Humans can still beat AI at video games
Ask anyone to track the progression of artificial intelligence (AI) models over the past few decades and you’ll likely hear a reference to its ability to play games. IBM shocked the world in 1997 when its Deep Blue model defeated chess grandmaster Garry Kasparov in his own field. Nearly two decades later, Google’s AlphaGo model beat a human champion in the game of Go, a feat some thought impossible at the time.
Since then, increasingly data-rich AI models have moved from board games to video games. Various models have used a training method called reinforcement learning – a technique that also plays a key role in training AI chatbots like ChatGPT – to teach machines to learn and outperform humans in a range of Atari games. More recently, reinforcement learning has taught machines to master incredibly complex strategy games, including Dota 2 and Starcraft II.
But there remains one area of ​​gaming – at least for now – where computers still can’t hold their own against flesh-and-blood humans. They are still not good at quickly learning different, more open-ended types of play. When it comes to picking up a random title from a game store that they’ve never seen before and understanding the gist of it, human gamers still learn the ropes much faster than even the most advanced AI models.
That’s the key argument made in a recent paper by Julian Togelius, a computer science professor at New York University, and his colleagues. They note that this distinction is not just a pat on the back for Homo sapiens. It could also shed light on a key part of what makes human intelligence so unique and why AI still has a long way to go before it can truly claim human-level intelligence – let alone surpass it.
“If you launch an LLM [large language model] against a match never seen before, the result is almost certain failure,” the authors write.
AI is addicted to games from the start
Games have been useful testbeds for AI models for decades, because they typically have predictable rules, defined goals, and variable mechanics. These basic principles fit particularly well with reinforcement learning, in which a model plays a simulation game over and over again, sometimes millions of times, using trial and error to gradually improve until it achieves mastery. This is, in a fundamental sense, how DeepMind was able to master Atari games in 2015. This same logic influences today’s popular large language models, even as the entire Internet serves as training data.
And yet, this method runs into problems when asked to generalize. AI models crush humans in board games and some video games because the constraints are clear and the objectives relatively simple. Ultimately, Togelius and his colleagues say that these models, as impressive as they are, still perform exceptionally well in a very specific task – and not much more. Even small variations in a game’s overall design can result in total failure. A model can be superhuman when playing a specific game, but be quite inept when asked to improvise.
This distinction becomes even clearer given the broader trend in modern gaming toward more open-ended and abstract titles. Take chess against a big-budget third-person adventure game like the open-world western “Red Dead Redemption.” Although both are games in the fundamental sense, what it means to succeed or win in each is very different. “Red Dead Redemption” offers many missions with clearly defined resolutions: shoot the bad guy, steal the horse. However, the overarching objective of the game is much less straightforward. What does it mean to earn when the central objective is to play a morally troubled Western outlaw?
Related Stories
Human players can intuit this; machines, not so much. Even in simpler games like “Minecraft,” the researchers note, an AI model can know how to jump from one block to another without having absolutely no idea what jumping actually means.
“In summary, all well-designed games are expertly attuned to human capabilities, intuition, and common sense,” the authors write.
Lived experience seems to be our biggest advantage when playing against machines. The average gamer who downloads a new version may not have been scrupulously trained by an office full of well-paid, Patagonia-clad engineers, but they have years of interaction and understanding with more abstract objects and concepts that they will later encounter in the game. The authors note that human babies learn to recognize and identify individual objects around 18 to 24 months of age, simply by existing in the world. Machines need more hands.
All of this translates to humans learning new games faster. Previous studies show that a gaming AI model using curiosity-based reinforcement learning can require four million keyboard interactions to complete a game. This translates to approximately 37 hours of continuous gameplay. The average human player, on the other hand, will typically understand completely new mechanics in less than 10 hours.
That said, the game’s AI continues to improve, even in more general contexts. Last year, Google DeepMind unveiled a model called SIMA 2, which the company describes as a significant advance in teaching AI to play 3D games in a way more similar to humans, including games it hasn’t been specifically trained to play. The key breakthrough was taking an existing model and integrating the reasoning capabilities of Google’s large Gemini language model. This combination helped him better understand and interact with new environments.
Togelius and his colleagues say these models still have a long way to go before they can be considered comparable to a human player. Their proposed benchmark involves taking a model and having it play and win the top 100 games on Steam or the iOS App Store, without having been previously trained on any of them – and doing it in about the same time it would take a human. It’s a big challenge.
“General video game playing, in the sense of being able to play any game in the top 100 on Steam or the iOS App Store after only the same amount of play time that a human would need, is a very difficult challenge that we are far from solving and are not even seriously tempted,” the authors write. “It is not at all clear that current methods and models are suitable for this problem. »
Meeting this challenge is not only of interest to the world of video games. Togelius argues that a machine capable of generalizing in this way would likely have to excel at true creativity, forward planning, and abstract thinking, all qualities that seem far more distinctly human than those possessed by current AI models.
In other words, the real test of AI’s ability to achieve “human-level intelligence” may not come from generating deepfakes or writing mundane novels, but from playing lots of games.
The best new PopSci 2025 releases
