Ethical and social risks of harm from Language Models

This paper aims to help structure the risk landscape associated withlarge-scale Language Models (LMs). In order to foster advances in responsibleinnovation, an in-depth understanding of the potential risks posed by thesemodels is needed. A wide range of established and anticipated risks areanalysed in detail, drawing on multidisciplinary expertise and literature fromcomputer science, linguistics, and social sciences. We outline six specific risk areas: I. Discrimination, Exclusion andToxicity, II. Information Hazards, III. Misinformation Harms, V. MaliciousUses, V. Human-Computer Interaction Harms, VI. Automation, Access, andEnvironmental Harms. The first area concerns the perpetuation of stereotypes,unfair discrimination, exclusionary norms, toxic language, and lowerperformance by social group for LMs. The second focuses on risks from privatedata leaks or LMs correctly inferring sensitive information. The thirdaddresses risks arising from poor, false or misleading information including insensitive domains, and knock-on risks such as the erosion of trust in sharedinformation. The fourth considers risks from actors who try to use LMs to causeharm. The fifth focuses on risks specific to LLMs used to underpinconversational agents that interact with human users, including unsafe use,manipulation or deception. The sixth discusses the risk of environmental harm,job automation, and other challenges that may have a disparate effect ondifferent social groups or communities. In total, we review 21 risks in-depth. We discuss the points of origin ofdifferent risks and point to potential mitigation approaches. Lastly, wediscuss organisational responsibilities in implementing mitigations, and therole of collaboration and participation. We highlight directions for furtherresearch, particularly on expanding the toolkit for assessing and evaluatingthe outlined risks in LMs.

Further reading#

• Access Paper in arXiv.org