AI & Machine Learning

Microsoft received 7 patents covering a broad range of AI-assisted gaming experiences. Three of them cluster around in-game player assistance: one automatically detects when players are struggling by analyzing aggregate gameplay data across populations and delivers proactive help before players even think to ask for it, another trains machine learning models on expert gameplay footage to create AI assistants that can actively play through difficult sections in real time on a player's behalf, and a third restricts what actions a human coach can perform during a co-op session based on the current game state, so helpers can guide without spoiling puzzles or bypassing intended challenges. On the creative and social side, Microsoft also filed 2 patents for real-time voice transformation in role-playing games: one automatically matches a player's voice to their character's physical appearance using computer vision and voice blending, and another dynamically adjusts that synthesized voice every frame based on in-game conditions like running, being underwater, or other character states. The remaining 2 patents venture into visual and physical animation, with one replacing traditional inverse-kinematics solvers with a neural network trained on human hand biomechanics to generate lifelike hand poses for games, VR, and robotics training data, and another using a statistical prior model trained on human appearance to constrain a Gaussian splatting system, producing photorealistic avatars from sparse input data for video conferencing and virtual environments.

Bandai Namco received 2 patents, both centered on making AI-controlled characters feel less mechanical. The first moves beyond the standard practice of copying a single player's behavior to control an NPC, instead blending movement and decision patterns from multiple players using weighted combinations, which produces far more variety and unpredictability than any single behavioral sample could. The second automates avatar creation entirely, pulling from stored user profile data alongside live text, image, and voice inputs to generate multiple personalized character options simultaneously, removing the need for players to manually navigate customization menus.

Sony's 41 patents span an unusually wide range of AI applications. Several filings address how games communicate with players through language and voice: an LLM-powered virtual teammate delivers real-time strategy advice and health monitoring through in-character voices, a voice-activated assistant lets players issue complex team commands that the system interprets using live game engine data, and a separate system allows players to ask in-game characters for story recaps or navigation tips that adapt to their individual progress. Sony also filed 3 patents around personalized audio content, with 2 covering AI-generated podcasts narrated in the voices of players' favorite game characters to deliver recommendations and gaming news, and a third that learns individual sound preferences over time to automatically mute, modify, or replace in-game audio during multiplayer sessions. A cluster of patents targets player assistance and adaptive difficulty: one detects frustration patterns to provide haptic-guided help through controller vibrations while tracking assisted and unassisted achievements separately, another tailors difficulty specifically to a player's behavioral style and uses emotional state detection to ease off before frustration tips into disengagement, and a third passively predicts what questions a player might have by analyzing live video frames and answering them without being asked. Multiple filings address how games recap and onboard returning players, including systems that use AI to curate story moments procedurally, generate personalized tutorials based on why a player stopped playing, and strip games down to essential narrative beats for time-pressed sessions. Sony's NPC patents are also numerous, covering a system that blends multiple players' operational data to train single tunable agents via behavioral style sliders, a bidirectional NPC dialogue system where action results feed back into the language model to prevent hallucinated responses, a system where NPCs listen to real co-op player conversations to generate timely contextual dialogue, one that adapts NPC responses across sessions by recognizing player behavioral patterns, and another that delivers real-world notifications like texts and emails through NPCs at contextually appropriate gameplay moments. On the content creation side, Sony filed patents for a text-to-character system that generates avatars from plain chat descriptions, an AI pipeline that converts gameplay telemetry into storybooks, highlight reels, and memes, a system that auto-generates smooth transitional video content between gameplay moments using a generator-discriminator feedback loop, a content generation pipeline where designer keyframes act as hard constraints on generative model output, and a NeRF-based system that produces personalized 3D in-game items in under 2 minutes from player data. Additional filings cover AI-driven broadcast production that automates camera angles and live narration for esports, a machine learning system that infers which controller buttons were pressed just from watching video footage, an audio description system that generates and inserts narration into content gaps for accessibility, a route-aware storytelling system that uses navigation data and window displays to create interactive in-car narratives, a crowd participation system where stadium cameras capture thousands of audience members' gestures to collectively control games on large displays, a system that monitors engagement signals to indicate through wearable lights whether a player is available to be interrupted, an eye-tracking system that suggests player-to-player chat based on gaze and game context, a multimodal sensor fusion system that decodes player intent from camera, IMU, and microphone data, an ML pipeline that reconstructs 3D overlays of popular player routes from 2D gameplay video, a system that analyzes gameplay clips and viewer comments to auto-generate game quality reports and trigger code changes, a closed-loop QA pipeline that detects performance issues, diagnoses them with machine learning, applies fixes, and commits them to version control, a pitch-control system for game dialogue that uses binned phoneme representation to adjust character voice without re-recording actors, and an AI system that personalizes trophy appearances based on how and in what context a player earned them.

Nvidia filed 5 patents addressing distinct gaps in AI-assisted gaming infrastructure. One enables large language models to process multichannel spatial audio by converting it to a fixed B-format representation, allowing NPCs to realistically respond to the direction sounds come from rather than just their content. A second decouples image retrieval from language model inference by pre-indexing text-to-image associations and using vector similarity search, letting chatbots and NPCs insert relevant images into responses without requiring expensive multimodal model architectures. A third uses natural language processing on game-generated text logs to extract gameplay events for highlight reels and activity feeds, avoiding the computational cost of analyzing video frame by frame. The fourth applies real-time neural network inference to filter abusive language during live voice and text chat, and the fifth introduces an auto-regressive auto-encoder that tokenizes 3D mesh faces more efficiently, enabling generation of meshes with more than 8,000 faces at artist-quality topology.

EA received 4 patents covering a range of AI-driven improvements to gameplay and accessibility. One creates lightweight "difference models" that capture only the idiosyncratic movement deviations of real athletes, allowing game characters to move like specific individuals without storing full volumetric capture data, and the system can even apply those movement signatures to athletes for whom no direct capture footage exists. Another uses geometric feature extraction from the virtual environment itself to determine when and how to assist players, making help contextually aware of the spatial challenges a player actually faces rather than relying on performance metrics alone. A third replaces human guides and inadequate screen readers with an AI model that converts live game state data into audio descriptions including character positions, distances, and environmental context for visually impaired players. The fourth uses a ring-based assignment algorithm to holistically match multiple characters to multiple objectives simultaneously in real time, avoiding the computational expense of brute-force coordination approaches.

Adeia received 2 patents, both using AI to manage game pacing and character behavior at a systemic level. The first embeds time-window constraints directly into a Monte Carlo Tree Search algorithm, allowing the AI opponent to simultaneously optimize for winning and for fitting the session into a user-defined duration without switching models or degrading difficulty mid-session. The second uses a graph-based system to propagate personality traits across all NPCs based on their relationship strengths, then generates their dialogue using large language models while constraining responses to what each character would plausibly know given player progression, scaling interactive characters far beyond what manual scripting allows.

Wizards of the Coast received 1 patent for an AI assistant that maintains a persistent profile of each player across sessions, accumulating knowledge about their preferences, skill level, and play patterns over time. Rather than starting from scratch each session, the system draws on this history to provide increasingly personalized guidance, rules explanations, and narrative suggestions during both tabletop and digital game play.

Roblox received 2 patents targeting opposite ends of the game creation and integrity pipeline. One applies diffusion model generative AI to avatar creation, allowing players and developers to describe characters in plain text prompts and receive fully generated models, replacing the rigid slider-based customization tools that previously required manual configuration. The other moves cheat detection from the client side to server-side machine learning inference, where models analyze game state data and physics violations in real time across multiple cheat types simultaneously and can also automatically generate cheat-resistant game scripts using language model integration.

NetEase received 1 patent for an AI system designed for virtual sports games that automatically identifies optimal attacking zones and matches virtual player objects to those positions, shifting the game engine from a passive stat-enhancement model to an active spatial coordination system that pairs players with opportunities rather than relying solely on user input.

Truist Bank received 1 patent for a player re-engagement system that combines in-game behavioral signals like session frequency and gameplay duration with real-world financial transaction data from a partner banking application to train churn prediction models, fusing cross-domain data in a way that departs from purely game-data-driven approaches.

Binary Kinetics received 1 patent for a stateless AI Game Master architecture that reconstructs conversational context from compact transaction logs rather than storing full session history, cutting per-session memory requirements by roughly 99%. The system also supports automatic failover across multiple LLM providers, and game rule enforcement is handled statelessly, creating an architecture designed for scalable AI-driven interactive fiction without the memory overhead traditional approaches require.

Meaning Machine received 1 patent for a system that allows NPCs in multiplayer games to improvise dialogue and then automatically merges that newly created lore back into the master world context database using semantic vector comparison. Unlike retrieval-augmented generation systems that only query static knowledge, this approach keeps the shared world consistent across concurrent sessions even as characters generate original content in real time.

Playtika received 1 patent for a real-time player segmentation system that decouples the segmentation logic from the underlying game architecture, allowing operators to define granular, behavior-driven player groups and deploy differentiated experiences rapidly without requiring deep code changes. The system goes beyond static difficulty settings by enabling temporary or permanent experience modifications at scale across large player populations.

GDM Holding received 2 patents. The first addresses multi-agent reinforcement learning matchmaking, assigning each learner in a training pool its own distinct distribution over opponents rather than using uniform self-play, which prevents the pool from collapsing to a single dominant strategy and allows collective performance to ratchet upward systematically. The second is a companion agent that observes a game's video stream from outside the game itself, requiring no developer integration, and combines proactive video sampling, session memory, and tool-use within a single ML pipeline to answer player questions and overlay visual hints directly on screen.

Google received 2 patents in this period. One is an AI coaching system that processes actual gameplay video alongside natural language dialogue, allowing players to ask follow-up questions and receive chain-of-thought explanations for why specific moves or strategies are recommended, rather than receiving opaque action outputs. The other detects in real time where a player is stuck, searches for existing tutorial content, and if none exists, automatically generates new tutorial material from the player's own captured gameplay footage, switching between play, search, and creation modes as needed.

Atombeam Technologies received 1 patent for a Persistent Cognitive Machine architecture that gives AI agents continuous thought processes independent of external prompts, biologically-inspired sleep cycles for memory consolidation, and the ability to retain cognitive state across system restarts, moving beyond the standard prompt-response model toward AI teammates that accumulate experience over time.

AMD received 1 patent for a hierarchical reinforcement learning framework in which experienced "leader" NPCs train "follower" NPCs through direct demonstration, rather than training every NPC from scratch independently. This structure reduces the computational resources required while allowing the system to produce NPCs across a range of skill levels and roles.

Triumph Labs received 1 patent for a matchmaking system that dynamically adjusts its own matching parameters based on real-time analysis of player availability, preemptively loosening or tightening skill-based criteria before wait times become long enough to affect retention, rather than applying fixed thresholds regardless of conditions.

Nintendo received 1 patent for a companion NPC system that automatically detects when a player character boards a moving platform or object and repositions allied NPCs accordingly, placing them in the correct state for cooperative attacks without requiring the player to manually re-pair with companions after the transition.

Tencent received 4 patents. One enables players to generate custom in-game clothing by entering keywords, with the system using existing garment technical maps as templates to produce new designs that remain compatible with the game engine. A second separates NPC behavior control into individual and group layers that communicate through status parameters, allowing complex and dynamic AI responses to player actions without the computational overhead of unified control systems. A third replaces manual playtesting for user-generated levels with an AI-driven pathfinding system that tests all routes simultaneously, categorizes specific error types, and offers one-click geometric fixes. The fourth automates NPC combat decision-making using reinforcement learning combined with LSTM networks that process both real-time state data and historical context, replacing manual behavior tree programming.

AT&T received 1 patent for a content filtering system that goes beyond age-gate blocking by dynamically modifying game visuals, audio, and gameplay mechanics in real time based on individual user profiles and the severity of detected mature content, adapting the experience rather than simply restricting access.

Acco Brands received 1 patent for a game controller with an integrated AI assistant that uses a built-in camera to visually identify the game being played, listens to gameplay audio, interprets natural language commands from the player, and provides real-time strategic guidance, including the ability to execute complex in-game actions through voice input.

Lemon received 2 patents both focused on lowering the barrier to game creation. The first combines a large language model for code and parameter generation with a diffusion model for visual asset generation inside a unified chat-driven pipeline, allowing iterative adjustments to both logic and art in real time through conversational input. The second addresses the visual consistency problem that arises when assets are generated independently, by first deriving a unified style description from a single user prompt and then applying it across all characters, backgrounds, and environments before any images are generated.