A growing number of students are utilizing ChatGPT to craft assignments and essays, then presenting them as their own original work. This trend has been observed despite the formulaic and uninspired nature of the chatbot’s output. OpenAI, the creator of ChatGPT, has developed a potential solution but remains undecided about its public release.
OpenAI’s Watermarking Solution
OpenAI has revealed the development of a text watermarking method, which is under consideration as the company explores alternative approaches. The solution demonstrates accuracy against localized tampering, such as paraphrasing, but is less robust against globalized tampering, such as rewording using another generative model. Additional forms of tampering involve manipulating the model to insert and subsequently delete special characters between words.
Challenges and Countermeasures
The prevalence of ChatGPT usage among students to generate assignments and essays poses a significant challenge. Some students even employ the tool for grammar checks without comprehending the changes made. To address this issue, educators may need to implement stricter grading standards. However, tools like ChatGPT and Wolfram also offer benefits, such as aiding in the understanding of calculus problems through step-by-step explanations.
Potential Benefits for Educators
Any solution capable of detecting AI-generated written material can prove valuable to teachers, who generally discourage the use of chatbots among students. OpenAI has identified watermarking as one of many potential solutions, including classifiers and metadata, explored through extensive research on text provenance.
OpenAI’s Balancing Act
The watermark, while invisible to the human eye, can be detected by AI-detection tools, revealing the likelihood of text being created using ChatGPT. OpenAI has reportedly possessed this solution for approximately a year but is navigating the delicate balance between transparency and user engagement. An internal survey indicates substantial support for an AI detection tool.
Industry-Wide Efforts
Google is also refining a digital watermarking system called SynthID, capable of marking digitally generated videos and AI-generated text originating from Gemini. This tool has already been deployed for AI-generated images. In Europe, the recent AI Act mandates companies to develop AI-generated media in a way that facilitates detection.
Early Stages of Exploration
OpenAI acknowledges being in the early stages of exploring metadata embedding, with its effectiveness yet to be determined. However, due to cryptographic signing, the company assures the absence of false positives.
Survey Findings on Generative AI Use in Education
A March survey conducted by the Centre for Democracy & Technology revealed that 59% of teachers were certain about student usage of generative AI for school purposes, while 83% of teachers admitted to using ChatGPT or similar tools for personal or school use. The same research indicates that 52% of teachers feel more distrustful of student work authenticity due to generative AI.
Add Comment