Claude's Watermark Proves It Was Used, Not Who Wrote It

For the past three years, school leaders have kept asking me some version of the same question: how can we know if a student used AI?It sounds reasonable. But it’s the wrong place to begin. Schools want a clean technical answer to what is really a messy question about human behavior, assessment design and trust. Now Anthropic has given them something that may look very much like that answer.On August 11, Anthropic quietly updated a Claude Help Center article to confirm that text produced by supported Claude models will carry an invisible, machine readable watermark. It will survive copying and pasting. It may survive editing. It will appear whether someone uses Claude itself, Claude Code, Claude Cowork, Claude Tag or the Claude API. The same applies when supported models are accessed through AWS, Google Cloud or Microsoft Foundry.There is no switch to turn it off. The marking happens inside the model.I can already see where this is heading in education. A teacher checks an essay. Claude's mark is found. The result gets described as proof that AI wrote the work. A misconduct process begins.Except that is not what the mark proves. Why Anthropic Is Doing This NowThe change follows the arrival of Article 50 of the European Union's AI Act, which took effect on August 2. Providers of generative AI systems must make AI generated or manipulated content detectable through machine readable marking. Failure to meet the rules can lead to fines of up to 15 million euros or 3% of global annual turnover.Anthropic has signed the associated transparency code. So have Google, Meta, Microsoft and OpenAI. In total, about 190 organizations had signed by the end of July. New Claude models launched on or after August 2 support marking from launch. Anthropic says it is working backwards to add the same capability to earlier models.None of this required Anthropic to mark text produced outside Europe. It chose to apply the system wherever Claude is available. What...