"Was this written by a human?"

As generative AI becomes a common writing tool, that question is becoming harder to answer.

US company Anthropic is introducing invisible watermarks into text generated by its Claude models, a move aimed at making AI-written content easier to identify.

The watermark cannot be seen by readers and is designed not to change the meaning or quality of the text. Instead, Claude slightly changes how it selects words as it generates a response, creating a hidden pattern that detection tools can recognize.

The move comes as transparency rules under the European Union's AI Act took effect on August 2. The rules require providers of certain AI systems to make generated or manipulated content detectable in a machine-readable form.

Anthropic has said the watermarking will be built into Claude at the model level and applied globally.

The technology could change how AI-generated writing is identified.

Most AI detectors currently examine a finished piece of text and look for features often associated with AI writing, such as patterns in word choice and sentence structure. They then estimate how likely it is that AI was used.

But current detection software is not 100% accurate.

A recent dispute over the bestselling novel Daggermouth highlighted that problem. A study that has not yet been peer-reviewed analyzed more than 14,000 e-books and estimated that about 60% of the novel's text was AI-generated.

The author denied using generative AI, while publisher Simon & Schuster said it stood behind the work.

Watermarking takes a different approach. Instead of trying to determine where a text came from after it has been written, the AI system leaves a hidden signal while generating it.

That could give publishers, schools and online platforms another way to check whether AI was involved in producing a piece of writing.

However, watermarks are not foolproof.

Anthropic has acknowledged that extensive rewriting, translation or mixing Claude-generated text with other material could make the watermark difficult or impossible to detect.

The absence of a watermark also does not prove that a text was written entirely by a human. A piece of AI-generated content may have been heavily edited, produced by a system without such a watermark, or had its marker removed.

For images and other files, Anthropic is also adopting C2PA, a standard that can carry information about how digital content was created. Such information, however, can sometimes be lost when a file is edited, converted or uploaded to another platform.

Anthropic is not alone in trying to make AI-generated content easier to trace.

Google has expanded its SynthID watermarking technology across AI-generated text, images, audio and video. OpenAI and Meta have also explored watermarks and other tools that can indicate when AI has been involved in creating digital content.

Regulation is adding urgency to these efforts. Under Article 50 of the EU AI Act, providers of generative AI systems must make certain AI-generated or manipulated content detectable where technically feasible.

For most users, invisible watermarks may make little difference to how they use an AI chatbot. They can still ask it to draft an email, summarize a report or help write a story.

What may change is how easily that use can later be traced.

But watermarking cannot answer every question about AI authorship. People increasingly use AI not only to generate entire texts, but also to edit, translate, summarize or brainstorm.

As human and AI writing become more closely mixed, the larger challenge may be not simply deciding whether something was "written by AI," but understanding how AI was used and who is ultimately responsible for the final content.