Anthropic Will Now Watermark AI-Generated Content – Who Wins?

Anthropic will now watermark AI-generated content produced by its models, in order to comply with the EU AI Act that came into effect at the beginning of this month.

The general public have increasingly soured towards AI-generated content. While some studies suggest consumers don’t mind as long as content is relevant to them, watermarks could make AI content obvious and open creators up to criticism.

Users deserve transparency around what has been involved in creating a piece of content. Not all crimes committed with Claude are equal, but we could see users start to prioritize authenticity and originality when interacting online.

Anthropic Plans to Watermark AI-Generated Output from Claude Models

Leading AI lab Anthropic announced plans earlier this week to watermark AI-generated content, coming from its Claude models. The marking applies to models launched in the EU on or after August 2, and the company have plans to add the same feature to previous models.

The move will help Anthropic comply with the EU AI Act, which came into effect earlier this month. The regulation includes a Code of Practice on Transparency of AI-Generated Content.

As well as watermarking content, Anthropic said in a supporting post it would “support users and other third parties to detect Claude’s marks.” Watermarking is done in two ways, the lab further explained: embedded watermarks in text, and signed provenance metadata for files, including images.

A Potential Solution to Allegations of AI Use

Anthropic isn’t the only company to pledge to abide by the new European AI act. OpenAI, Meta, and Microsoft have all committed to this legislation, the first of its kind when it comes to policing AI models. So, while the terms may be different for each company, similar moves around transparency of AI-content will continue.

For the record, the watermarking of AI content isn’t a new thing. If you create an image using Google’s Gemini, for example, the lab’s logo appears in the bottom corner.

However, if all AI-generated, or I should say, AI-tampered content (based on Anthropic comments I’ll talk about later) is watermarked as more companies adapt to the new legislation, it will provide a new avenue of accuracy to detect AI content.

Consumers will receive this well, especially considered AI writing isn’t always easy to identify, and AI detectors aren’t always that accurate. Yes, tools like Pangram boast a false positive rate of only 1 in 10,000, but that hasn’t stopped Reddit users from complaining about their work being wrongly flagged as AI-generated.

Similarly, with clear indications of what content has traces of AI, users won’t be wrongfully pointing fingers. A recent study found a significant number of LinkedIn posts are AI-generated, which may or may not have led to LinkedIn adding a ‘Seems like AI slop’ button to the platform.

While the tool is useful at a glance, it leaves the user to decide whether content is AI-generated or not. It’s likely to deter some people from using AI in posts, however it’ll also mean those who aren’t using AI, but maybe enjoy the occasional em dash, are at risk of being penalized. A watermark in this case becomes a stamp of innocence, or an admittance of guilt.

An Uplift for Transparency in Today’s Content Landscape

Hopefully, AI companies identifying what content came out of their models will increase trust online. Users can interact with content generated by humans, if that’s their preference, and avoid content that doesn’t suit them.

However, as more and more people use AI in content, the practice will be normalized. What will be more important is how someone has used AI. In my eyes, not all crimes committed with Claude are equal, and there’s a big difference between light edits and a fully generated text.

In some cases, using AI tools like Claude gives people the option to communicate better and share their ideas. These people shouldn’t be penalized as much as the person plugging prompts all day, and copying and pasting to publish.

Unfortunately, it doesn’t look like the watermarking system at Anthropic will consider how AI was used. In its statement, Anthropic said: “People often use Claude to proofread, translate, summarize, or convert files. The output can carry a Claude mark even if the underlying ideas, text, or data originated from another source.”

Ultimately, users have a right to know if a piece of content is AI-generated. For brands operating online, being authentic while valuing the human voice might become a badge of honor.