Research Summary: On Seeding Watermarks to Detect Verbatim LLM Copy-Paste Responses

The research introduces an input-side watermarking method for Large Language Models (LLMs) to detect verbatim copy-pasting of LLM responses in academic settings. This approach aims to address the limitations of existing AI-text detectors, which are often unreliable and can disadvantage non-native English speakers. Unlike output-side watermarks that require model provider cooperation, this new method empowers educators with direct control over detection.

Why it matters

This development is crucial for maintaining academic integrity and evaluating genuine student engagement in an era of prevalent AI tools. It offers an alternative detection mechanism that bypasses the limitations of existing methods and reduces reliance on external providers, providing a more robust and equitable solution for assessing original work.

Key insights

• LLMs enable instant generation of essays, code, and quiz answers, posing challenges for educators.
• Educators primarily seek to identify verbatim submission of LLM output, not all LLM use.
• Current post-hoc AI-text detectors are unreliable and may unfairly penalize non-native English writers.
• Output-side watermarking requires collaboration with LLM providers.
• The proposed solution is an input-side watermark, directly controllable by the educator, which embeds an invisible instruction.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2605.16336