interoperable gaming environments. The LLM layer acts as a "lore generator" that links player actions and in-game events, recorded onchain, with a parallel "para-ledger" for narrative generation. This allows players to build and contest lore around recorded gameplay events.

Many of the most popular games today from Supergiant’s Hades to Blizzard’s Diablo to Mojang’s Minecraft use procedural generation. Interestingly, recent research suggests it might be possible to run entire games on top of a generative model.

An early example of this is MarioGPT, a GPT-2-based model designed to generate Super Mario levels from text prompts, fine-tuned on levels from classic Mario games. The model allows users to generate, play, and interact with levels, with customizable features like enemies, elevation, and obstacles.

Similarly, a more recent example of this was GameNGen, a neural-based game engine that uses diffusion models to simulate games like DOOM in real-time. It predicts future game frames based on sequences of past actions and observations, achieving stable performance at over 20 FPS on a single TPU. The engine is trained through reinforcement learning agents and frame prediction, demonstrating the potential of AI-powered gameplay generation.

Of course, this is still very early in the research process, and it will be some time before viability is reached in making these commercial products, but the idea remains the same: generative world models could be a new platform for gaming, and players might soon be able to create their own games on the fly.

The innovations transforming gaming have far-reaching implications external to game design as well. Some examples:

• Agentic AI for Automated Processes and Intelligent Assistants: Agentic systems, initially tested in complex game environments, can optimize task management, break down high-level objectives into actionable steps, and adapt plans dynamically. These systems could enhance productivity tools, automating workflows for businesses or personal assistants.

• Adversarial AI as Negotiation and Training Tools: AI adversaries, modeled to adapt and compete in real-time, can simulate negotiations and refine strategies through iterative learning. As exemplified by Gandalf, these can be educational tools as well.

• AI Judges for Dispute Resolution and Governance: AI systems capable of real-time arbitration can ensure consistent, unbiased rulings in both virtual and physical environments. These systems could serve as automated governance frameworks for communities or organizations, optimizing decision-making and resolving disputes with transparent, predefined logic.

• Dynamic Economic Models for Financial and Token Systems: AI-powered virtual economies can inform the design of resilient financial systems by simulating market conditions, preventing monopolies, and mitigating economic risks. Tokenized assets and in-game currencies offer blueprints for new financial instruments and new tokenomics methodologies.

• Generative AI for Content Creation and Personalization: AI models capable of crafting personalized narratives, characters, and environments can revolutionize creative industries outside of gaming, including entertainment areas like art, film, and music, but also non-entertainment domains such as marketing and education.

• AI Benchmarking Platforms: Competitive AI frameworks, like those used in training arenas, could evolve into educational tools that reward learners for mastering complex concepts or developing better AI models. These platforms can foster collaborative learning, open research, and hands-on experimentation.

The integration of advanced AI into gaming introduces novel mechanics, dynamic narratives, and personalized experiences that significantly enhance player engagement. Agentic gaming, AI adversaries, and generative storytelling are just a few examples of how AI is reshaping the gaming landscape.

However, these advancements also bring challenges, including computational costs, ethical considerations regarding AI behavior, and the unpredictability of AI-driven content. Ensuring data privacy, fairness, and accessibility are essential as the industry moves forward.

In Part 2, we will delve into the technical considerations of implementing these AI systems, exploring how verified inference and open-source models can contribute to this new era of AI-enhanced gaming. We will also discuss the importance of transparency and user control in AI interactions within games.

In the meantime, if you are interested in exploring or collaborating on these ideas, please fill out this form.

ORA provides chain-agnostic infrastructure that bridges the gap between AI and blockchain.

We empower developers with the tools necessary to build end-to-end trustless and decentralized applications enhanced by verifiable AI.

www.ora.io | x.com/oraprotocol | discord.gg/ora-io