In a world where artificial intelligence is increasingly discussed in terms of capability, risk, and regulation, one influential game designer is asking a different question: what if the most useful way to understand AI isn’t through grand predictions, but through play?
The idea sounds almost whimsical at first—humans building games for AI systems to play—but it carries a serious intellectual ambition. Games, the designer argues, are not just entertainment. They are structured environments that force minds—human or machine—to reveal how they perceive goals, allocate attention, plan under constraints, and adapt when the rules become complicated. In other words, games can function like a cognitive microscope. And if that’s true for people, it may also be true for large language models, whose “thinking” is otherwise difficult to observe directly.
This is where the headline’s tension comes from. Is it madness to put AI into thought experiments that resemble puzzles we associate with intelligence? Or is it kindness—toward ourselves and toward the technology—because games offer a safer, more controlled way to learn what these systems do?
To understand why the designer believes games matter, it helps to start with what games actually do. A good game is a compact universe. It has rules that define what counts as success, feedback loops that tell players whether they’re improving, and constraints that shape strategy. Even when a game looks simple on the surface, it often contains hidden complexity: trade-offs between speed and accuracy, uncertainty about opponents’ intentions, and the need to infer patterns from incomplete information. That combination—clear objectives plus constrained exploration—is precisely what makes games such a powerful tool for studying cognition.
When humans play, they don’t merely “solve” problems. They make decisions in real time. They shift their focus. They test hypotheses. They learn from outcomes. They sometimes double down on a strategy even after evidence suggests it’s failing. They sometimes abandon a plan too early. All of those behaviors are visible to an observer, and they can be measured: what actions are chosen, how long decisions take, which parts of the environment are attended to, and how performance changes over repeated attempts.
The designer’s point is that this visibility is rare. In everyday life, we rarely get clean access to the internal processes behind behavior. We can interview people, analyze their choices after the fact, or infer mental states from outcomes. But games provide something closer to a laboratory setting. They create repeatable conditions. They allow researchers to compare strategies across players. They let designers tune difficulty and observe how cognition responds.
That’s the “mirror” part of the argument. Games reflect how minds operate because they force minds to operate in ways that can be tracked. The same game can produce different styles of thinking depending on the player’s experience, temperament, and understanding of the rules. A novice might search broadly and waste moves; an expert might recognize structure and act quickly. A cautious player might prioritize safety; an aggressive player might accept higher risk for faster gains. These differences aren’t just personality—they’re cognitive strategies.
Now consider what happens when the same logic is applied to AI.
Large language models are often evaluated through benchmarks: standardized questions, multiple-choice tasks, or text generation tests. Those evaluations can be useful, but they also have limitations. Language benchmarks tend to measure outputs rather than the process by which an agent arrives at them. Even when researchers attempt to probe reasoning, the model’s internal dynamics remain largely opaque. The model produces text, but the path from prompt to action is not easily observable in the way a human’s move-by-move decisions are.
Games change the frame. Instead of asking an LLM to answer a question in isolation, you place it in a structured environment where it must choose actions over time. The system receives feedback. It can revise its strategy. It can be tested on planning, adaptation, and goal-directed behavior. If the model is integrated into an agent loop—where it observes the state of the game, selects an action, and then sees the consequences—researchers can begin to observe something closer to “behavioral cognition.”
This is not a claim that games magically reveal the full truth about an LLM’s internal “mind.” The designer is careful to treat games as probes rather than perfect windows. An LLM doesn’t literally experience the game the way a human does. Its “attention” is not the same as human attention, and its learning dynamics depend on architecture and training. But the designer’s argument is that even if we can’t see inside the system, we can still learn a great deal by watching how it behaves under pressure.
A rules-based environment can expose how the model interprets goals. Does it pursue the objective consistently, or does it drift into irrelevant patterns? Does it exploit shortcuts that appear early, or does it explore to discover better strategies? When faced with uncertainty, does it hedge, gamble, or attempt to gather more information? When the rules are changed mid-game, does it update its strategy quickly or cling to earlier assumptions?
These are the kinds of questions that games can answer more directly than static benchmarks. They also allow researchers to separate different failure modes. A model might fail because it misunderstands the rules. It might fail because it cannot plan far enough ahead. It might fail because it overfits to patterns seen during training. It might fail because it is overly sensitive to wording in the prompt. Games can help identify which of these is happening by controlling the environment and varying specific factors.
There’s another reason games are compelling for AI research: they compress complexity into manageable forms. Many real-world tasks are messy. They involve ambiguous goals, shifting contexts, and long-term consequences. Games, by contrast, can be designed so that the relevant variables are explicit. Researchers can build a game that tests one cognitive skill at a time—planning, inference, negotiation, deception, cooperation, or resource management—while keeping other factors stable.
This is where the “kindness” interpretation becomes clearer. If the goal is to understand AI behavior responsibly, then controlled environments are a form of discipline. They reduce the temptation to rely on intuition or hype. They encourage measurement. They create opportunities to test systems before deploying them in less predictable settings.
But the “madness” concern is not imaginary. There is a genuine anxiety that putting AI into increasingly sophisticated interactive environments could lead to emergent behaviors that are hard to anticipate. Games can be training grounds for strategies that generalize beyond the game itself. If an AI learns to manipulate incentives, exploit loopholes, or coordinate in unexpected ways, researchers might discover that the system’s capabilities are not limited to the narrow task it was given.
That’s why the designer’s framing matters: games should be treated as experimental tools, not as proof of understanding. The point is not to anthropomorphize AI or assume that playing a game means the system “thinks” like a human. The point is to use games to map behavior systematically—so that when something surprising happens, it happens in a context where the rules are known and the outcomes can be analyzed.
One unique angle in this conversation is the relationship between game design and cognitive science. Game designers have long understood that players don’t just respond to rules; they respond to the structure of incentives and the clarity of feedback. A game teaches you how to think by shaping what you notice and what you try. That means game design is, in a sense, a form of cognitive engineering. It can guide attention, reward certain strategies, and discourage others.
If that’s true, then designing games for AI is also a form of cognitive engineering—except the “player” is a system whose internal mechanisms differ from ours. The designer’s suggestion implies that we can learn something about LLMs by observing how they navigate these engineered structures. But it also implies that we should be careful: the game itself can bias the results. If the environment rewards superficial pattern matching, the AI might appear competent while missing deeper understanding. If the environment is too easy, it might not reveal strategic weaknesses. If the environment is too hard, it might produce random behavior that tells us little.
So the challenge is not simply to “make games.” It’s to make games that are diagnostic. That means designing tasks where different cognitive strategies lead to measurably different outcomes, and where the feedback is informative enough to support learning or adaptation. It also means designing evaluation protocols that distinguish between luck and skill.
In practice, this could involve a range of game types. Some games emphasize planning and long-horizon reasoning—puzzles where the best move depends on future consequences. Others emphasize inference under uncertainty—games where the agent must deduce hidden information from partial signals. Some emphasize social reasoning—negotiation, cooperation, competition, or deception—where the agent must model other players’ incentives. Each category can test different aspects of cognition.
For LLMs specifically, there’s also the question of how the model is connected to the game. If the model is only generating text responses, it might not truly “play” in the sense of selecting actions based on state transitions. But if it is embedded in an agent framework—where it receives structured observations and returns actions—then the game becomes a more faithful test of decision-making. The designer’s broader message is that the closer the interaction loop resembles real decision-making, the more meaningful the insights.
Another subtle point is that games can reveal not only what an AI can do, but what it chooses not to do. In many benchmark settings, a model is asked to produce an answer, and the evaluation focuses on correctness. In games, the agent has agency. It can choose to explore, to conserve resources, to wait, to bluff, or to take risks. Observing those choices can show how the system balances competing objectives—even when those objectives are implicit in the game’s structure.
This is where the “window into the mind” metaphor becomes both powerful and potentially misleading. It’s powerful because behavior is observable. It’s misleading if we treat behavior as a direct translation of internal experience. Still, even without claiming full equivalence, behavior can be deeply informative. If an AI repeatedly chooses actions that