Artificial Intelligence (AI) is fundamentally reshaping the landscape of gaming, serving as a superhuman competitor, a novel development tool, a sophisticated in-game agent, and a fertile training ground for advancing AI research itself. The strategic domain has been a key crucible for these advancements, with deep reinforcement learning (DRL) models like Google DeepMind's AlphaGo, AlphaGo Zero, and MuZero achieving superhuman performance in complex games such as Go, Chess, and Shogi. These models, often tested in environments like Atari games , evolved from requiring human expert data to learning purely from self-play, and ultimately to mastering games without any prior knowledge of their rules. The introduction of these superhuman AI-Powered Go Programs (APGs) provides a unique lens to study human-AI learning; analysis of over 749,000 professional moves reveals that human decision-making quality significantly improved post-APG release. Players demonstrated a genuine learning effect, showing higher alignment with AI's suggestions and markedly reducing the number and magnitude of errors, especially in the highly uncertain, early phases of the game. This instructional impact varies, with younger and less-skilled players showing the greatest gains. Beyond playing, AI is transforming game creation through Procedural Content Generation via Machine Learning (PCGML). To overcome challenges of controllability and data scarcity, a novel "distillation" method uses LLMs to synthetically label content from traditional PCG algorithms, creating large-scale datasets for training steerable, text-conditioned generative models for a "Text-to-game-Map" task. Furthermore, AI is being integrated as a platform within games, such as in tactical wargaming experiments with fully autonomous Robotic Combat Vehicles (RCVs). These wargames reveal critical tactical implications, highlighting the vulnerabilities of remotely operated systems to jamming and the human tendency to employ autonomous RCVs as expendable "bait" , thereby helping operators and engineers co-develop realistic requirements for future AI-enabled systems
Paper
Full text
A Comprehensive Review of the Use of Artificial Intelligence in Gaming
Semantic Scholar · 2025
Abstract
Artificial Intelligence (AI) is fundamentally reshaping the landscape of gaming, serving as a superhuman competitor, a novel development tool, a sophisticated in-game agent, and a fertile training ground for advancing AI research itself. The strategic domain has been a key crucible for these advancements, with deep reinforcement learning (DRL) models like Google DeepMind's AlphaGo, AlphaGo Zero, and MuZero achieving superhuman performance in complex games such as Go, Chess, and Shogi. These models, often tested in environments like Atari games , evolved from requiring human expert data to learning purely from self-play, and ultimately to mastering games without any prior knowledge of their rules. The introduction of these superhuman AI-Powered Go Programs (APGs) provides a unique lens to study human-AI learning; analysis of over 749,000 professional moves reveals that human decision-making quality significantly improved post-APG release. Players demonstrated a genuine learning effect, showing higher alignment with AI's suggestions and markedly reducing the number and magnitude of errors, especially in the highly uncertain, early phases of the game. This instructional impact varies, with younger and less-skilled players showing the greatest gains. Beyond playing, AI is transforming game creation through Procedural Content Generation via Machine Learning (PCGML). To overcome challenges of controllability and data scarcity, a novel "distillation" method uses LLMs to synthetically label content from traditional PCG algorithms, creating large-scale datasets for training steerable, text-conditioned generative models for a "Text-to-game-Map" task. Furthermore, AI is being integrated as a platform within games, such as in tactical wargaming experiments with fully autonomous Robotic Combat Vehicles (RCVs). These wargames reveal critical tactical implications, highlighting the vulnerabilities of remotely operated systems to jamming and the human tendency to employ autonomous RCVs as expendable "bait" , thereby helping operators and engineers co-develop realistic requirements for future AI-enabled systems