GPT-5 Launch Gets Off to a Rocky Start with Major Mistakes
353OpenAI just rolled out GPT-5, the latest big step in AI tech that CEO Sam Altman called a major move toward artificial general intelligence. But the debut didn’t go quite as planned. During a livestream showcasing GPT-5, the company shared some performance charts that turned out to be full of errors. Instead of impressing viewers, the graphs caused quite a bit of confusion and embarrassment.
Chart Fails Steal the Spotlight
The most talked-about mistake was a set of bar graphs comparing GPT-5’s coding skills to older models. These charts looked professional at first glance but fell apart when people looked closer. The graph showed GPT-5 with a score of 52.8 percent accuracy, but the bar was nearly twice as tall as a bar for an older model with a 69.1 percent score. Even stranger, another bar showing GPT-4o’s 30.8 percent accuracy was the same size as the 69.1 percent bar. How does that make any sense? It’s a clear error, and people on social media pointed it out quickly. The Verge highlighted the mistake, and it became a running joke among AI fans.
OpenAI didn’t confirm whether GPT-5 made the graphs or if someone else did. Either way, it’s a pretty embarrassing slip for a company valued at around half a trillion dollars. It’s kind of poetic that a company so focused on AI performance could mess up something so simple.
AI’s Growing Pains and Hilarious Failures
This isn’t just about bad charts. Some recent research suggests that newer AI models might actually be getting worse in some ways. Reports show they hallucinate more often — meaning they make up facts — and that the longer they “think,” the more their accuracy drops. Other studies point to training data issues, where bad information pollutes the AI’s learning process.
After GPT-5 launched, users quickly tested its abilities beyond just numbers. Many found that GPT-5 still struggles with basic tasks like drawing maps or recognizing real-world locations. One user asked ChatGPT to map out two cities in Virginia with neighborhoods labeled, but the AI returned bizarre and nonsensical names. Ed Zitron, writing for the “Where’s Your Ed At?” newsletter, tested GPT-5’s ability to draw a simple map of the US and was met with amusingly incorrect results. Instead of real place names, it suggested places like “West Wigina,” “Delsware,” “Fiorata,” and “Rhoder land.” Tellingly, it also mangled states’ names, calling Tennessee “Tonnessee” and Mississippi “Mississipo.”
AI’s Promises vs. Reality
Earlier this year, OpenAI claimed that an update to GPT-4o improved its ability to generate text within images. They boasted that ChatGPT could now produce accurate text in images, showing off examples where the AI seemingly nailed it. Yet, the reality is quite different. Many users have pointed out that GPT-5 and its predecessors still produce errors, especially in visual tasks. The supposed leap in capabilities doesn’t seem to match what users are experiencing.
This disconnect raises questions about how much progress AI has really made. Some experts argue that these models are actually regressing in certain areas, hallucinating more, and making more mistakes than before. The ongoing issues with training data quality and the complexity of AI reasoning may be contributing factors. As AI systems get more sophisticated, they also seem to become more prone to errors and falsehoods.
In the end, GPT-5’s rocky start highlights how challenging AI development really is. Even companies with massive resources and high hopes can stumble on basic tasks, reminding us that AI still has a long way to go. The mistakes serve as a reminder that progress in AI isn’t always linear, and sometimes, the biggest leaps come with unexpected setbacks.
What do you think?
It is nice to know your opinion. Leave a comment.