Anthropic has announced that its upcoming Claude AI models will have watermarking in their text, in accordance with the EU's AI Act's Transparency Code, as per a blog post from the company's website published on Sunday.
The watermarking system will also be implemented for older models, with the changes rolled out in the months to come.
The watermarking is invisible to the naked eye, but is created via a select pattern in its responses, which is used when it makes 'low-stakes' choices when generating words, which can be detected by the key that encodes that pattern. Those that possess the key will be able to verify whether or not the text was generated via Claude AI.
ALSO READ: Anthropic CEO Dario Amodei Defends AI Regulation As Shield Against Corporate Monopolies
The firm will also be releasing a watermark detection API and said it was in the midst of "working out the details of its implementation."
The company further stated that this watermarking method will not interfere with the quality of the AI's output. It will also not affect its speed and pricing, as the process does not involve the consumption of additional tokens.
Anthropic also stated that editing the text would not be enough to get rid of the watermark entirely, the firm stated that it would only be removed if every word in the text was replaced.
"In the latter case, of course, it's arguable whether the text can any longer be described as AI-generated," the blog post stated. The watermark also does not make distinctions between writing text from scratch and editing it as the output is generated in either case will contain it. This also applies to translations.
"It cannot distinguish 'Claude wrote this' from 'Claude heavily edited this.'", the post said.
Anthropic's watermarking is based on the 'Synth-ID Text' system developed by Google DeepMind, which it had recently open sourced for use by businesses and developers.
It differs itself from other AI detection programs like Pangram, which usually look for writing patterns that AI models tend to overutilise in the text that they generate. For example, overusing the phrase,“this isn't [X], it's [Y]” and using the word “quietly” quite a lot.
ALSO READ: Anthropic Eyes $6 Billion Decart AI Acquisition In Its Largest Known Deal
The watermarking system will also not affect the ownership of the output and the user's rights over it.
"It doesn't say anything about ownership or authorship, and doesn't change a user's rights under our terms. We only apply the watermark when Claude was involved in processing the content or file," the blog post said.
Essential Business Intelligence, Sharp Market Insights, Practical Personal Finance Advice, Daily Fuel, Gold and Silver Prices and Latest Stories — On NDTV Profit.
