Even Super-Smart AI Can Get Beaten by a 45-Year-Old Video Game
Hello everyone, John here! Welcome back to the blog where we break down the big, confusing world of AI into bite-sized, easy-to-understand pieces. Today, we’ve got a story that’s both hilarious and incredibly insightful. Imagine the world’s most advanced AI, a tool that can write poetry, code websites, and answer complex questions, getting completely outsmarted in a game of chess. Now, imagine it’s not being beaten by a grandmaster, but by a video game from 1979!
It sounds like a joke, but it’s exactly what happened. Not once, but twice! First with the famous ChatGPT, and now with Microsoft’s own AI assistant, Copilot. Let’s dive into this fascinating tale of David vs. Goliath, or perhaps more accurately, dusty old video game cartridge vs. billion-dollar AI.
Meet the Unlikely Champion: Atari 2600 Video Chess
Before we talk about the AI getting checkmated, let’s talk about the champion. The hero of our story is a game called Video Chess, released for the Atari 2600 console. My assistant, Lila, looks a bit puzzled.
Lila: “John, I’m lost already. What on earth is an Atari 2600? And the article mentioned an ’emulator’—what’s that?”
Great questions, Lila! For our younger readers, the Atari 2600 was one of the very first popular home video game consoles. It came out in 1977! Think of it as the great-great-grandfather of the PlayStation 5 or the Nintendo Switch. The graphics were just simple blocks, and the sounds were basic beeps and boops. It was incredibly simple technology compared to today.
And an emulator is just a special piece of software that lets a modern computer pretend to be an old one. So, a researcher named Robert Caruso didn’t need to find a 45-year-old console in an attic; he just used an emulator on his PC to run the classic Video Chess game.
The chess game itself was revolutionary for its time, but by today’s standards, it’s a very weak player. It makes basic mistakes and isn’t very strategic. It’s programmed with the rules of chess and a few simple strategies, and that’s it. Keep that in mind, because it makes what happens next even more surprising.
Round 1: ChatGPT Faces the Ancient Beast
A little while ago, Mr. Caruso had the brilliant idea to pit ChatGPT against this ancient chess program. He would tell ChatGPT what move the Atari made, and then input ChatGPT’s requested move back into the game. The result? Total humiliation for the super-advanced AI.
ChatGPT made moves that were not just bad, but completely illegal. It would try to move pieces in ways they aren’t allowed to, or it would forget where pieces were. It was a mess! The simple, rigid programming of the Atari game, which could never make an illegal move, easily won.
Lila: “Wait a second. I thought ChatGPT was supposed to be a genius! Why would it lose so badly? Doesn’t it know the rules of chess?”
That’s the million-dollar question, Lila. And it gets to the heart of what these AIs really are. ChatGPT is what’s known as a Large Language Model, or LLM. The best way to think of an LLM is like a super-powered autocomplete or a parrot that has read the entire internet.
- It’s trained on trillions of words and sentences from books, websites, and articles.
- Its main goal is to predict the next most likely word in a sentence.
- It’s fantastic at generating text that sounds human because it has seen countless examples of how humans write and talk.
So when you ask it to play chess, it isn’t “thinking” strategically like a dedicated chess computer. It’s just trying to predict what a common next move would be based on all the chess game descriptions it has read online. It knows what a chess game looks like in text, but it doesn’t truly understand the underlying rules or logic. It’s like asking a brilliant poet to fix a car engine—they can describe the engine beautifully, but they have no idea how the pistons and spark plugs actually work together.
Round 2: Microsoft’s Copilot Tries Its Luck
You’d think one public defeat would be enough. But Robert Caruso, not satisfied with stumping one AI, decided to try again with Microsoft’s Copilot. Copilot is another powerful LLM, deeply integrated into Microsoft’s products like Windows and its search engine, Bing. It’s also based on cutting-edge technology from OpenAI (the creators of ChatGPT). Surely, it could beat a game from the 1970s, right?
Wrong. So, so wrong.
Copilot stepped up to the board and, according to the report, its “confidence was… misplaced.” It fell into the exact same traps as ChatGPT. It started making bizarre, nonsensical moves. For example, it would try to move its king into a position where it was already under attack (a totally illegal move known as moving into check). It was essentially walking into a trap with its eyes wide open.
The Atari 2600, with its simple, unyielding logic, just patiently played by the rules until Copilot had either made a series of illegal moves or put itself in an unwinnable position. Another victory for the old-timer!
Why Do These Modern AI Marvels Keep Failing?
This is the most important lesson from this funny experiment. It perfectly highlights a key truth about the current state of artificial intelligence: not all AI is the same.
The AI in ChatGPT and Copilot is a generalist. It’s designed to be a jack-of-all-trades for language. It can chat, summarize, write, and brainstorm. But it’s not a specialist.
On the other hand, a chess-playing program—even a ridiculously old one like Atari’s Video Chess—is a specialist. Its entire existence is dedicated to one thing: executing the rules of chess. It doesn’t know how to write a poem or tell you a joke, but it knows that a pawn can’t move backward and a bishop must stay on its color.
Lila: “So are you saying the AIs aren’t really ‘playing’ chess at all? They’re just… talking about chess?”
That’s a fantastic way to put it, Lila! They are essentially generating a text response that looks like a chess move. They’re mimicking the patterns they’ve seen in data, but they lack a true, internal model of the game board and its strict rules. The Atari game, for all its faults, has that internal model. It’s simple, but it’s unbreakable.
This is why AIs like Deep Blue and AlphaGo, which famously beat world chess and Go champions, are so different. They were highly specialized AIs, built from the ground up with the sole purpose of mastering the logic and strategy of their respective games. Comparing ChatGPT to AlphaGo is like comparing a Swiss Army knife to a surgeon’s scalpel. One is useful for many things, but you know which one you want for a specific, critical job.
A Few Final Thoughts
John’s Perspective: I honestly love stories like this. They bring a much-needed dose of reality to the AI hype. It’s a humbling and hilarious reminder that these powerful new tools have very real limitations. It teaches us to be critical thinkers and to use the right tool for the right job, rather than assuming “AI” can magically do everything.
Lila’s Perspective: I find this really encouraging! It makes AI seem a lot less scary. Knowing that these “genius” systems can be flummoxed by my dad’s old video games makes them feel more like a tool I can learn to use, not some all-knowing being to be afraid of. It’s actually pretty funny!
So, the next time you hear about the incredible power of AI, just remember the mighty Atari 2600, our reigning, undefeated champion. It’s a great lesson that sometimes, newer isn’t always better for every single task.
This article is based on the following original source, summarized from the author’s perspective:
Microsoft Copilot joins ChatGPT at the feet of the mighty
Atari 2600 Video Chess