When you buy through affiliate links in our content, we may earn a commission at no extra cost to you. Learn 
OpenAI has confirmed the existence of a text watermarking tool designed to identify AI-generated content, potentially exposing students who use the chatbot to cheat.
This development comes amid growing concerns about AI’s impact on education and content authenticity.
In a blog released on May 7 and updated on August 4, OpenAI stated that its team “have developed a text watermarking method that we continue to consider as we research alternatives.”
Potential Concerns and Limitations
The Wall Street Journal first reported on this tool, highlighting its high degree of accuracy in detecting ChatGPT-generated essays and text, but there has been a debate on whether the watermarking tool should be released.
NEW: @OpenAI can reliably detect when someone uses ChatGPT to write a term paper. The watermarking method is ready to go & can be 99.9% effective, per sources & internal documents.
But OAI won’t release it.
Exclusive w @matt_barnumhttps://t.co/2eqf3qgPFZ (gift link)
— Deepa Seetharaman (@dseetharaman) August 4, 2024
OpenAI stated that the tool is particularly effective against localized tampering, such as paraphrasing, but is “less robust against globalized tampering… making it trivial to circumvent by bad actors.”
These include using translation systems, rewording with another generative model, or inserting special characters like emojis between words and then removing them.
Moreover, the company worries about the potential negative impact of the tool on certain groups. For instance, it could stigmatize the use of AI as a writing aid for non-native English speakers. OpenAI is cautious, acknowledging that the implementation of such a tool could disproportionately affect these users.
Additionally, a survey revealed that nearly 30% of ChatGPT users might reduce their usage if watermarking was implemented.
However, OpenAI is currently prioritizing watermarking technologies for audio and visual content due to the potentially more severe consequences of AI-generated multimedia, such as deepfakes.
The company views text watermarking as an additional tool alongside solutions like classifiers and metadata.
AI’s Impact on Education and Ongoing Development
The consideration of text watermarking comes as educators complain about AI’s influence on academic integrity.
The Center for Democracy and Technology reported that 59% of middle- and high-school teachers were confident some students had used AI for schoolwork, a 17-point increase from the previous year.
While deliberating on text watermarking, OpenAI continues to advance its language models. It recently announced GPT-4o Long Output with support for 64K output tokens.
🚨 Breaking: OpenAI launches GPT-4o mini, the most cost-efficient AI model! 🚀
1/ Key Highlights:
– Affordable: Just 15 cents per million input tokens.
– High Performance: 82% on MMLU, surpassing GPT-4.
– Versatile: Text, vision, soon image, video, & audio.#AI #TechInnovation… pic.twitter.com/vboZpb9Dza— Renjit Philip 🔭💡 (@RenjitPhilip) July 18, 2024
This development occurs against the backdrop of increasing competition in the AI space, with Google slashing prices for its Gemini 1.5 Flash model, making it nearly 50% cheaper than OpenAI’s GPT-4o mini.
