Making Tough Choices: Watermarking Of AI Text Might Increase The Chances Of Getting AI Hallucinations

In today’s column, I examine an interesting twist associated with the latest trend of AI makers opting to include text-based watermarking in their latest versions of generative AI and large language models (LLMs). The twist is that new research suggests that there is a chance of increasing the odds of getting AI hallucinations due to the efforts of the watermarking process. Yes, believe it or not, the internal efforts of devising hidden text-oriented watermarks seem to also come with the potential for generating AI hallucinations.

You might already know that AI hallucinations are the bane of AI makers and all users of generative AI. An AI hallucination is when the AI generates a made-up, ungrounded response to your prompt. For example, you might ask when Halloween takes place each year, and the AI responds that it occurs annually on October 13 (the correct answer is October 31). The AI has produced a confabulation, something that is not based on actual facts. All users of AI need to be alert to the possibility of AI hallucinations. Now, the advent of text-based watermarking could worsen the AI hallucinations predicament by providing a new avenue for producing unwanted and potentially unsafe AI hallucinations.

Let’s talk about it. This analysis of AI breakthroughs is part of my ongoing Forbes column coverage of the latest in AI, including identifying and explaining key AI complexities (see the link here).

Watermarking Is Challenging

First, some foundational aspects of the topic of watermarks (for my in-depth analysis of AI watermarking, see the link here). We are all aware of watermarking when it comes to paper-based materials and likewise for any tangible artifact that exists in a definitive physical form. A dollar bill can contain a watermark, allowing the naked eye to tell whether it is real or counterfeit. Watermarks can also be hidden from visual inspection, requiring some other means to detect the watermark.