Text-watermarking is now added to ChatGPT after Claude. Who is really benefiting from it?
A couple of weeks ago, Anthropic announced that Claude-generated texts will now have invisible watermarks, which caused quite an uproar. Meant to label synthetic content and for professors and teachers to track AI usage across assignments, the response to watermarking was divided. And it still is.
OpenAI is following through on that, even though it’s mostly to abide by the EU’s AI Act.
ChatGPT and Codex will now possess OpenAI’s text watermarking technology, textGrain. This will add invisible, machine-readable watermarking to the generated content/texts.
According to the announcement made by OpenAI, this isn’t the global default and is merely applicable to users based within the EU.
Some part of the market was hopeful. Readers and platforms should know when content is AI-generated, right? For transparency and ownership. Owing to this, even Substack now has Pangram added to its platform for users to gauge if a piece or any part of it is AI-generated.
But the benefits aren’t entirely limited to users. There’s another angle to the reformation: help identify recursive content for future AI models and filter out low-quality recurrent information.
Much of the conversation around AI-generated text has boiled down to its quality. AI companies had an idea- let’s focus on feeding AI only high-quality content. Not AI content fed back into AI models.
Now there’s the other side of the market, which isn’t happy.
Their reigning concern is: what if human work (that’s merely proofread or edited slightly by AI) is also marked as AI-generated?
These watermarks aren’t a measure of accuracy or ownership, and neither do they indicate how much of the content is truly human. There’s no variable to truly differentiate. And if wrong, can lead to unfair penalization.
OpenAI’s only advice for now? Publishing a disclaimer.
