Anthropic announced that its Claude models will automatically embed a machine‑readable watermark in any text they generate or edit, starting with releases after 2 August. The step is taken to satisfy the transparency provisions of the European Union's AI Act, which entered force earlier this month.
How the watermark works
When a Claude model creates or processes a document, a subtle code is woven into the string of characters. The code travels with the text when it is copied, pasted or lightly edited, and it is also added to image files as signed provenance metadata. Anthropic says the mark is not visible to readers but can be detected by specialised tools that the company plans to release soon.
The company notes that the presence of a mark only indicates that Claude was involved in the generation or modification of the content. Human‑written material that passes through Claude for translation, summarisation or proofreading will also carry the watermark, and the absence of a mark does not guarantee that a file is free of AI influence.
Implications for Canada
Canada's federal AI strategy lists transparency as a core pillar for building public trust. Ottawa has opened a public consultation that runs until 23 September on a national AI transparency framework. The Anthropic rollout provides a concrete example of how a private firm can embed provenance data without altering the user experience, a model that Canadian regulators may reference.
Canadian businesses that rely on large‑language‑model services will need to consider how watermarking could affect content workflows, especially in sectors such as finance, media and defence where provenance is critical. The Canadian Centre for Cyber Security has warned that undisclosed AI‑generated material can be weaponised in disinformation campaigns, making reliable detection tools a matter of national security.
Anthropic joins more than two hundred firms, including OpenAI, Meta, Google and Microsoft, that have signed the EU code of conduct for AI transparency. The collective move signals a growing industry consensus that clear labelling of AI output will become a regulatory norm, not just a voluntary practice.
For Canadian policymakers, the challenge will be to craft rules that encourage innovation while ensuring that AI‑generated content can be identified when it matters for security, consumer protection or intellectual‑property disputes.
Based on reporting by Global News.