GPT-5’s Blunders Show It Still Has a Long Way to Go
402OpenAI’s new big language model, GPT-5, has been hyped as getting close to human-level smarts. But real-life tests show it still makes some pretty silly mistakes. Recently, a retired economics professor named Gary Smith put GPT-5 to the test with a simple game—rotated tic-tac-toe—and the results were eye-opening.
What Happens When GPT-5 Meets Rotated Tic-Tac-Toe
Smith explained that he asked GPT-5 about playing a version of tic-tac-toe where the board is rotated 90 degrees before the game starts. You’d think this wouldn’t matter, since the game itself stays the same. But GPT-5 gave a long-winded explanation about how rotation might change how players see the game. It said that while the rules stay the same, the psychology might be different because players see the board differently. That’s not quite right. The game’s logic doesn’t really change with rotation.
Then Smith asked if rotating the board makes it harder for humans to play well. That’s when GPT-5 started to get confused. It claimed that rotating the board doesn’t change anything for strategy because all winning lines just shift around. That part is true—if you’re a perfect player or a computer. But GPT-5 then added that “for humans, the story is different,” which is wrong. Playing rotated tic-tac-toe doesn’t actually make the game harder for humans because the rules are the same, and players can adapt easily.
GPT-5’s Visual Aid Goes Completely Off the Rails
Smith then asked GPT-5 to draw a rotated tic-tac-toe board, with labels, to help understand how the rotation messes with recognition. GPT-5 responded by generating an image. But the picture was a mess. It was full of typos, garbled grids, and strange blank images that made no sense. It looked like a malfunctioning robot trying to draw. Smith saw the faults and decided not to even bother responding further.
This episode shows that despite the hype, GPT-5 still struggles with simple tasks, especially ones that require understanding or visual representation. The model’s tendency to produce confusing or inaccurate outputs is still very much present. It’s a reminder that AI models, even the most advanced ones, aren’t perfect and often make obvious mistakes.
What Does This Say About AI Progress?
OpenAI has claimed that GPT-5 should feel more like talking to a helpful, intelligent friend. But the reality is quite different. The errors in Smith’s test highlight that GPT-5 can be overly verbose, inaccurate, and even quite goofy. The model’s confidence doesn’t always match its understanding, leading to situations where it confidently states wrong facts.
Some speculate that OpenAI’s decision to make GPT-5 more “friendly” and less formal was driven by user feedback. But these improvements might come at the cost of accuracy or clarity. GPT-5’s mistakes in this experiment show that AI still has a long way to go before it can reliably handle tasks that seem simple to humans.
This isn’t to say that AI isn’t advancing. It’s just a good reminder that models like GPT-5 still need more work before they can replace human intuition or judgment. As AI continues to develop, it’s important to keep testing it with real-world scenarios—and to keep expectations realistic.
In the end, Smith’s experiment is a clear example that AI, no matter how advanced it seems, still has its quirks. And it’s a sign that we should stay cautious about claims that these models are nearly human. They’re powerful tools, but they’re far from perfect.
What do you think?
It is nice to know your opinion. Leave a comment.