OpenAI has developed a powerful tool capable of detecting GPT-4’s writing output with an impressive 99.99% accuracy, offering a potential solution to the growing problem of plagiarism in the age of AI. However, the company has been reluctant to release this tool to the public, citing concerns about its potential impact on the broader ecosystem.
OpenAI acknowledges the complexities involved and the potential for unintended consequences. The company is particularly worried about the susceptibility of its watermarking system to circumvention by malicious actors. They are also concerned that the tool could disproportionately impact groups like non-English speakers.
The text watermarking system works by subtly incorporating a specific pattern into the model’s written output, invisible to the end user but detectable by OpenAI’s tool. Despite its effectiveness in identifying GPT-4 generated content, this method cannot detect outputs from other large language models like Gemini or Claude. Moreover, the watermark itself can be easily removed by a simple process of translating the text to another language and then back, rendering the tool ineffective.
This isn’t OpenAI’s first attempt at building a text-detection tool. Last year, they quietly abandoned a similar detector due to its low accuracy and high rate of false positives. This earlier tool required users to manually input at least 1,000 characters of text before making a determination, and it achieved only a 26% accuracy rate in correctly classifying AI-generated content. Furthermore, it incorrectly labeled 9% of human-generated content as AI-derived, leading to a situation where a Texas A&M professor mistakenly failed an entire class for allegedly using ChatGPT on their final assignments.
OpenAI’s hesitation to release the new tool stems partly from a fear of user backlash. A Wall Street Journal survey revealed that 69% of ChatGPT users believe such a tool would be unreliable and likely result in false accusations of cheating. Another 30% reported they would abandon ChatGPT for a different model if OpenAI were to release the feature. Additionally, the company is concerned that developers could reverse engineer the watermark and create tools to negate it.
While OpenAI continues to weigh the merits of releasing its watermarking system, other AI startups are moving forward with their own text detectors, including GPTZero, ZeroGPT, Scribbr, and Writer AI Content Detector. However, the overall lack of accuracy in these tools leaves the human eye as our best defense against AI-generated content, a reality that offers little reassurance in the ongoing battle against plagiarism in a rapidly evolving technological landscape.
