The developers of ChatGPT, OpenAI, have announced that they are removing the AI classifier from their chatbot service due to significant inaccuracies. The online tool, which was used to determine if text inputted into ChatGPT had been generated by other AI text generators, is no longer available for use.
Previously, individuals interested in verifying if an AI-powered program had generated the text they were reading could utilize the AI classifier on OpenAI’s website. This tool was especially useful for checking the authenticity of text-based content such as emails, blog posts, and essays.
OpenAI acknowledged the lack of accuracy in their AI classifier, admitting that it would sometimes make incorrect predictions with great confidence. In certain cases, it flagged content as AI-generated when it was actually created by a human. OpenAI emphasized that while the tool could assist in differentiating between human-written and AI-generated content, it should not be the sole determining factor.
The AI Text Classifier was trained on human-written text from various sources, but OpenAI acknowledged that this dataset might not represent all types of human-written text. They cautioned that the classifier had not been thoroughly evaluated for specific targets such as student essays, automated disinformation campaigns, or chat transcripts. Therefore, its predictions may not be reliable outside of its training data.
OpenAI issued a warning about the performance of neural network-based classifiers, stating that they are often poorly calibrated beyond the data they were trained on. For inputs significantly different from their training data, the classifier could make highly confident yet incorrect predictions.
Given these major concerns about accuracy, OpenAI made the decision to remove the AI classifier from ChatGPT. The developers aim to foster conversations about distinguishing between human-written and AI-generated content but recognize the limitations of their current approach.
In summary, OpenAI has removed the AI classifier from ChatGPT due to the tool’s inaccuracies. While the classifier could detect AI-generated text, its performance was often unreliable and sometimes flagged human-generated content as AI-generated. OpenAI acknowledges the need for further evaluation and caution when determining the origins of text.