Authors: Bradley Emi, Max Spero

Published on: February 21, 2024

Impact Score: 8.0

Arxiv code: Arxiv:2402.14873

Summary

• What is new: this http URL classifier significantly outperforms existing zero-shot methods and commercial tools in identifying text authored by AI, with a much lower error rate.
• Why this is important: Distinguishing text written by AI from text written by humans is increasingly challenging.
• What the research proposes: A transformer-based neural network, trained with a novel algorithm called hard negative mining with synthetic mirrors.
• Results: Achieves over 9 times lower error rates across various text domains and is effective against both open- and closed-source large language models without bias against nonnative English speakers.

Technical Details

Technological frameworks used: Transformer-based neural network

Models used: DetectGPT, other leading commercial AI detection tools

Data used: Text domains include student writing, creative writing, scientific writing, books, encyclopedias, news, email, scientific papers, Q&A

Potential Impact

This advancement could impact markets relying on content authenticity, including educational platforms, publishing industries, and online content creation businesses.

Want to implement this idea in a business?

We have generated a startup concept here: Authentext.