OpenAI Touts ChatGPT Fix for Letter Counting, Other Logic Errors Remain

OpenAI announced on Tuesday that ChatGPT can now correctly answer a long-standing logic question about the number of times the letter R appears in the word strawberry. For months, the chatbot had frequently provided the wrong answer to the question and would sometimes insist the word did not contain three Rs. The issue is one of several examples of confident mistakes, where large language models provide incorrect answers while presenting them as fact. OpenAI posted on X that at long last, ChatGPT could answer the strawberry question correctly. The company also noted that another common logic prompt, asking whether to walk or drive to a car wash 50 meters away, was now handled properly. Multiple users responded to the announcement by pointing out that other similar logic questions still produced incorrect results. For example, when asked how many Rs are in cranberry, ChatGPT reportedly continued to answer that the word has one R, which is incorrect. The fixes may be hardcoded rather than representing a fundamental improvement in the model’s reasoning capabilities. Hardcoded solutions for known failure cases are not new in AI chatbots, but OpenAI’s framing of the update as a general fix drew criticism from users who noted the underlying problem persists. Confident mistakes remain a significant challenge for large language models. When presented with incorrect information, chatbots often fail to notice the error and may double down on the wrong answer even when challenged by users. The issue has been cited as both a practical annoyance and a broader concern about the reliability of AI tools as they consume increasing computational resources.