# On the Possibilities of AI-Generated Text Detection

**Authors:** Souradip Chakraborty, Amrit Singh Bedi, Sicheng Zhu, Bang An, Dinesh Manocha, Furong Huang  
**arXiv:** [2304.04736](https://arxiv.org/abs/2304.04736)  
**Date:** 2023-04-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.04736v2/

## Key points

- Authors address the challenge of distinguishing outputs generated by Large Language Models (LLMs) from those generated by humans
- Detecting AI-generated text is crucial in various applications
- It is almost always possible to detect AI-generated text unless the distributions of human and machine-generated texts are exactly the same across the entire support
- Precise sample complexity bound for AI-generated text detection is derived
- Designing more complex detectors that consider multiple samples for detection is highlighted as an area for future research
- Empirical evaluations conducted on various real and synthetic datasets to support claim about better detectors
- Theory and results align with empirical findings from OpenAI regarding sequence length
- Provides solid theoretical justification for outcomes, making authors pioneers in this aspect
- Feasibility of achieving AI-generated text detection should be feasible in most scenarios with appropriate detector design and sufficient sample sizes.

## Summary

In their work titled "On the Possibilities of AI-Generated Text Detection," authors Souradip Chakraborty, Amrit Singh Bedi, Sicheng Zhu, Bang An, Dinesh Manocha, and Furong Huang address the challenge of distinguishing outputs generated by Large Language Models (LLMs) from those generated by humans. The ability to detect AI-generated text is crucial in various applications, but its feasibility has been a topic of debate within the community. The authors provide evidence that it is almost always possible to detect AI-generated text unless the distributions of human and machine-generated texts are exactly the same across the entire support. This observation is based on standard results in information theory and relies on the fact that as machine-generated text becomes more human-like, more samples are required for detection. They derive a precise sample complexity bound for AI-generated text detection, which quantifies the number of samples needed for detection. This finding presents additional challenges in designing more complex detectors that consider multiple samples for detection instead of just one. The authors highlight this as an area for future research. To support their claim about the existence of better detectors, they conduct empirical evaluations on various real and synthetic datasets. The authors' theory and results align with empirical findings from OpenAI regarding sequence length. Importantly, they provide a solid theoretical justification for these outcomes, making them pioneers in this aspect. Overall, their work contributes to understanding and addressing the challenge of detecting AI-generated text. It suggests that achieving AI-generated text detection should be feasible in most scenarios with appropriate detector design and sufficient sample sizes.