Language Generation Models Can Cause Harm: So What Can We Do About It? An Actionable Survey

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3299–3321 May 2-6, 2023 ©2023 Association for Computational Linguistics Language Generation Models Can Cause Harm: So What Can We Do About It? An Actionable Survey Sachin Kumar∗,♣ Vidhisha Balachandran∗,♣ Lucille Njoo♡ Antonios Anastasopoulos♢ Yulia Tsvetkov♡ ♣Language Technologies Institute, Carnegie Mellon University, Pittsburgh PA ♢Department of Computer Science, George Mason University, Fairfax, VA ♡Paul G. Allen School of Computer Science & Engineering, University of Washington, Seattle WA {sachink,vbalacha}@cs.cmu.edu, lnjoo@cs.washington.edu, antonis@gmu.edu, yuliats@cs.washington.edu Abstract Recent advances in the capacity of large lan- guage models to generate human-like text have resulted in their increased adoption in user- facing settings. In parallel, these improvements have prompted a heated discourse around the risks of societal harms they introduce, whether inadvertent or malicious. Several studies have explored these harms and called for their miti- gation via development of safer, fairer models. Going beyond enumerating the risks of harms, this work provides a survey of practical meth- ods for addressing potential threats and societal harms from language generation models. We draw on several prior works’ taxonomies of language model risks to present a structured overview of strategies for detecting and ame- liorating different kinds of risks/harms of lan- guage generators. Bridging diverse strands of research, this survey aims to serve as a practical guide for both LM researchers and practition- ers, with explanations of different mitigation strategies’ motivations, their limitations, and open problems for future research. 1 Introduction The new wave of large language models (LMs; Brown et al., 2020; Chowdhery et al., 2022; Zhang et al., 2022b) capable of generating text with human-like fluency, coherence, and realism (Zellers et al., 2020; Ippolito et al., 2020) has caused a paradigm shift in our society.1 With appli- cations like OpenAI’s ChatGPT, Microsoft’s Bing, and Google’s Bard, bringing such LMs directly to users, we are beginning to see the impact in fields like education (Schulten, 2023; Gleason, 2022), healthcare (Patel and Lam, 2023), law (ChatGPT and Perlman, 2022), science (Stokel-Walker, 2023), and more. Since language is inherently a tool of *Equal contribution 1While the majority of these models are trained on English, recent studies have also obtained similar advancements in other languages (Lin et al., 2021; Shliazhko et al., 2022). Figure 1: Overview of Intervention Strategies. A typical ML/NLP model development process involves data col- lection/curation, model training and design, inference, and finally application deployment. For each phase of this development cycle, different techniques can be adopted to mitigate harms. Our survey presents a tax- onomy of intervention strategies organized around the different phases where they can be applied. power—the primary means by which people and so- cieties perpetuate stereotypes and manipulate opin- ions (Bar-Tal et al., 2013; Chong and Druckman, 2007, inter alia)—LMs that are deployed to mil- lions of users also hold similar power, but our un- derstanding of their risks/harms has lagged behind (Bender et al., 2021). Indeed, LMs have been shown to introduce vul- nerabilities and threats, both inadvertent and …

Content truncated to 500 words