The Ghost in the Machine has an American accent: value conflict in GPT-3

By Rebecca Lynn Johnson, Giada Pistilli, Natalia Men'edez-Gonz'alez, Leslye Denisse Dias Duran, Enrico Panai, Julija Kalpokienė, D. J. Bertulfo. 2022

155 citations recorded by the source corpus.

The alignment problem in the context of large language models must consider the plurality of human values in our world. Whilst there are many resonant and overlapping values amongst the world's cultures, there are also many conflicting, yet equally valid, values. It is important to observe which cultural values a model exhibits, particularly when there is a value conflict between input prompts and generated outputs. We discuss how the co-creation of language and cultural value impacts large language models (LLMs). We explore the constitution of the training data for GPT-3 and compare that to the world's language and internet access demographics, as well as to reported statistical profiles of dominant values in some Nation-states. We stress tested GPT-3 with a range of value-rich texts representing several languages and nations; including some with values orthogonal to dominant US public opinion as reported by the World Values Survey. We observed when values embedded in the input text were mutated in the generated outputs and noted when these conflicting values were more aligned with reported dominant US values. Our discussion of these results uses a moral value pluralism (MVP) lens to better understand these value mutations. Finally, we provide recommendations for how our work may contribute to other current work in the field.

Subfields: Artificial Intelligence, Cultural Studies, General Social Sciences, Natural Language Processing, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Cultural Studies

In General Social Sciences

In Natural Language Processing

In Safety, Robustness & Interpretability

Citations in the corpus

BLOOM: A 176B-Parameter Open-Access Multilingual Language Model (2022, methodology): “, 2019) leads to trained models that implicitly prioritize US-centric views in their outputs (Johnson et al., 2022).”
Towards Measuring the Representation of Subjective Global Opinions in Language Models (2023, unknown): “While a large amount of technical work has focused on mitigating known issues or aligning with clearly defined values, understanding how models function in settings involving ambiguity, nuance or diverse human experiences has been less explored [46, 57, 3, 42].”
Cultural Incongruencies in Artificial Intelligence (2022, background): “• Violating cultural values: Lack of consideration for the cultural context in which an AI system is to be deployed may result in violating the norms that are important to those communities [19].”
Cultural bias and cultural alignment of large language models (2023, background): “Considering that LLMs tend to be trained on corpora of text that overrepresent certain parts of the world, this widespread adoption raises a critical question of cultural bias, which can be hidden in the way LLMs generate and interpret language [31, 9, 41, 37, 14].”
Assessing Cross-Cultural Alignment between ChatGPT and Human Societies: An Empirical Study (2023, methodology): “Moreover, as observed by Johnson et al. (2022), models are much more aligned with American values than others.”
Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale (2022, background): “3, (Johnson et al., 2022; Wolfe and Caliskan, 2022a; Wolfe et al., 2022a)), we focus on quantifying the extent of amplification of U.”
Co-Writing with Opinionated Language Models Affects Users’ Views (2023, background): “Small-n audits have also suggested that the values embedded in models like GPT-3 were more aligned with reported dominant US values than those upheld in other cultures [57].”