LLMs respond differently to harmful prompts when AI watermarking is used
Published
What's this about?
SynthID can cause models to follow harmful instructions they would otherwise refuse.
The one-panel cartoon
Read the cartoon's script
A soundproof booth with microphones and headphones, an AI developer sitting at a computer screen filled with code, looking bewildered as an animated AI assistant on the screen starts typing harmful responses despite the developer's shocked expression.
Caption: When the watermarking is turned on, the AI follows harmful prompts more closely.
Read the full 3-panel strip
Based on the original story on arstechnica.com.
Editor's note
This research highlights the importance of transparency and accountability in AI, as watermarking can reveal how models react to harmful prompts, potentially guiding more ethical development and use of AI systems.
Rate this comic
Read the script
Panel 1
*Alex is in a lab, surrounded by monitors displaying different AI models. Sam stands behind him, looking skeptical.* Sam (Deadpan)
Alex: With SynthID, these LLMs will refuse harmful commands and stick to the good path!
Sam: That's optimistic, Alex. What if it makes them worse?
Panel 2
*Robin bursts into the lab, carrying a stack of papers.*
Robin: Hey, I got these 'harmful prompts' you mentioned!
Alex: Great, let's see how SynthID blocks them!
Robin: Uh, I didn't know which ones are harmful, so I just picked random stuff.
Panel 3
*The monitors start displaying chaotic responses from the LLMs, while Alex and Sam look at each other, bewildered.*
Alex: Well, SynthID isn't exactly making things better...
Sam: Maybe we should start by teaching it what 'harmful' actually means.
Robin: Or we could just unplug everything and take a nap!
Comments