Artificial Intelligence in Video Games

While the relationship between artificial intelligence (AI) and games far exceeds the parameters that a short review can even sketch out, I hope to propose a series of representational, historical, aesthetic, technical, experimental, and processual ways to approach the intersection between these two terms.Representations of AI and intelligent robots have been common throughout video game history, ranging from nonplayer to player characters. Some representations of AI antagonists include AM in I Have No Mouth, and I Must Scream (Cyberdreams and the Dreamers Guild, 1995), GLaDOS in Portal (Valve, 2007), and P03 in Inscryption (Daniel Mullins Games 2021). Depictions of helpful nonplayer AI familiars include Rush in Mega Man 3 (Capcom, 1990), Claptrap in Borderlands (Gearbox Software, 2009), and Nick Valentine in Fallout 4 (Bethesda Game Studios, 2015). Another important category comprises AI player-characters whom a player controls or roleplays as but that do not depend on actual AI systems, including Mega Man throughout the Mega Man series (Capcom, 1987–), the unnamed android of The Talos Principle (Croteam, 2014), and SAM in Observation (No Code, 2019). The centrality of science fiction as a popular narrative genre in video games—a medium that emerged during the Cold War when speculative narratives were assuming a more central form within popular culture—might account for the prevalence of AI characters in games.Unlike literature, film, or television series, however, the history of AI in video games exceeds representation. The earliest history of AI includes uses of games to test AI programs. Already in the late 1940s, mathematicians Alan Turing and Claude Shannon both used chess-playing programs as a theoretical route into the problem of how a computer could think. Some of the earliest examples of video games, such as Tennis for Two (William Higinbotham, 1958), Spacewar! (Steve Russell, 1962), and Pong (Atari, 1972), were two-player games in which the challenge derived from the other player, thereby eliminating the need for AI. The development of AI in popular video games initiated an era of single-player engagements, starting with games such as Computer Space (Syzygy Engineering, 1971) and Space Invaders (Taito, 1978). Enemy behaviors in such single-player games, which depend on character-based scripting and pathfinding algorithms, are more basic than the supervised and unsupervised learning of later neural networks. Historically, developers have often spent more resources on graphics than AI systems in order to appeal to consumers through visual delights. Even so, AI has become more important to video games in the early twenty-first century.Though the larger topic of AI uses within video game development exceeds the scope of a short review, it is helpful to think through the aesthetic effects that AI can have in games at the intersection of interactive art and computer science. For example, as I have argued elsewhere (Jagoda 2020), difficulty is an aesthetic dimension of interactive games. Some games use AI to achieve adaptive difficulty. By contrast to variable difficulty, which allows a player to choose a difficulty setting that might include easy (or story), normal, or difficult modes, adaptive difficulty responds to a player’s performance in real time. Crash Bandicoot 2 (Naughty Dog, 1997) was one of the earliest video games to experiment with dynamic difficulty, providing a player who failed often with added checkpoints or stronger power-ups. Another example comes with the practice of “rubberbanding” in multiplayer video games in which a game increases or decreases difficulties based on players’ real-time performances, as with the infamous “spiny” or “blue” shell that targets the first-place racer in all games in the Mario Kart series (starting with Mario Kart 64 [Nintendo EAD, 1996]) to maintain parity.Other games use more sophisticated AI systems to produce emergent gameplay, which includes adaptive difficulty but also exceeds it. The Left 4 Dead (Valve, 2008, 2009) games offer one compelling example. These cooperative horror first-person shooter action games have up to four players control survivors who are immune to a virus that has turned people into zombie-like “infected” creatures. This game enables infinite replay not because of its familiar narrative premise but because of a dynamic underlying system called the “AI Director.” This system introduces unpredictable and variable elements in real time during each gameplay session. Such responsive adjustments depend on collective player positions and performance. The AI Director uses procedural generation to determine the positions and frequency of weapons, health items, enemies, and storms that influence difficulty. Beyond instrumental or threatening elements, this system also influences purely aesthetic components that do not impact gameplay difficulty, such as dynamic music, character dialogue, and visual effects. All of these elements generate opportunities for emergent gameplay in which players can improvise within a new environment in each round of play.In addition to adjusting dimensions of the overall game state, AI can also guide the behavior of a single boss-level enemy. Beyond difficulty, such a system can contribute to horror aesthetics. The game Alien: Isolation (Creative Assembly, 2014) uses AI to bring to life the creature from the Alien series, first featured in Ridley Scott’s 1979 titular film. In addition to an AI Director that tracks the game’s global state (much like in Left 4 Dead), this AI makes decisions based on behavior trees and audiovisual sensors within the game’s world. Instead of a prescripted game, this experience becomes unique in each playthrough by organizing it around an unpredictable and terrifying nonhuman alien. Other games have combined AI representation and systems. For example, the first-person survival horror game Soma (Frictional Games, 2015) has the player face an AI called the Warden Unit (WAU), which both is an AI in the game’s diegesis and behaves according to algorithms that might remain largely unknown to the player, thereby simulating unpredictability and an emergent sense of horror.The connection between AI and games moves beyond the application of the former to improve the experience of the latter. Increasingly games have become an experimental testing ground for AI developers, especially with human users. This kind of AI testing became visible and popular when the chess-playing program Deep Blue defeated world champion Garry Kasparov in 1997. Nearly two decades later, AlphaGo used Monte Carlo tree search algorithms and supervised machine learning to defeat top Go players such as Lee Sedol in 2016 and Ke Jie, in 2017; this led to the subsequent development of AlphaZero, an AI that can excel at Go by reinforcement learning that develops through neural net self-play without human supervision. Beyond the spectacle of such matches, and their narrativized metaphysical stakes of clashes between human and nonhuman intelligence, these events demonstrate the use of games as a means of AI development. The connection between AI and games has surpassed well-known and complex games, extending, for instance, to initial testing of self-driving cars in driving simulator environments.Beyond gameplay, AI has even contributed to game design. Through everything from path-finding algorithms to neural networks, human developers can guide autonomous tools to generate elements of game spaces. For example, for the tactical shooter game Ghost Recon Wildlands (Ubisoft Paris, Ubisoft Milan, 2017), autonomous techniques produced large quantities of the game’s terrain. While human designers modified and augmented material that was procedurally generated by algorithmic processes, this example suggests emergent possibilities for collaborations between humans and AI in game design. Such work between human and nonhuman systems is likely to have not only computational consequences but also aesthetic, cultural, ethical, and sociopolitical implications in the 2020s and beyond.