A naukri.com initiative
Analyticsindiamag
1M
36
Image Credit: Analyticsindiamag
Trying to Watermark LLMs is Useless
- Watermarking AI-generated content has become a popular method to identify language model outputs (LLMs) but it is flawed and essentially useless.
- The goal of watermarking, often misunderstood, is to identify text from a specific model but it cannot distinguish reliably between AI-generated and human text.
- Many companies have launched their own watermarking tools however, research from Carnegie Mellon University shows that watermarking cannot deal with AI-driven misinformation.
- Watermarking raises several difficulties including compatibility concerns with temperature settings and a total lack of accuracy and is therefore, not enough to keep pace with AI or reduce its spread.
- The problem with this approach is that only capable LLMs can be watermarked, and disingenuous actors will always access unmarked models, despite even the tiniest of hindrances in watermarked models.
- Dominik Lukes, lead business technologist at the AI/ML support competency centre at the University of Oxford claimed “Outside a school exam, the use of an LLM is no longer a reliable indicator of fraud.”
- Watermarking suffers from several restraining factors like robustness and detection difficulty, tempting bad actors to use open-source models for privacy, rendering API-based watermarking ineffective.
- Furthermore, authors may edit AI-generated text created with the help of LLMs, making the necessary distinction between human-written and AI-written text even harder.
- Even an LLM based on a model that has abolished the API will not be enough to avoid detection by AI-detection technology.
- Trivially, if watermarking would miraculously work, it would not solve the problem as not all AI-generated text is harmful and AI-generated and human-written texts are often intertwined.
Read Full Article
2 Likes
For uninterrupted reading, download the app