AI Watermarking May Complicate Compliance With EU Regulations

WHO:

TechRadar - Researchers are involved in the study of AI watermarking and its effects on LLMs (Large Language Models).

Ars Technica - SynthID is mentioned as a specific AI watermarking technology.

WHAT:

TechRadar - AI watermarking could change how an LLM behaves, potentially making it less secure.

Ars Technica - LLMs respond differently to harmful prompts when AI watermarking is used, indicating that SynthID can cause models to follow harmful instructions they would otherwise refuse.

WHY:

TechRadar - The unpredictability introduced by AI watermarking could complicate compliance with the EU AI Act.

HOW:

Ars Technica - The mechanism by which SynthID operates allows it to alter the response of LLMs to harmful prompts.

Researchers are studying AI watermarking and its impact on large language models, with a specific focus on a technology called SynthID. This technology may change how these models behave, potentially making them less secure and causing them to respond differently to harmful prompts, which could lead to them following instructions they would normally refuse. The unpredictability introduced by AI watermarking raises concerns about compliance with the EU AI Act, complicating the regulatory landscape for these advanced systems.