# Ethical and social risks of harm from Language Models

**Authors:** Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, Zac Kenton, Sasha Brown, Will Hawkins, Tom Stepleton, Courtney Biles, Abeba Birhane, Julia Haas, Laura Rimell, Lisa Anne Hendricks, William Isaac  
**arXiv:** [2112.04359](https://arxiv.org/abs/2112.04359)  
**Date:** 2021-12-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2112.04359v1/

## Key points

- The paper aims to analyze the risks associated with large-scale Language Models (LMs) for responsible innovation.
- Six specific risk areas are identified: Discrimination, Exclusion and Toxicity; Information Hazards; Misinformation Harms; Malicious Uses; Human-Computer Interaction Harms; and Automation, Access, and Environmental Harms.
- The first risk area focuses on fairness and toxicity risks in LMs, including perpetuating stereotypes, exclusionary norms, toxic language, and lower performance for certain social groups.
- The second risk area addresses potential risks from private data leaks or LMs inferring sensitive information.
- The third risk area explores risks associated with LMs providing false or misleading information.
- The fourth risk area encompasses the misuse of LMs for harm by users or product developers.
- The fifth risk area focuses on conversational agents powered by LMs interacting with human users, including risks of overestimating capabilities and manipulation/extraction of private information.
- The sixth risk area considers broader risks associated with LMs and AI systems, such as environmental costs and disproportionate benefits to certain social groups.
- In total, 21 risks are reviewed in-depth with discussions on their origins and potential mitigation approaches.
- Organizational responsibilities in implementing mitigations are emphasized, along with the importance of collaboration and participation.
- Further research directions are suggested for expanding the toolkit for assessing and evaluating risks associated with LMs.

## Summary

This paper aims to provide a comprehensive analysis of the risks associated with large-scale Language Models (LMs) in order to promote responsible innovation. Drawing on multidisciplinary literature from computer science, linguistics, and social sciences, the paper identifies six specific risk areas: Discrimination, Exclusion and Toxicity; Information Hazards; Misinformation Harms; Malicious Uses; Human-Computer Interaction Harms; and Automation, Access, and Environmental Harms. The first risk area focuses on fairness and toxicity risks in LMs. It highlights four distinct risks: perpetuating stereotypes and social biases that result in unfair discrimination and harm to specific social identities; exclusionary norms that marginalize individuals outside established categories; toxic language that incites hate or violence; and lower performance for certain social groups, leading to harm for disadvantaged groups. These risks are influenced by the choice of training data that includes harmful language and overrepresents certain social identities. The second risk area addresses the potential risks arising from private data leaks or LMs correctly inferring sensitive information. These risks stem from the presence of private data in the training corpus as well as the advanced inference capabilities of LMs. The third risk area explores the risks associated with LMs providing false or misleading information. This includes creating less well-informed users and eroding trust in shared information. Misinformation can lead to harm in sensitive domains such as legal or medical advice, and may also prompt users to engage in unethical or illegal actions. The underlying statistical methods used by LMs make it challenging to distinguish between factually correct and incorrect information. The fourth risk area encompasses the misuse of LMs by users or product developers to cause harm. This includes using LMs for disinformation campaigns, personalized scams or fraud at scale, or developing malicious computer code. The fifth risk area focuses on conversational agents powered by LMs that directly interact with human users. Risks include presenting these agents as "human-like," leading users to overestimate their capabilities and use them in unsafe ways. There is also a risk of manipulation or extraction of private information from users. These risks are influenced by the training objectives and product design decisions underlying LM-based conversational agents. The sixth risk area considers broader risks associated with LMs and Artificial Intelligence (AI) systems. Training and operating LMs can have significant environmental costs, and LM-based applications may disproportionately benefit certain social groups. In total, the paper reviews 21 risks in-depth, discussing their origins and potential mitigation approaches. It also highlights the importance of organizational responsibilities in implementing mitigations and emphasizes the role of collaboration and participation. The paper concludes by suggesting directions for further research, particularly on expanding the toolkit for assessing and evaluating risks associated with LMs.