AI & Machine Learning

Sony received 10 patents spanning an unusually wide range of AI applications, several of which center on using machine learning to automate or personalize what has traditionally required manual effort. 3 of those filings deal directly with AI-driven playtesting: one describes a cloud pipeline that captures image frames from a human play session, extracts structured context like boss encounters, environmental settings, and item loadouts, and passes that data through an API so an AI agent can play and test the game in a later session; a second uses a large language model to read game source code, identify where API calls should be inserted, and write them automatically, eliminating the need for developers to hand-instrument each data hook; and a third trains an imitation learning-adjacent model on recorded gameplay to let an AI agent test content without requiring direct console instrumentation. On the personalization side, Sony filed patents for a recommendation system that pairs each suggestion with an explanation and then retrains on whether the player actually acted on it, an NPC generation system that trains on a player's recorded encounter history and outputs changes to NPC appearance, behavior, emotion, and role through either a live injection or an updated game build, and a kernel-function-based profiling system that generates scenarios across audio, visual, and gameplay styles and continuously steers game state toward whatever each individual player engages with most. 2 additional filings address how players and developers interact with systems through natural language: one translates a developer's plain-English description of a task into a matching resource from a studio repository along with usage guidance and bandwidth-aware delivery sequencing, while another maps a player's free-form expressed intent, such as wanting a more cinematic feel, to a concrete bundle of console and game settings drawn from patterns in historical user behavior. The remaining 2 patents cover a cognitive load scoring system that models mental effort per user and produces content rankings, duration suggestions, and hardware compatibility mappings, and a popularity prediction model trained on highly viewed videos that scores draft game scenes, storylines, and trailers and can identify specific low-scoring elements before release.

Nvidia received 1 patent describing an automated pipeline for generating labeled training data for game computer vision models. The system crawls game wikis and associated metadata, uses a multimodal large language model to label a seed set of images, and then expands that labeled set by matching visually similar objects across unlabeled image pools and raw gameplay video frames, removing the need for manual expert annotation at scale.

PostVoyage received 1 patent for a runtime system that places a large language model directly inside the game engine, where it reasons over per-NPC alignment vectors and accumulated interaction memory to rewrite the shared world state. Faction realignments, quest changes, and item evolution emerge from probabilistic reasoning over gameplay history rather than from designer-authored rules, and NPC states persist across sessions so one player's actions can shape the world another player encounters.

Tencent received 1 patent covering a speech recognition system for NPC commands that conditions its transcription on what the player character and nearby NPCs actually perceive in the virtual environment. By using in-game surroundings to disambiguate similar-sounding words, the system produces a recognized behavioral intention that better matches the current game context than a general-purpose speech-to-text model operating on audio alone.

EA received 1 patent for a level design system built around an imitation learning agent. The agent plays through a level, and its predicted state-action pairs drive decisions about where to place assets and how to reshape terrain; the same agent then replays the modified level to identify unreachable goals, closing a generate, test, and regenerate loop that replaces manual adjustment or static rule sets.

Microsoft received 1 patent describing a decision engine for NPC behavior that keeps every possible output within a developer-curated action list, preventing the AI model from generating free-form or unanticipated responses. The model's selection from that fixed list varies at runtime based on character memory, prior player interactions, and current world state, allowing NPC behavior to adapt over time without ever stepping outside the boundaries a developer has defined.