Claude is getting ambitious with watermarking, and I can smell the problems from a mile away
Anthropic is trying to solve the problem of identifying AI-generated text, but a persistent watermark could create another one if AI assistance is mistaken for AI authorship.
Anthropic is testing an invisible watermark that can be embedded directly into text generated by Claude, a large language model. The company believes this could help identify AI-generated content, which is prevalent in today's digital landscape. By altering how Claude selects words to create patterns, the watermark aims to be detectable later.
However, there's a concern about how persistent the watermark would be once the text has been modified. For instance, if someone writes an essay in Spanish and asks Claude to translate it into English, the resulting text might still carry the watermark. The same issue could arise with proofreading or shortening a paragraph. While Anthropic clarifies that the watermark indicates Claude's involvement, not the creation of the original work, the distinction might be difficult for users to comprehend, especially when dealing with AI detection systems that have known issues.
The potential for misuse is high, as students may turn to AI humanizers to evade detection, further complicating the situation.
Written by urgent.news from Digital Trends's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
