tech
Why AI Watermarks and Detectors Could Backfire
Claude now watermarks AI-generated text to comply with European Union transparency rules. OpenAI and Google add invisible fingerprints to AI-generated images. And Substack is touting a feature that scans pieces for signs of AI. Will we finally be able to tell what’s real on the Internet? My take: not even close.
TL;DR
- AI watermarks and detectors are being implemented for transparency, but are easily circumvented.
- These tools can create a false sense of security, making users believe content is genuine if it lacks a watermark.
- Past methods of detecting fake online content based on visual clues have failed as technology advanced.
- Open-source AI models and workarounds like humanizer tools make it difficult for detectors to remain accurate.
- The inverse illusion, believing the absence of a signal proves the opposite, is a key problem for AI detection.
- Focusing on the reputation and context of information sources is a more reliable method for determining credibility than AI detectors or visual cues.