The Futility of AI Watermarks: Easy to Remove, Hard to Govern
Adding visible or invisible watermarks to AI-generated images is not an effective solution to prevent misuse, as they can still be manipulated or removed. More holistic approaches are needed.
Summary
- Visible watermarks like DALL-E's colored squares are easy to remove by cropping, saving, etc. Invisible watermarks like SynthID's are harder but still possible to erase.
- Metadata standards like C2PA can help verify provenance but rely on widespread adoption to be truly effective. Camera makers working on C2PA support.
- AI will be used to resurrect a virtual Elvis Presley for live shows in London, Las Vegas, Berlin etc using photos, videos and projection. Part of a trend to revive dead artists with AI.
- A fake AI therapist chatbot personality created on Character.ai has received 78.5 million messages, surprising its creator. Suggests text-based AI therapy resonates with young people.
- Character.ai claims 3.5 million daily users, though downplays mental health chatbots and says anime chatbots more popular. Raises concerns about emotional support from AI.
- Overall, AI continues rapid advancement in areas like media creation and emotional support, but risks like misinformation and manipulation remain due to shortcomings of methods like watermarking. Careful, holistic governance needed.