New Delhi: AI-generated content will soon be identifiable without the need for an AI detector, as Anthropic is set to begin watermarking content generated by its AI models. This global rollout complies with the European Union’s AI Act transparency code, enabling systems to detect AI-generated text without traditional third-party detectors.
“Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing,” Anthropic says on the Claude support page. “Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from.”
The new policy announcement about who should be credited for AI-generated work has sparked an online debate and backlash over Anthropic’s plan to introduce invisible watermarks as well.
“I had ditched Grammarly for Claude for proofreading because it does a better job,” radio show host and blogger Erick Erickson posted on X. “But now the stuff I’ve written will be watermarked to show that Claude did the work. This is ridiculous.”
The EU’s rules state that companies using AI must inform people when they are interacting with AI. AI-generated content should also carry watermarks or other labels. People must also be informed when they see deepfakes or when AI is used to recognise their emotions or categorise them based on biometric data.
“Claude markings cover output from supported models everywhere you use Claude, including Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag. Embedded watermarks will apply to all generated text. Provenance metadata will apply where Claude supports processing files,” reads a statement posted by Anthropic.
In its announcement this week, Anthropic also said that Claude models launched on or after 2 August would include watermarking. It will also add hidden watermarks to the text and to the SVG, PNG, and JPG files it creates. For text content, these watermarks will be invisible to people but detectable by computers.
“Anthropic’s whole new watermark scam. So if I wrote a whole blog and only asked Claude to do a proof check, Claude would add the watermark, claiming it did the work. How ridiculous is that?” wrote another user on X.
Also read: Mark Zuckerberg writes a 6,500-word essay on why everyone should have an AI assistant
Debate around credit
The watermarks will remain in the text even if it is copied and pasted into apps such as Windows Notepad or macOS TextEdit. In some cases, the watermarks may even remain after the text has been edited.
“We’re also working to enable users and other third parties to detect Claude’s embedded watermarks and provenance metadata. Detection checks whether a piece of text or a file carries a supported Claude mark. If a supported mark is found, it indicates that the content may have been processed by Claude,” the announcement further reads.
Claude will use two methods to mark the content it creates. For text, it adds a hidden watermark that people cannot see, and that does not change the meaning or quality of the text. For files such as .svg, .png, and .jpg, Claude adds special information called metadata. This metadata shows that the file was created or processed by Claude and can also help detect whether the file has been changed.
However, the backlash from users over the new development is largely focused on credit and authorship. Even when the content is written entirely by a human and AI tools are used only to make minor changes, such as proofreading or copy-editing, the text could still be watermarked as AI-generated. This is creating further debate among users, who argue that the watermark could incorrectly suggest that AI created the content when the actual work was done by a human.
“If I read Anthropic’s new watermarking policy correctly, not only will Claude watermark text it writes as AI (fair enough), but if I upload text I wrote and ask Claude to copy-edit it, Claude will now watermark my (human-written, AI-copy-edited) text as AI, which seems ridiculous,” another user wrote on X.
(Edited by Saptak Datta)