Speaker

Dr Adam Lopez, Reader in Natural Language Technologies, School of Informatics: : Large Language Models are Normal Language Technology

Large language models (LLMs), exemplified by ChatGPT, have sparked an overwhelming amount of conversation about their future, much of it sensationalist. I argue that the technical trajectory of LLM technology currently resembles — and will very likely continue to resemble — the past trajectory of a closely related language technology: automated translation. This is because LLMs are simply a repackaging of the same set of technical ideas, whose development spanned much of the past century (belying narratives that cast them as a result of recent “breakthroughs”). But automated translation has been in the wild for decades, quietly expanding its reach into new social contexts. So, critically assessing its history in these contexts should allow us to understand the range of possible impacts from LLMs. In this talk I’ll sketch some of the technical and social histories of translation and discuss what I (tentatively) think they tell us about LLMs.

Adam is researcher in natural language processing who has worked on many scientific, mathematical, and engineering problems related to NLP. He has held roles in both academia and industry, as an engineer, scientist, lecturer, and as director of artificial intelligence at a startup. I am currently an Reader at the University of Edinburgh.

Zee  is a Chancellor’s Fellow in Responsible Machine Learning and Artificial Intelligence at the Centre for Technomoral Futures and the School of Informatics at the University of Edinburgh, where they lead the ZEST Lab and a also a faculty fellow at DAIR. Zee works on the intersection between machine learning, science and technology studies, and media studies with a focus on  decolonial and anti-capitalist critiques of ML. Zee develops evidence that current ML methods (i.e., the mathematics of it all) are constructed in a way that naturally afford authoritarianism, colonialism, and forms of hegemony (i.e., white supremacy, patriarchy, ableism, and so on) and thinking about and developing ML methods that rethink the ways in which ML is applied to offer resistance to such notions. After a Ph.D. at the University of Sheffield, Zee did post-docs at the Digital Democracies Institute, and Mohamed bin Zayed University of Artificial Intelligence.

Since 2016 there has been a large body of work aimed at making language technologies, and machine learning more widely, fair and non-discriminatory. This line of work has been given an increasing amount of attention as time has passed. However, a line work has argued that it is impossible for machine learning for human and social data–and language technologies more specifically–to be unbiased, fair, or ethical. In spite of these concerns, research and development has continued to engage with the myth of ethical language technologies. In this talk, I will argue that not only are unbiased language technologies impossible, the notion itself relies entirely on an unhuman conceptualisation of language. In this light, recent attempts towards mitigating biases in language models, and for models to act as agents with some imbued notion of what it means to be “ethical” can be understood as distractions from the question of what capacity we want language technologies to act within.