# 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/2304.04736v1/

## Key points

- Distinguishing between outputs generated by Large Language Models (LLMs) and humans is important in numerous applications
- The possibility and impossibility of such discernment have been subjects of debate within the community
- It should almost always be possible to detect AI-generated text unless the distributions of human and machine-generated texts are exactly the same over the entire support
- A precise sample complexity bound for AI-generated text detection tells us how many samples are needed to detect
- If machine text becomes more like human text, we need more samples to detect it, which gives rise to additional challenges of designing more complicated detectors that take in n samples to detect than just one, which is an area for future research.
- Empirical evaluations support their claim about the existence of better detectors demonstrating that AI-generated text detection should be achievable in most scenarios.
- Detecting AI-generated text has significant implications for various applications where trustworthiness is crucial.
- Continued research in this area is important.

## Summary

The ability to distinguish between outputs generated by Large Language Models (LLMs) and those generated by humans is of utmost importance in numerous applications. However, the possibility and impossibility of such discernment have been subjects of debate within the community. In their paper titled "On the Possibilities of AI-Generated Text Detection," authors Souradip Chakraborty, Amrit Singh Bedi, Sicheng Zhu, Bang An, Dinesh Manocha, and Furong Huang provide evidence that it should almost always be possible to detect AI-generated text unless the distributions of human and machine-generated texts are exactly the same over the entire support. The authors derive a precise sample complexity bound for AI-generated text detection which tells us how many samples are needed to detect. They observe that if machine text becomes more like human text, we need more samples to detect it. This gives rise to additional challenges of designing more complicated detectors that take in n samples to detect than just one, which is an area for future research. Their empirical evaluations support their claim about the existence of better detectors demonstrating that AI-Generated text detection should be achievable in most scenarios. The authors' results emphasize the importance of continued research in this area as detecting AI-generated text has significant implications for various applications where trustworthiness is crucial. Overall, this paper provides valuable insights into detecting AI-generated text and highlights potential avenues for further research on developing more sophisticated detectors capable of distinguishing between human and machine-generated texts with greater accuracy.