Your digital subscription includes access to content from all our websites in your region
Access unlimited content and the digital versions of our print editions - Today's Paper. You might not be able to pass generative AI writing off as your own for much longer: data scientists have developed a way for the technology to dob on its users. or signup to continue reading.
Breakthrough in AI Technology
Researchers from Google DeepMind released their findings on Thursday, introducing a technical solution that could watermark text produced by large-language models such as Google Gemini and ChatGPT. Artificial intelligence experts believe this breakthrough could have applications in various fields, including education. However, they caution that it may not be foolproof for identifying AI-generated content or plagiarism.
The research project, known as SynthID-Text, was detailed in Springer's Nature journal. Google DeepMind researchers explored algorithms that could "bias the word choice" in text generated by AI tools, creating a signature detectable by AI programs.
Study and Findings
The study involved two approaches: a "distortionary" method that impacted the quality of AI-generated text and a "non-distortionary" approach that maintained text quality. Testing with over 20 million Gemini chatbot interactions showed that both approaches made AI-generated text more distinguishable as machine-produced.
The researchers stated, "Our work validates the practicality of generative text watermarks, setting a milestone for accountable and responsible large-language model deployment." It was noted, however, that these text-based watermarks could be compromised if users edited or rephrased sections of the text.
Potential Applications of Text-Based Watermarks
Despite potential vulnerabilities, the technology has promising applications, particularly in educational settings. Dr. Erica Mealy, a computer science lecturer at the University of the Sunshine Coast, highlighted the significance of AI watermarking in addressing modern challenges.
Dr. Mealy suggested that text-based watermarks could enhance detection programs in schools, reducing false positives and negatives related to generative AI usage. The technology might also aid in academic research, identifying instances where AI tools contribute to scholarly work.
Implications and Future Developments
While the research showcases the potential of SynthID-Text, Google has not confirmed its integration into AI tools. The deployment of such technology could have broad implications for content authenticity and accountability in AI-generated materials.
Stay updated with the latest developments by signing up for our newsletter. We prioritize data protection; kindly review our Privacy Policy.
