ChatGPT still hasn’t fully recovered from its letter counting mistakes

It was noteworthy that ChatGPT fixed a simple bug that had long been the subject of ridicule. The artificial intelligence chat bot can now correctly answer the question of how many letters “R” are in the word “strawberry”, which means “strawberry” in English. Despite this, the system’s tendency to confidently produce false information has not disappeared.

One of the most frequently criticized aspects of big language models is that they produce false information but present it with high confidence. When users point out the error, it seems that the model continues to defend the wrong answer instead of taking a step back. This situation both increases the reliability debate and points out the problems these vehicles create in daily use. It is especially noteworthy that even simple logic or basic grammatical errors appear in this way.

One of the most well-known examples of this situation for ChatGPT, developed by OpenAI, was the number of letters “R” in the word “strawberry”. For a long time, the chatbot could not correctly detect that there were three “R”s in this word, and often argued with users, defending its incorrect answer. Similarly, the same types of errors were encountered in different artificial intelligence models.

In its post on the X platform, the company announced that ChatGPT now provides the correct answer to this question. Likewise, it is observed that more consistent answers are given to basic logical questions such as whether to walk or drive to a car wash within a short distance. Such examples reveal that the system is improved in certain scenarios.

Why does the problem of incorrect answers persist in artificial intelligence models?

However, it is thought that these corrections may have been specifically addressed for certain examples rather than a radical solution. In users’ posts, it seems that similar questions are still answered incorrectly. For example, it is stated that ChatGPT can still give an incorrect answer when asked about the number of letters “R” in the word “cranberry”. This suggests that improvements may be limited adjustments rather than a generalized logic development.

In addition, adding special solutions to certain questions, called “hardcoded” in artificial intelligence systems, is not a new approach. However, it is known that this method does not improve the general reasoning ability of the model and causes similar errors to be repeated in different examples. This creates an inconsistent picture in terms of user experience.

In addition to all this, the increasingly widespread use of artificial intelligence tools increases the impact of such errors. Wrong but convincing answers can create a serious problem for users who rely on these systems in their daily information acquisition processes. Despite this, developers continue to improve both training data and control mechanisms to improve model accuracy.