If your eye is trained for spotting AI-generated content - and it probably is, given how much of it you see online - you can probably recognize it right away. Of course, some AI content may be harder to spot, depending on how much effort a person who generated it put in, but most of it isn’t very skillfully camouflaged.
Beyond the visual test conducted by the human eye, there’s another identification layer that exposes AI-generated content. I’m talking about the watermarks, of course. Nearly every major company is putting some kind of watermark on content generated by their AI tools. They do this for various reasons, mostly legal ones; but also, for better transparency, and in the end, self-promotion.
Now, watermarking AI-generated images or text is straightforward. An LLM will just slap its logo in the corner of an image, making it instantly recognizable as being generated by that particular service. But did you know that AI companies are also watermarking text?
At first thought, it seems impossible to watermark text. You can’t put a physical watermark over it. Anthropic attaches signed metadata to actual files its models generate. But that’s only step one; what if you just copy the AI generated text, either from a file or straight from a chat interface, and paste it somewhere as plain text? There’s no way that can carry a watermark with it.
