Can AI Text Be Caught? I Built a Watermark Detector and Tried to Break It

OpenAI says it will watermark ChatGPT text by default in the EU, and Anthropic says Claude's text is already watermarked worldwide. So can AI-written text actually be caught? I explain how a text watermark works (green and red word lists plus a z-score), go through OpenAI's own numbers, then build a small watermark and detector in Python and try to break it by editing the text. What I found: my toy detector flagged untouched text at z = 10.47. After I swapped 25% of the words it still caught it (z = 6.20). After 50% it dropped to z = 3.59, below the threshold of 4. OpenAI's own figure is 17%, not the 70% I say once in the video (the on-screen note fixes it). Chapters 0:00 OpenAI is watermarking text 0:59 What OpenAI announced 2:46 Claude is already watermarked 3:57 How a text watermark works 5:44 OpenAI's own numbers 7:10 Let's build one (Python) 8:50 Running the detector 10:51 What it can't tell you 12:01 What this means for you Sources OpenAI announcement: https://openai.com/index/eu-text-prov... OpenAI help article: https://help.openai.com/en/articles/8... textGrain paper (PDF): https://cdn.openai.com/pdf/e9508624-d... Anthropic: https://www.anthropic.com/news/claude... Claude support article: https://support.claude.com/en/article... Kirchenbauer et al., "A Watermark for Large Language Models": https://arxiv.org/abs/2301.10226 TechCrunch: https://techcrunch.com/2026/10/05/ope... Ars Technica: https://arstechnica.com/ai/2026/10/op... The Hindu: https://www.thehindu.com/sci-tech/tec... The demo is a toy green-list watermark for learning. It is not OpenAI's or Anthropic's method. Follow me X: https://x.com/panditdhamdhere LinkedIn:   / panditdhamdhere   GitHub: https://github.com/panditdhamdhere #AI #ChatGPT #Claude #OpenAI #Watermark #Python