OpenAI and Content Detection: Towards a New Era of Transparency?

OpenAI is working on a tool to detect text generated by ChatGPT. Despite a claimed 99.9% effectiveness, this tool remains confidential, sparking debates and concerns. Here's an overview of the reasons for this decision and its potential implications.

OpenAI's Detection Tool: A Promising Innovation Awaiting Release

For a year, OpenAI has been developing a detection tool designed to identify text generated by ChatGPT with a claimed accuracy of 99.9%. However, this content detector remains inaccessible to the public. According to the Wall Street Journal, although the tool has been ready for about a year, OpenAI hesitates to deploy it. Following this report, the company confirmed the tool's existence and its limitations while stating that it continues to work on the matter.

OpenAI's Concerns and Limitations

One of OpenAI's main concerns is that users might turn away from ChatGPT if they knew it was easily detectable. An internal survey revealed that this technology could deter a third of users, especially in educational, academic, and professional contexts. ChatGPT does not watermark its content with a "digital watermark" or "watermark." However, this would change with the detection tool, which uses invisible watermarks that can be detected and attributed to ChatGPT with a 99.9% probability when the text contains enough words.

Nevertheless, there are ways to bypass these watermarks, such as translating the text via Google or adding special characters before removing them. Moreover, over time, the watermarking method could be discovered, raising other concerns, including potential discrimination against non-English-speaking users.

Detection and Evasion: A Technological Challenge

OpenAI is not the first to face this challenge. In the past, the company tested another detection tool without watermarks, which had a success rate of only 26%. Other companies are also working on watermarking technologies. For example, Google has launched its own tool, SynthID, to detect texts generated by its AI Gemini. Other solutions like Originality AI, Content at Scale, and ZeroGPT are attempting to tackle this challenge, but the question remains: are these tools truly reliable?

In my opinion, all AI-generated content detection tools are circumventable due to user creativity and the immense number of users. It is likely that explanatory videos will flourish on YouTube showing how to bypass these systems. An ethical question also arises: should we always disclose who created the content? Furthermore, detection concerns only fully AI-generated texts, but if a text is modified and adapted by a user, what will the detection rate be then? The reliability of these tools remains uncertain in the face of numerous strategies to obscure the origin of content.

Towards a Future Defined by Digital Watermarks

The rise of digital watermarks as a standard for tracking AI-generated works seems inevitable. Despite technological and ethical challenges, OpenAI and other industry players are working tirelessly to provide effective solutions. In the near future, these tools could become standards to attempt to ensure greater transparency and accountability in the use of generative AIs.

By Brice Matter

____________________________

To stay updated on generative AI news, subscribe here to the Gennn Newsletter!