Subscribe to the Popular Science daily newsletter for breakthroughs, discoveries, and DIY tips delivered six days a week. Artificial intelligence (AI) models’ progress is often illustrated by their gaming prowess. IBM’s Deep Blue stunned the world in 1997 by defeating chess grandmaster Garry Kasparov, and nearly two decades later, Google’s AlphaGo bested a human champion in Go—once thought impossible. Since then, AI has advanced from board games to video games, using reinforcement learning, a technique also crucial for training chatbots like ChatGPT, enabling machines to master Atari games and complex strategy titles such as Dota 2 and Starcraft II. However, AI still struggles with quickly learning various more open-ended games—a domain where humans excel. When faced with an unfamiliar game, human players rapidly grasp its basics, while AI models often fail, as highlighted in a recent paper by NYU computer science professor Julian Togelius and colleagues. This gap underscores a fundamental difference between human intelligence and AI’s current capabilities, highlighting that AI has a long journey ahead before achieving or surpassing true human-level intelligence. Games have long served as ideal testing grounds for AI due to their predictable rules, defined goals, and mechanics, which fit well with reinforcement learning: models repeatedly play games in simulations to improve via trial and error. This approach enabled DeepMind’s 2015 mastery of Atari games and influences today’s large language models trained on massive internet data. Yet, these models excel only at specific tasks with clear constraints; slight changes to game design can disrupt AI performance. While AI may achieve superhuman skill in a particular game, it struggles with improvisation. This limitation is more apparent as modern games become increasingly open-ended and abstract. Unlike chess, games like the open-world “Red Dead Redemption” have complex objectives tied to embodying a morally conflicted character rather than straightforward goals.

Humans intuitively grasp such nuances; machines do not. Even in simpler sandbox games like “Minecraft, ” AI may perform basic actions like jumping without understanding their context. The authors emphasize that well-designed games align closely with human intuition, common sense, and lived experience—which humans accumulate over years of real-world interaction. For example, babies learn to recognize objects by about 18 to 24 months simply through experience, while machines require much more guided input. This experiential advantage allows humans to learn new games faster. Research shows curiosity-driven reinforcement learning AI may need about four million keypresses—or roughly 37 continuous hours—to complete a game, whereas average human gamers often grasp new mechanics in under 10 hours. Nevertheless, AI is advancing in general gameplay. In 2023, Google DeepMind introduced SIMA 2, a model combining existing AI with reasoning capabilities from its Gemini large language model, enabling better understanding and interaction with 3D games—even those it wasn’t specifically trained on. Yet, Togelius and colleagues caution that AI still has significant ground to cover before matching human adaptability. They propose a benchmark where a model could play and win the top 100 games on Steam or the iOS App Store without prior training on any of them—and do so in roughly the time a human would take. This remains a formidable challenge that current AI methods are neither close to solving nor seriously attempting. Achieving this level of generalization would require AI to demonstrate true creativity, forward planning, and abstract thinking—qualities uniquely characteristic of human intelligence. Ultimately, the real test for AI reaching “human-level intelligence” may lie not in creating deepfakes or writing shallow novels but in its ability to master a wide array of diverse games with human-like learning speed and understanding.