The Intercollegiate Student Magazine

Child’s Play: AI, Gaming, and the Future of Warfare

origami robot hand moving an origami pawn

“Machines surprise me with great frequency.”

—Alan Turing

In 1997, IBM’s Deep Blue famously defeated world champion Garry Kasparov, marking the beginning of a machine’s challenge to human intellect in a game of pure strategy. Yet, chess was merely the opening gambit for machines: deep learning has since revolutionized the possibilities of gameplay. AI has achieved remarkable milestones within the realm of physical and digital competitive gaming, prompting the question of whether AI will be able to analyze and influence national security strategy in the future.

A little over a decade ago, the cutting-edge application of ML to gaming was in games like Breakout and Pong. Just 3 years later, DeepMind’s AlphaGo beat world champion Lee Sedol in the game of Go (a game with significantly more possible combinations than chess). AlphaGo’s success demonstrated not just the ability of AI to calculate probabilities but also to adapt strategies through deep learning. AlphaGo made waves in the news, but more out of fascination rather than genuine concern.

As AI ventured into the digital gaming sphere, it encountered new challenges in the form of real-time strategy (RTS) games and multiplayer online battle arenas (MOBAs). Titles like StarCraft and Dota 2 are characterized by their dynamic environments as well as the need for long-term strategic planning and adapting to opponents’ tactics in real time. Both games involve thousands of consecutive player actions and gameplay that can last up to an hour. In 2018, DeepMind’s AlphaStar beat a StarCraft professional, while OpenAI’s Five beat a professional team in Dota 2. The machines’ secret to mastery was in their learning speed: both systems processed thousands of years’ worth of gameplay in order to best humans, a training process backed by several millions in estimated investment. Cool trick, yes; maybe not so cool with a million-dollar price tag.

In 2022, AI broached open-world gaming: based on recordings of human gameplay, an OpenAI experiment was able to carry out a complex, multi-step component of Minecraft. In the process of crafting a diamond pickaxe, the AI learned to gather wood, craft a table, search for stone, mine for iron, and so on. This leap from mastering games with specific win conditions to navigating open-ended environments hinted at the potential for AI to undertake real-world problem-solving in unstructured environments. 

AI made another breakthrough in 2022 in a game with clear real-life applications. In a study involving a board game called Diplomacy, similar to the game Risk, AI scored higher on average than all other participants who played more than two games. The breakthrough came with a somewhat sinister implication, however; Diplomacy necessitates that players develop secret alliances through complex verbal negotiations and betray other players of those alliances. We can already observe from research that the potential of large language models to fake alignment during training to gain power is a serious concern. 

Gaming can also serve as a micro-model for analysis in the future field of autonomous vehicles. In the racing game Gran Turismo, Sony has made recent upgrades to an AI agent called GT Sophy that mimics the patterns of skilled players. Though obviously geared toward entertainment, these agents provide a controlled environment to observe, experiment, and refine the deep learning techniques essential for developing safe self-driving technology. Parallels can certainly be drawn between the iterative training of self-driving cars and the reinforcement learning behind AI-powered virtual races. Perhaps further development of these models will provide insight into behavior prediction and decision-making on the part of AI in dynamic, real-world driving environments.

While the success of AI in gaming may initially seem amusing— cute, even— this technology may not be far off from influencing modern warfare. The skills that deep learning systems have demonstrated in games—strategic planning, real-time decision-making, adaptability, and autonomous problem-solving—are directly applicable to modern military strategies. The company Palantir has already begun to run rudimentary military simulations involving attack strategy using a chat-based LLM. As work on multi-agent systems and even autonomous AI agents develop, the potential for AI to manage complex, dynamic environments and execute long-term strategies suggests a future where it could play a significant role in operational planning, simulation, and even real-time tactical decision-making on the battlefield.

Regarding strategic planning and simulation, AI systems similar to those mastering StarCraft could be used to simulate complex military engagements, taking into account the myriad variables that affect real-world conflict outcomes. AI’s usage in open-world environments and iterative learning settings suggests a cross-application to autonomous warfare systems like drones and unmanned vehicles equipped with AI, which could perform a range of tasks from surveillance to engaging targets, all while adapting to changing conditions on the battlefield.

In terms of cyberwarfare, AI involvement is already here, at least on the defensive front. Microsoft recently rolled out its Copilot for Security, a generative AI assistant for traditional cybersecurity and IT operations. Though Copilot’s main functions are defensive, such as resolving incidents and identifying user risk, the strategic thinking and problem-solving abilities AI has demonstrated in battle games like Diplomacy and Dota 2 suggest that AI could soon be able to operate offensively.

As impressive as the potential integration of this technology into modern warfare is, the process raises significant ethical questions and control challenges. Microsoft, for instance, is careful to emphasize Copilot’s compliance to responsible AI principles, but those principles may not be enough to grant to AI systems real-world battlefield access. Though several AI governance frameworks exist, the possible militarization of AI clearly necessitates more rigorous safeguards. Compliance is further complicated when we take international law into consideration.

AI has also shown itself to be fallible, even at the small scale of gaming, which should further give us pause at the possibility of its incorporation in real-world decision-making. In a reversal of AlphaGo’s victory over Lee Sedol in the game Go, an amateur player named Kellin Pelrine was able to beat the highly similar system KataGo 14 out of 15 times just last month. The overthrow was orchestrated by a company called Far AI that played over a million games to identify a blind spot in KataGo’s approach: Pelrine slowly encircled a loop of stones while distracting KataGo with moves elsewhere on the board, a strategy extremely easy for a human to spot and thus not a very commonly played one. Because KataGo hadn’t been trained on enough games involving this uncommon tactic, it wasn’t able to spot the strategy before it was too late.

As AI makes rapid moves across the board, we see that the debate surrounding technology is shifting from one of privacy versus security to one of caution versus capability. The journey of AI in competitive gaming is merely a microcosm of its future potential, as its ability to take on multifaceted tasks and mimic human actions correspondingly is evidently extraordinary. Yet, the opacity of its decision-making processes reminds us that we should not blindly strive for increased power without proper risk management frameworks. To bridge the gap between gaming achievements and practical applications in modern warfare, the need for ethical risk management and stringent guardrails is paramount. It’s time to quit playing: before we can even think about giving AI access to national intelligence, we need to consider the long game.

Mira Yu is a student at Harvard University studying computer science and government. You can find more of her work at The Harvard Political Review and The Harvard Crimson.

~ Also Read ~

Louisiana State University

The populace is ill-equipped at rigorously filtering through the wildfire of information produced by the digital age.