# Machine Generated Text: A Comprehensive Survey of Threat Models and Detection Methods

**Authors:** Evan Crothers, Nathalie Japkowicz, Herna Viktor  
**arXiv:** [2210.07321](https://arxiv.org/abs/2210.07321)  
**Date:** 2022-10-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2210.07321v1/

## Key points

- Advances in natural language generation (NLG) have made machine-generated text indistinguishable from human-authored text.
- Access to generative models is becoming more democratized due to open-source models and user-friendly tools.
- Detecting machine-generated text is crucial to counteract the risk of abuse.
- The survey provides a comprehensive analysis of threat models posed by contemporary NLG systems.
- It offers a complete review of machine generated text detection methods, highlighting the need for attention to open problems in this area.
- Existing detection methodologies often lack realism, transparency, and fairness methods, potentially causing harm themselves.
- Collaboration between AI researchers, cybersecurity professionals, and non-technical experts is necessary to prevent widespread abuse of NLG models.
- Usage and disclosure policies for online platforms can help address the issue, such as bans or enforced rules mandating public disclosure of AI-generated content.
- Adjusting licenses for released models to require disclosure can also be beneficial.
- Adoption or development of licenses like the Responsible AI License (RAIL) can improve best practices around handling powerful NLG models.
- Coordinated efforts across technical and social domains are essential to harnessing the benefits of high-capacity NLG systems while minimizing societal damage caused by their misuse.

## Summary

Advances in natural language generation (NLG) have led to the development of machine-generated text that is becoming increasingly difficult to distinguish from human-authored text. With the availability of powerful open-source models and user-friendly tools, access to generative models is becoming more democratized. However, along with the great potential of state-of-the-art NLG systems comes the risk of abuse. To counteract this abuse, detecting machine-generated text is crucial. This survey provides a comprehensive analysis of threat models posed by contemporary NLG systems and offers the most complete review of machine generated text detection methods to date. It places machine-generated text within its cybersecurity and social context and provides guidance for future work in addressing critical threat models. The survey highlights the need for attention to open problems in machine generated text detection. Existing detection methodologies often do not reflect realistic settings or incorporate sufficient transparency and fairness methods, which can potentially cause harm themselves. The prevention of widespread abuse of NLG models requires collaboration between AI researchers, cybersecurity professionals, and non-technical experts. In terms of future development, usage and disclosure policies for online platforms are worth considering. These policies can take the form of bans or enforced rules that mandate public disclosure of AI-generated content. Researchers can also adjust licenses for released models to require disclosure. The adoption or development of licenses like the Responsible AI License (RAIL) can help improve best practices around handling powerful NLG models. Overall, this survey emphasizes the urgent need for attention to open problems in machine generated text detection in order to provide suitable defenses against widely-available NLG models. Coordinated efforts across technical and social domains are essential to harnessing the benefits of high-capacity NLG systems while minimizing societal damage caused by their misuse.