Anthropic will soon add a hidden watermark to text generated by its AI assistant, Claude. The change aims to help companies and educators identify machine-written content.
How the watermark works
The watermark inserts a statistical pattern that specialized tools can detect. Anthropic states the method is highly reliable, though not perfect. If a student submits a Claude-written essay, detectors should flag it. The same applies to LinkedIn posts or internal memos.
The rollout date remains unannounced, but the company confirms the feature is coming. It follows other AI developers testing detection methods, though few have made them mandatory. OpenAI removed its own detection tool last year after reliability issues.
Related: Spin Master founder says 20s are best for startups
Attitudes toward AI-generated content
The announcement has renewed discussions about authorship and transparency. Some professionals still view AI-assisted writing as embarrassing. Y Combinator CEO Garry Tan adopted the label during a recent talk, admitting to using AI tools without hesitation. His response suggests workplace norms may be evolving.
Most AI writing occurs privately. Employees use these tools for emails, meeting summaries, and reports—tasks rarely scrutinized. For many, the reaction depends on context. A coworker discovering an AI-written Slack message may not care. The writing serves a purpose, not artistic expression. The same applies to customer service scripts or marketing copy, where efficiency matters more than originality. The focus has shifted from whether AI can write well to whether anyone notices or objects.
Some firms believe detection still matters. Pangram, an AI detection startup, is testing a Gmail integration that labels incoming emails as AI-generated. Users could filter these messages automatically. The feature, uncovered by researcher Jane Manchun Wong, awaits Google’s approval. Pangram’s CEO, Max Spero, confirmed to Business Insider that they are indeed hoping to launch it soon (pending some permissions with Google).
The challenge of detection
Watermarking and detection tools may trigger an ongoing battle. As AI improves, so do methods to hide its use. Some users already paraphrase or edit text to avoid detection. Anthropic acknowledges no system is flawless—errors will occur.
Related: Buffett’s Heir Starts Spending Berkshire’s Cash Mountain
The debate also questions whether detection is the right approach. If AI becomes as routine as spellcheck, will people still view it as a shortcut? Or will it blend into editing tools like grammar suggestions? Norms are still developing and may differ by field. A novelist using AI for ideas might face criticism, while a sales team generating emails likely won’t.
Academia and creative writing remain the most resistant areas. Professors are unlikely to allow AI-assisted essays, and readers may reject AI-authored books. For most writing—emails, memos, notes—the stakes feel lower. If watermarks improve detection, some users will adapt. Others will treat AI as just another tool.
Anthropic’s move does not resolve the debate. It forces the issue: if AI writing becomes harder to hide, will anyone keep pretending it’s a secret?
Leave a Reply