Anthropic adds invisible watermarks to Claude that track copy and pasted text
An AI company has rolled out a new watermark system that will impact copy and pasted text from the model.
Antropic introduced the move to its Claude models at the start of August after a ruling under the European Union (EU) was brought into effect.
Under the new AI Act, the Commission, which is the executive arm of the EU, will now inspect AI models that are being introduced to the market.
The act came into effect on August 2, with providers of models put on the market having until August 2, 2027, to comply with new obligations.
Advert
The Council of the European Union describes the new act as being a way to ensure that AI systems are ‘developed and used responsibly’.
It goes on to say: “The rules impose obligations on providers and deployers of AI technologies and regulate the authorisation of artificial intelligence systems in the EU single market.
“The law addresses risks linked to AI, such as bias, discrimination and accountability gaps, promotes innovation and encourages the uptake of AI.
“As the world’s first law regulating AI, the EU’s rules could set a global standard in AI regulation, just as the general data protection regulation (GDPR) has done for data privacy, promoting ethical, safe, and trustworthy artificial intelligence worldwide.”
So, what does this mean for the likes of Anthropic’s Claude? Well, the AI firm will now place invisible watermarks that will track text that copy and pasted.
This decision has been made in a bid to comply with Article 50 of the new AI Act which lists transparency requirements.
However, despite the ruling pertaining to countries in Europe, the watermarks will apply for users around the world.
While not all AI models currently use watermarks, it is expected to become more commonly used in the future as more legislation is outlined for the technology.
Google’s Gemini and Anthropic’s Claude are two examples of firms adding a digital fingerprint to their models and while ChatGPT does not yet include its own watermark, it does include hidden Unicode space characters which can act as a fingerprint.
Under the new ruling from the EU, there will be four levels of risk with corresponding regulations, including minimal to no risk, limited risk, high risk, and unacceptable risk.
According to the council, high risk AI systems include those that are ‘used in disease diagnoses, autonomous driving and the biometric identification of individuals involved in criminal activities or investigations, must meet strict requirements and obligations to gain access to the EU market’.
The pressure appears to be ramping up for other AI companies to abide by these new rules or risk being shut out from the entire European market.