AI & Machine Learning

Sony received 13 patents covering a wide range of AI-driven gaming technologies. Several patents address how games respond to players in real time: one system detects input errors before failure occurs and delivers proactive assistance based on how a player's button presses deviate from optimal sequences, while another uses machine learning to predict or delay commands so they sync perfectly with intended game moments. A third patent describes technology that infers cognitive state continuously from controller or headset sensors, then adjusts game content to match tiredness, stress, or engagement levels without requiring dedicated biometric equipment. Two patents focus on using biometric data more directly: one feeds heart rate and physiological signals into machine learning models that auto-generate or refine game segments in response to measured engagement, and another reads player and spectator emotions in real time to create sentiment-tagged gameplay fragments matched to available playtime. On the content creation side, Sony patented an AI image generation system that uses live game state data for semantic scene understanding, allowing players to modify visuals with text while preserving depth and 3D structure, plus a dynamic world generation system that uses nested templates and player preferences to populate environments in real time. The company also patented a system that clones individual playstyles into NPC squadmates by training models on specific players' gameplay data, letting users game with digital versions of friends or themselves. For development and moderation, one patent describes a quality assurance NPC that autonomously stress-tests other AI-generated characters in a simulated environment, while another monitors player inputs in real time to detect toxic behavior and apply graduated consequences automatically. Two additional patents address content beyond direct gameplay: an AI system that automatically generates sports highlight reels by computing an excitement score for each moment using shot event metadata, and a cognitive load-based evaluation system that matches games and media to users based on required mental effort. A final patent describes technology that turns smart home devices into gameplay output channels, using multi-modal machine learning and latency compensation to make lights, fans, and thermostats react to in-game events with contextually appropriate responses learned from user feedback.

EA received 3 patents focused on automating game development workflows. One patent describes a text-to-speech system that lets developers control NPC speech emotion and prosody using reference audio, achieving expressive character dialogue in a single unified architecture by separating content and style into distinct feature embeddings. Another applies machine learning to automatically fill in missing motion capture data, replacing the manual process of interpolating occluded frames with automated predictions based on movement patterns. The third patent covers visual bug detection technology that uses a large language model to analyze live video output frames in real time, catching coding errors during testing by observing the game as a black box through its display output rather than requiring engine integration or screen capture software.

Nvidia received 2 patents addressing animation and content generation. One patent describes a neural network system that learns realistic forces for animating objects by training on multi-aspect motion data, producing force-based control signals that address the brittleness and limited movement variety of earlier neural animation approaches. The other patent covers an AI system that auto-generates live commentary for sports, gaming, and events by feeding video frames into a language model pipeline, creating dynamic narration through a structured system that translates visual events into natural language without human commentators or manual scripting.

Hasbro received 2 patents related to AI-powered trivia gameplay. One patent describes a question "runway" system that pre-generates and queues trivia questions during gameplay, using predicted user interaction time as a trigger to eliminate wait times between rounds by managing a proactive question pipeline rather than responding reactively to each query. The other patent covers a multi-AI validation layer that uses independent models to fact-check and moderate generative AI responses in real time, injecting pre-loaded query context into validators that enforce factual and behavioral conditions through a consensus mechanism before trivia content reaches players.

Beijing Zitiao Network Technology received 1 patent for dynamic AI teammate chat timing that calculates when bot companions should send messages based on game state and player activity. The system replaces static timed intervals with adaptive behavior that prevents message spam during active player communication and prompts encouragement during lulls, making AI agents feel more human through context-aware communication period calculations.

AviaGames received 1 patent for an automated system that scores how much skill versus chance drives a game's outcomes, then feeds that classification into platform rules, provisioning, and anti-cheat systems. The mechanism uses a mixed effects model with decision thresholds set by outcome factor distributions from a training dataset, placing games on a skill-versus-chance spectrum automatically rather than relying on manual per-title judgment.

Cygames received 1 patent describing a method to train game AI on replay logs by weighting player history, converting game states and actions into text, and generating order variants to predict the next move. The system uses weighted text-order augmentation where history groups receive weights from user information, and each game state yields multiple reordered text samples in proportion to that weight, with randomized negative action pairs added to distinguish selected actions from merely selectable ones.

Disney Enterprises received 1 patent for AI that trains virtual character motion controllers using game controller signals rather than motion capture data alone. The system matches features extracted from player input control signals to pre-recorded motion sequences during training, ensuring the neural network encounters data distributions that mirror live inference conditions and resolving domain shift without requiring separate adaptation modules or motion capture dataset re-labeling.