On Seeding Watermarks to Detect Verbatim LLM Copy-Paste Responses

Educators can now detect verbatim LLM submissions without relying on unreliable AI detectors. The paper introduces SteganoPrompt, a tool that embeds invisible instructions into assignment prompts. When students copy-paste the prompt directly into an LLM, the model reads the hidden instruction and adds a detectable signature to its response, revealing the copy-paste pathway specifically.

This approach avoids penalizing non-native English speakers. Unlike existing post-hoc AI-text detection tools that have been shown to unfairly flag non-native English writing, this input-side watermarking method focuses on detecting the specific behavior of verbatim copying rather than analyzing writing style or language patterns.

The invisible watermark is robust and practical for real-world classroom use. The encoded watermark survives common document formats (Word, Google Docs, PDF, Markdown, Slack, email) and learning management systems, meaning educators can embed it in prompts without worrying about it being accidentally stripped away during normal classroom workflows.

The tool puts control directly in educators' hands. Unlike output-side watermarks that require cooperation from AI companies like OpenAI, SteganoPrompt is a free, single-page web tool that educators can use independently—no special permissions or partnerships needed.

The method works across major LLM providers. Testing showed compliance across eight different LLM families, suggesting the watermarking approach is broadly effective rather than limited to one particular chatbot or model.