The Confusing Conundrum of GPT-5

OpenAI's CEO, Sam Altman, has been touting GPT-5 as a breakthrough in achieving human-like intelligence. However, a recent experiment has revealed that this advanced language model can still stumble spectacularly, especially when faced with even the simplest challenges.

The Experiment That Exposed the Flaws

In a baffling demonstration by Gary Smith, an economics professor, GPT-5 was asked to play a game of "rotated tic-tac-toe." The rules are straightforward: the grid is rotated 90 degrees before play begins. Logically, this shouldn’t change the game at all — it’s still the same three-by-three setup with identical win conditions.

GPT-5’s Overcomplicated Response

Instead of grasping this simple concept, GPT-5 launched into an elaborate explanation about how players' perceptions might be subtly altered. It claimed, "Rotating the board 90° doesn’t change the set of possible wins... but psychologically, it could feel different." Really?

Muddled Logic on Winning Strategies

The confusion escalated when GPT-5 misrepresented basic game theory by asserting that starting in the center is the best first move, despite established strategies favoring the corners. It even suggested that players might misjudge edge versus corner moves after orientation changes — a statement that left many scratching their heads.

The Breakdown with Image Generation

Things went further awry when Smith asked if rotating the grid made it harder for humans to play. GPT-5 maintained that from a strategic viewpoint, rotation makes no difference, yet went on to contradict itself by saying the human experience varies. This verbosity echoed characteristics more typical of its predecessor, GPT-4o, suggesting a shift in design philosophy toward a more congenial tone — likely in response to user feedback.

Disastrous Drawings: A Low Point