Anthropic to Add Invisible Markers to Claude

Anthropic plans to integrate invisible markers into texts produced by Claude to comply with new European regulations, with detection via a dedicated tool. In practice, a history professor in Mississippi describes another approach: a hidden instruction in an assignment that caused the word "Madagascar" to appear in AI-generated papers, making his demonstration go viral.

Anthropic announces invisible signals for its texts in Europe

Anthropic announces that it will begin inserting invisible signals into the texts produced by its artificial intelligence to comply with new European regulations. According to the company, these markers, when paired with a detector, will allow identification of whether content originates from or has been modified by Claude.

A hidden instruction triggers a keyword in AI responses

On TikTok, a teacher describes a method of embedding an invisible instruction in white font on a white background within the assignment prompt. In an exercise on the Industrial Revolution, the instruction asked students to insert the word "Madagascar" in a nonsensical manner. It does not appear on students' screens unless selected, but it is read by a chatbot if the full prompt is copied and pasted. The presence of the term "Madagascar" in papers then served as a clue to identify AI-generated responses.

Virality on TikTok and exam failures reported by the teacher

The video posted by Jason Gibson, a history professor at Alcorn State University in Mississippi, has been viewed over 2 million times on TikTok. In this same testimony, he states that thirty-two of his thirty-five students, spread across two classes, failed part of their midterm exam, all having, according to him, relied on AI to write their entire responses.