OpenAI says it has a new tool to watermark text produced by AI models like ChatGPT, addressing academic dishonesty by flagging AI-created content. OpenAI initially discussed the tool in a blog post in May, and the Wall Street Journal spotted the reveal.
The watermarking method introduces tiny word choice changes in ChatGPT’s output, embedding an undetectable signature. Although it’s effective against basic alterations, it struggles with more advanced modifications, such as using translation tools or rephrasing with different AI models. OpenAI acknowledges the tool’s limitations and the possibility of sophisticated techniques to bypass it.
Ethical Considerations and Wider Implications
The company is being cautious about deploying this tool, considering the ethical ramifications. A key concern is its effect on non-native English speakers who use AI for writing help.
Besides watermarking, OpenAI is working on verifying audiovisual content authenticity and exploring alternatives like classifiers and metadata. These initiatives aim to boost the transparency and dependability of AI systems.
Recent Developments and Competition
OpenAI has recently added features to its ChatGPT model, including memory functions for select users and a voice mode for ChatGPT Plus subscribers. Integration with Google Drive and OneDrive now allows the model to create interactive charts and tables. Meanwhile, Google has cut the price for its Gemini 1.5 Flash model, heightening competition in the AI field.
The Wall Street Journal notes that OpenAI is still debating the release of the watermarking tool. The complexities and potential impact on the ecosystem are prompting a careful approach. This tool is designed to identify text written specifically by ChatGPT, not other AI models. OpenAI previously shut down an AI text detector due to inaccuracy but notes improvements in the watermarking process can better handle localized tampering.
For end-users, these advancements mean better tools for identifying AI-generated text and enhanced features in AI models like ChatGPT. As AI technology progresses, companies like OpenAI and Google are looking to offer more trustworthy solutions.
Last Updated on November 7, 2024 3:24 pm CET
