ChatGPT-5’s ‘PhD-Level’ AI Struggles With Spelling and Geography

OpenAI launched ChatGPT-5 with bold claims, calling it a “PhD-level expert” in your pocket. But within hours, users began posting examples of its obvious errors. The AI model often miscounted letters, made spelling mistakes, and even invented U.S. states.

CEO Sam Altman introduced the update by saying ChatGPT had evolved from high school to PhD intelligence. Yet, early users weren’t convinced by the model’s accuracy or reliability in basic tasks.

‘Blueberry’ Has How Many Bs?

One of the most viral mistakes involved the word blueberry. A user on Bluesky asked ChatGPT-5 how many Bs it had. The AI confidently said three, describing a poetic “bb moment” that “makes the word feel extra bouncy.” Of course, there are only two Bs in “blueberry.”

Such mistakes raise serious questions about the chatbot’s core language comprehension. Despite its friendly tone, the model failed to count letters correctly, which should be a basic function.

Geography Gone Wild

Another user asked the model to list U.S. states containing the letter R. GPT-5 not only gave incorrect answers but also invented fictional states. These included “Krizona,” “Vermoni,” “Mitroinia,” and “New Jefst.” Even worse, it listed California twice and omitted valid states.

When asked to display a map, ChatGPT-5 misspelled Northern Territory as “Northan Territor” and claimed it had only three Rs instead of five.

What’s Causing These Mistakes?

OpenAI attributes some issues to the new “real-time router” built into GPT-5. This router dynamically selects which internal model to use based on context. If a user adds “think hard about this,” the system switches to a reasoning-optimized model.

While this routing concept sounds smart, it appears inconsistent in execution. At times, simple tasks are handled by less accurate sub-models.

Users Say the Model Hallucinates

Dan Shipper, CEO of media startup Every, tested the bot on literature passages. When he submitted an image from a novel and asked for an explanation, GPT-5 often hallucinated the plot. But when he added “think longer” to his prompt, the AI gave a more accurate summary.

This inconsistency suggests the router doesn’t always choose the right model for the job—especially for nuanced or creative tasks.

Still a Step Toward AGI?

Despite flaws, OpenAI immediately released GPT-5 to 700 million weekly users. Altman confirmed the system isn’t yet artificial general intelligence (AGI) but said it’s “generally intelligent” and marks a key milestone.

However, with basic spelling and geography errors so visible, many wonder whether this “PhD-level” intelligence is ready for real-world application.

From poetic miscounts in “blueberry” to fictional states like “New Jefst,” the ChatGPT-5 spelling geography flaws remind us that even cutting-edge AI has limits. While it represents progress in reasoning and language, the current version shows that human oversight remains essential—for now.