AI & Machine Learning

Sony received 6 patents covering a broad set of AI systems that connect player physiology, behavior, and identity to game content generation. 2 of these patents are rooted in biometric data: one feeds real-time physiological signals like heart rate directly into machine learning models that auto-generate or refine game segments based on measured engagement, while another reads cognitive state through existing controller and headset sensors to continuously adjust game content in response to tiredness, stress, boredom, or focus, using federated learning to build personalized models without sending raw data off-device. A third patent extends sentiment analysis further, combining multimodal biometric and behavioral data with game state information to generate emotionally-tagged, replayable gameplay fragments matched to a player's mood and available time window. On the social side, Sony also received a patent for an AI that replicates a specific player's in-game behavior as an NPC companion, with the ability to blend traits from multiple real players into a single character at adjustable ratios, and a separate patent for a real-time toxic behavior detection system that analyzes player inputs and game state to apply graduated in-game consequences automatically, without waiting for post-game reports or human moderators. The sixth patent describes a game world generation system built on nested template hierarchies, where multiple layers of structure constrain a generative model simultaneously while player preference data shapes how environments are ultimately populated.

Hasbro received 2 patents, both aimed at improving the reliability and pacing of AI-generated trivia gameplay. One addresses the latency problem inherent in large language models by introducing a pre-generation queue: the system predicts how long a player will spend on the current question and uses that window to generate and hold the next question in reserve, so players never wait between rounds. The other patent places a layer of independent AI validators between the generative model and the player, each pre-loaded with context about the rules and factual standards the game must meet, creating a consensus-based check that catches hallucinations and content violations before they reach the user.

Among the single-patent filers, Cygames received 1 patent for a method of training game AI on historical replay data by converting game states and player actions into text that transformer-style language models can process. The system assigns weights to different portions of a player's history, generates multiple reordered versions of each game state as training samples in proportion to those weights, and pairs selected actions with negative examples of merely available actions, allowing the model to learn meaningful move prediction rather than generic action selection.

AviaGames received 1 patent for a system that automatically measures how much a game's outcomes are determined by player skill versus chance. Rather than relying on manual, per-title evaluation, the system uses a mixed effects model whose classification thresholds are derived from a training dataset of outcome factor distributions, placing any given game on a skill-to-chance spectrum without human judgment, then feeding that classification directly into platform-level rules, provisioning logic, and anti-cheat mechanisms.