Anthropic adds invisible watermark to Claude AI-generated text

Some computer programs can write sentences just like people.

A company called Anthropic makes one of these programs, named Claude.

Sometimes it is hard to tell whether a person or a computer wrote something.

Anthropic has found a clever way to help: it hides a secret, invisible pattern inside the words Claude chooses.

You cannot see this pattern, like invisible ink.

When Claude writes, it often picks one word out of several that would all fit, and the hidden pattern guides which word it picks.

Later, special tools can check for the pattern and guess whether Claude wrote the text.

The pattern stays if you copy the text or fix small mistakes, but it can be lost if someone rewrites the whole text.

The mark does not tell who used Claude, and text without the mark is not automatically human-written.