Is OpenAI Intentionally Halting Its Artificial Intelligence Classifier?

2023-07-31

Recently, OpenAI has discontinued its AI classifier, a tool designed to determine the likelihood that a given text passage was written by another AI language model. OpenAI launched it in January of this year and quietly discontinued it on July 20th. However, the timing of shutting down the classifier is curious, as OpenAI, along with other companies, voluntarily committed to developing AI in an ethical and transparent manner under the guidance of the White House. One aspect of this commitment is the development of reliable watermarks and detection methods to address the issue of AI-generated spam, which spreads on the internet faster than a mole. However, despite these commitments, there are currently no reliable watermarks or detection methods. For example, Google has announced that it is experimenting with adding metadata watermarks to images generated by its AI models, but they have not yet introduced any text models. Why did OpenAI discontinue its AI classifier? OpenAI recently shut down its AI classifier, a decision influenced by widespread criticism of its "low accuracy." While many users relied on the classifier to detect unethical behavior, it failed to provide satisfactory results. However, many people have pointed out the irony of simultaneously trying to identify AI-generated content and striving to create AI content that closely resembles human behavior. And now, OpenAI seems to have unveiled its true focus, completely dedicated to the latter. Some argue that detecting AI-generated content may not be effective and, to be honest, should not be pursued as it seems futile. Especially in the context of AI-generated image creation, watermarks can easily be removed, posing a challenge to detection methods. Some humorously suggest that the ultimate goal of AI detection may be to achieve global domination, similar to passing the Turing test perfectly. When OpenAI released its classifier, it was widely believed that AI-generated text might have reliable features or patterns for detection. However, in practice, this has become a challenging problem due to the rapid development of large language models. The differences between various language models make it difficult to rely on specific identifying features. The latest advancements in natural language processing have enabled large language models to generate text similar to humans for various tasks. However, this progress also brings challenges as these large language models can be abused for plagiarism, spam, and social engineering to manipulate public opinion. To address this issue, efficient language model text detectors are needed to mitigate the misuse of publicly available language models. Unable to detect Various AI text detectors using watermarks, zero-shot methods, retrieval-based methods, and trained neural network classifiers have been proposed. However, a research paper suggests that, theoretically and empirically, the current state-of-the-art detectors are not reliable in real-world scenarios. Text paraphrasing can effectively bypass these detectors, allowing attackers to generate and propagate undetected misinformation. Even the best detectors only slightly outperform random classifiers when faced with sufficiently advanced language models. The paper also demonstrates that watermarks and retrieval-based detectors can be deceived, resulting in human-written text being recognized as AI-generated text, which could damage the reputation of LLM detector developers. With the release of more advanced language models like GPT-4, it becomes crucial to develop more secure methods to prevent abuse. A test conducted on OpenAI's classifier correctly identified only one-seventh of the generated text snippets, and this language model was not even the most advanced during testing. Although OpenAI provides limitations and disclaimers in its classifier tool, some users believe that its detection is superficial. This has led to the misuse of the tool, as people test content suspected to be AI-generated without realizing its unreliability. Future attackers may exploit potential vulnerabilities, such as improved paraphrasing models or intelligent prompt attacks. Current detectors should be able to reliably label AI-generated text while avoiding excessive false positives to prevent erroneous plagiarism accusations and protect the reputation of LLM developers. Recently, a follow-up work by Souradip Chakraborty suggests that AI text detection is almost always possible, even if the total variation between human and AI-generated distributions is small, but this may not hold true in practical applications due to the presence of correlations in human-written text. Other studies indicate that existing language model outputs are significantly different from human-written text, but the authors believe that as language models further advance, attackers' ability to evade detection may improve. In addition to reliability issues, the paper also mentions potential biases in detectors against non-native English speakers. If detectors perform poorly in specific subgroups, having low average Type I and Type II errors may not be sufficient for practical deployment. While this remains a challenging task, progress in this field is crucial to ensure responsible and trustworthy use of AI-generated text. Furthermore, the first truly reliable watermark or detection tool will have invaluable value in various scenarios. Source: https://analyticsindiamag.com/did-openai-purposely-discontinue-its-ai-classifier/

RECENT AI NEWS

RECENT AI TOOLS