AI & Machine Learning

Sony received 7 patents covering a broad range of AI-driven systems that touch nearly every layer of the gaming experience. 2 of those patents deal directly with how computing resources and controller inputs are handled at the platform level: one uses a trained AI model to predict a player's skill level from gameplay behavior and pre-allocate computing power before it's needed, while another detects physical limitations from a player's actuation patterns and automatically compensates for inputs the user cannot physically complete, removing the need for manual accessibility configuration. A third patent operates in a similar accessibility space but targets visuals and audio, applying a platform-level AI engine that transforms game output in real time to meet a player's needs without pausing gameplay or requiring any involvement from the original developer. On the social and competitive side, Sony's ML-powered matchmaking system gathers both behavioral data and explicit player preferences to pair users with compatible teammates and opponents, and a separate patent allows real players' movement styles and tendencies to be extracted from recorded gameplay and embedded into NPC companions, so solo players can be joined by AI partners that reflect how an actual human plays. Sony also filed a patent for a video search system that indexes gameplay footage using AI-generated metadata drawn from live game signals, such as controller input speed and completion times, making it possible to find clips by describing in-game scenarios rather than relying on manual tags. Rounding out the 7 is a closed-loop coaching system that analyzes both controller inputs and gameplay video to deliver real-time instructions and generate custom practice minigames tailored to a specific opponent's weaknesses.

Adeia received 1 patent for a coaching tool that pulls a skilled player's AI-trained behavioral model directly into a live session, allowing it to offer real-time, context-aware guidance that adjusts as the player deviates from recommended actions. The system can optionally deliver that advice in the skilled player's own voice using generative AI, adding a layer of authenticity that distinguishes it from static tutorials or recorded walkthroughs.

SoftBank Group received 1 patent for a mobile system that lets fans point a phone camera at a live sports game and receive commentary from a personalized virtual character. The character reads the viewer's emotional state, adapts its commentary style accordingly, and is animated in sync with game events using a multi-sport motion database, with generative AI powering the interactive dialogue.

SafeBear received 1 patent for a cyberbullying detection system designed for social networks and video games that runs multiple specialized AI models in parallel, each focused on a distinct category of toxic behavior, rather than relying on a single general-purpose classifier. When harmful content is detected, the system masks it in real time and records certified evidence on a blockchain, creating an end-to-end workflow that connects detection directly to documentation suitable for legal proceedings.

Bitpart AI received 1 patent for a multi-agent system that allows game designers to author NPC behavior at scale without scripting each character individually. The pipeline converts human-written narrative transcripts into a hierarchical task network of reusable, recomposable plans, and a runtime director draws from that network to sequence and coordinate behaviors across multiple NPCs simultaneously, enabling dynamic interactions between characters and players without relying on fixed scripts.

Nvidia received 1 patent for a QA system that uses vision-language models to detect bugs in gameplay video and audio automatically, without requiring pre-defined rules for each bug type. To keep the process computationally manageable, the system selects keyframes based on embedding dissimilarity scores, focusing analysis on moments that represent meaningful changes in game state rather than processing every frame.