When an article reads a little too neatly, someone is bound to say: “You can tell AI wrote that.”
Add a few conversational phrases and change a handful of words, and another reader may decide it sounds entirely human.
Judging the origin of a text by its tone involves a great deal of guesswork. As Claude adds invisible text watermarks to outputs from some models, the discussion gains another dimension: text can carry a machine-readable signal about its origin. How much does that signal actually prove?
Anthropic says supported Claude models can embed imperceptible watermarks in generated text. The marks may survive copying and pasting, as well as light editing. The feature is being expanded in stages, while text-watermark detection is currently a private preview for eligible organisations.
The easiest misunderstanding comes immediately after a watermark is detected.
Anthropic explains that a marker can indicate that text was processed by Claude, but cannot establish its original author. A user might write a draft themselves and ask Claude to proofread or translate it. Conversely, substantial rewriting, translation, mixed text or a short passage can affect detection. An absent marker does not prove that a human wrote the text.
For businesses producing content, that distinction matters.
Imagine an engineer writing about a real project and then using AI to improve the wording. Another author enters a topic and publishes the resulting AI-generated article in full.
Both pieces may involve AI, yet the knowledge, verification and responsibility behind them can differ significantly. A company that reviews content solely for traces of AI may overlook the factors that actually determine its quality.
Readers trying to understand how a service works still need accurate information, clear explanations and someone willing to stand behind the content. A watermark can offer a clue about origin. Figures, examples and conclusions still need to be checked individually.
Content teams would benefit from retaining source material and editing records: which information came from interviews, which conclusions are the author's analysis, and whether AI helped with translation, organisation or an initial draft. Those records provide a concrete basis for answering questions from clients and readers.
A future article may begin with a human idea, pass through AI-assisted organisation and then be rewritten by an editor. As the process becomes more mixed, any single marker can answer only a limited set of questions.
Claude's watermark raises a practical editorial question: when AI participates in writing, have we kept enough evidence to explain how the article was produced?
And before someone presses “Publish”, has a person actually read it, checked it and agreed to take responsibility under their name?
Source: Anthropic | How Claude marks AI-generated content. This article includes independent editorial analysis by SRAM.
Author: SRAM