Can AI Play Video Games? Why It's Slower Than a Kid

Last updated: October 2026

A teenage student of mine had an idea during a lesson: write a program where an AI plays a fast action game.

Not a slow puzzle game where it could take its time. The kind where things come at you and you have to react before you have finished thinking.

It is a good idea. It is also exactly the kind of idea that teaches you how these tools really work, because it fails for a reason almost nobody explains to kids.

So, can AI play video games? Yes, some can, and one racing AI has beaten four of the world's best players at its game. But not the AI most people picture: the chat-style model that writes essays, answers questions and writes code. Put that kind of AI in charge of a fast game and it is slower than any child holding a controller.

Every move is a whole conversation

Here is what I explained.

When you play a game, you look at the screen and your hands move. There is no gap you would notice. Seeing and doing feel like one thing.

A chat-style model does not work like that. Every time it has to decide something, it has to be handed the situation, read all of it, and then write a reply. Only when the reply is finished do you know what it wants to do. By then the game has moved on, so you describe the new situation and it starts again.

That read-and-reply cycle takes seconds. For writing a paragraph, a few seconds is nothing. For a game where an enemy crosses the screen in less than a second, it is an eternity.

I have tested this myself. I set up an AI to drive a car in a racing game, and it made one decision every 10 to 20 seconds. Picture driving a car when you are only allowed to touch the wheel once every fifteen seconds. That is not a driver. That is a passenger with opinions.

The engineering word for this gap between asking and getting an answer is latency. It is one of the most useful ideas in computing, and children almost never hear it named. Once you have the word, you start seeing it everywhere: in games, in video calls, in why a website feels slow.

The AI that beat the world's best drivers

There is another side to this, though, because "AI can't play games" is not true either.

In 2022, researchers published a paper in Nature about Gran Turismo Sophy, an AI driver for the PlayStation racing game Gran Turismo. It won a head-to-head competition against four of the world's best Gran Turismo drivers.

The detail that matters for my student's question is buried in one of the paper's figure captions. It compares a human driver's steering and braking in a time trial with the AI's, and describes "GT Sophy's 10-Hz action rate." Ten decisions every second.

Do the arithmetic. In the time my AI driver made one decision, GT Sophy would have made somewhere between 100 and 200. That is a rough comparison, not a fair race. Different games, different setups, different goals. But the gap is the whole point.

GT Sophy is not a chatbot. It was trained with deep reinforcement learning for one job: turning what is happening on the track into steering, throttle and braking. It was never built to explain a loop or write you a story. It races.

And being that good at one thing took serious machinery. Gran Turismo's own page on the technology says the training platform "had access to over 1,000 PlayStation 4 (PS4) consoles," each used to collect data for training or to evaluate a trained version.

So the answer I gave my student holds up: real-time play needs much smaller, faster models, or a different design. A different design could mean a game where the AI does not have to react in real time at all, where it takes turns, or plans between rounds, or runs the world while the human does the reacting.

"EXCELLENT. Very hands-on. They actually teach concepts and have you apply them immediately. I highly recommend for anyone who wants to learn from ground zero." Sarah, parent of a daughter learning Python

That is how I like a concept to arrive. Not as a lecture on inference time, but because a student's own idea ran straight into it.

What I tell every student now

Speed is a property of the tool, not a detail you sort out later. Every answer from a chat-style model costs time, and a bigger or more careful model usually takes longer. The same trade-off shows up in the effort settings on AI coding tools, where a maximum-effort run can take nine minutes and produce nothing better than a lower setting.

So before any AI idea, I now ask one question: how fast does this need to answer? If a human would happily wait a few seconds, a chat-style model is fine. If the answer has to arrive before the next frame of a game, you need something built for speed, and that is a very different project.

It also explains where AI already helps kids with games today, which is on the building side rather than the playing side, and why an AI coding agent is one of the routes a teen can take into making a first game. In one of my lessons, a 10-year-old directed an AI to add an obstacle to a game and saw it appear within minutes. Building can wait a few seconds. Playing cannot.

My student's idea was not wrong. It was aimed at the wrong kind of AI, and it ran into one of the real limits of the tools. That is the best kind of idea to have in a lesson, because the limit is the lesson.

If your child comes up with ideas like this, those are the lessons I most enjoy building, and a free Discovery Call is where we work out what they are ready to try next.

Frequently Asked Questions

Can AI play video games? Yes. Specialised game-playing AIs can play very well, and Sony AI's Gran Turismo Sophy beat four of the world's best Gran Turismo drivers. A general chat-style model is a different story: every decision is a full read-and-reply that takes seconds, which is far too slow for a fast game.

Why is a chatbot so slow at playing games? It has to read the whole situation and write out a reply before anything happens in the game. In my own test, an AI driver in a racing game made one decision every 10 to 20 seconds, while GT Sophy acts ten times a second.

What AI game project suits a beginner? One where waiting does not matter. A turn-based game, an AI opponent that plans between rounds, or letting AI help build the game rather than play it all avoid the latency problem.