# Perspectives on the Social Impacts of Reinforcement Learning with Human Feedback

**Authors:** Gabrielle Kaili-May Liu  
**arXiv:** [2303.02891](https://arxiv.org/abs/2303.02891)  
**Date:** 2023-03-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.02891v1/

## Key points

- -Reinforcement learning with human feedback (RLHF) is a promising approach to teaching machines to think like humans.
- RLHF provides feedback from a human teacher in addition to a reward signal, allowing agents to learn from human feedback in a naturalistic manner.
- RLHF has been utilized in high-profile AI applications such as OpenAI's ChatGPT, DeepMind's Sparrow, and Anthropic's Claude.
- The paper aims to provide a systematic study of the social effects of RLHF and identify key social and ethical issues associated with this technology.
- The paper describes seven primary ways in which RLHF-based technologies will positively transform human experiences with AI: misinformation, AI value-alignment, bias, AI access, cross-cultural dialogue, industry, and workforce.
- One important question that needs addressing is whether RLHF can be developed and used without negatively affecting human societies.
- The paper proposes steps we can take to counter potential risks associated with the use of RLHF technology.
- Cooperation and intelligence sharing between governments and industry parties will be key to developing early warning systems for disinformation campaigns enabling rapid response threat information sharing and cross-platform defense.
- Overall this paper suggests that while there are potential risks associated with the use of RLHF technology; it also has significant potential for positive societal impacts if used responsibly.

## Summary

Reinforcement learning with human feedback (RLHF) has emerged as a promising approach to teaching machines to think like humans. This method provides feedback from a human teacher in addition to a reward signal, allowing agents to learn from human feedback in a naturalistic manner. RLHF has already been utilized in high-profile AI applications such as OpenAI's ChatGPT, DeepMind's Sparrow, and Anthropic's Claude, which are capable of overturning our understanding of how AI interacts with humanity. However, the wide applicability and burgeoning success of RLHF also raise concerns about its social impacts. This paper aims to provide a systematic study of the social effects of RLHF and identify key social and ethical issues associated with this technology. While text-based applications of RLHF have received much attention, it is crucial to consider the diverse range of areas where it may be deployed when evaluating its social implications. The paper describes seven primary ways in which RLHF-based technologies will positively transform human experiences with AI: misinformation, AI value-alignment, bias, AI access, cross-cultural dialogue, industry, and workforce. One important question that needs addressing is whether RLHF can be developed and used without negatively affecting human societies. The paper proposes that while RLHF raises concerns similar to those of existing AI technologies such as disinformation campaigns or misuse for propaganda purposes, there are steps we can take to counter these risks. For instance, governance of infrastructure supporting disinformation campaigns could mitigate their impact on society. Moreover, methods to counter AI disinformation will likely be equally useful for RLHF. Cooperation and intelligence sharing between governments and industry parties will be key to developing early warning systems for disinformation campaigns enabling rapid response threat information sharing and cross-platform defense. Overall this paper suggests that while there are potential risks associated with the use of RLHF technology; it also has significant potential for positive societal impacts if used responsibly. As such stakeholders must remain aware and intentional in adopting this technology while taking into account its potential benefits as well as risks.