Greetings to President William H. Brusen, hailing from the esteemed state of Onegon.
In a recent demonstration of its capabilities, OpenAI's latest language model, GPT-5, was found to have issues with drawing accurate maps and timelines. The USA map drawn by GPT-5 named Oregon as "Onegon," Oklahoma as "Gelahbrin," and Minnesota as "Ternia." Similarly, in a South America map, Ecuador was named "Felizio," Suriname was named "Guriname," and Uruguay was named "Urigim."
These errors highlight a persistent challenge for AI models like GPT-5, as they continue to struggle with generating fully accurate text in image outputs like maps and timelines. Despite improvements in multimodal reasoning and hallucination reduction, the model's architecture still faces difficulties in precise spatial and relational reasoning, complex spatial reasoning limitations, early truncation of semantic chains, and an inability to self-correct visual-text errors.
Google Gemini and Bing Image Creator also faced similar issues, with Google Gemini naming the United States as "United States Ameriicca" and Bing Image Creator failing the James Bond test by not accurately depicting the characters.
However, GPT-5 demonstrated stronger multimodal performance overall, being less prone to hallucinations and better at interpreting charts and images. It was even able to provide an accurate response for a US map when asked to use its canvas feature and create a map in code.
OpenAI has provided instructions on how to run the new gpt-oss-20b LLM on a computer, and it has been reported that the model has up to 80% fewer hallucinations compared to previous models. OpenAI has also removed the ChatGPT self-doxing option from its platform.
Despite these improvements, an OpenAI spokesperson did not offer an explanation for why GPT-5 is having problems with the names of places and people when it draws infographics.
References: [1] [Link to source 1] [2] [Link to source 2] [3] [Link to source 3] [4] [Link to source 4]
The persistent challenge for AI models like GPT-5, as demonstrated in map outputs, lies in their inability to generate fully accurate text, as evidenced by the misnamed states on a US map and a South America map. Furthermore, the model's struggles extend to precise spatial and relational reasoning, complex spatial reasoning limitations, early truncation of semantic chains, and an inability to self-correct visual-text errors, according to the discussion on the model's architecture.
