Using large language models in psychology

Large language models (LLMs), such as OpenAI's GPT-4, Google's Bard or Meta's LLaMa, have created unprecedented opportunities for analysing and generating language data on a massive scale. Because language data have a central role in all areas of psychology, this new technology has the potential to transform the field. In this Perspective, we review the foundations of LLMs. We then explain how the way that LLMs are constructed enables them to effectively generate human-like linguistic output without the ability to think or feel like a human. We argue that although LLMs have the potential to advance psychological measurement, experimentation and practice, they are not yet ready for many of the most transformative psychological applicationsbut further research and development may enable such use. Next, we examine four major concerns about the application of LLMs to psychology, and how each might be overcome. Finally, we conclude with recommendations for investments that could help to address these concerns: field-initiated 'keystone' datasets; increased standardization of performance benchmarks; and shared computing and analysis infrastructure to ensure that the future of LLM-powered research is equitable.

Sections

How the models are fitted

LLMs are designed to reproduce word co-occurrence patterns that were found in their training data. They have become very good at this task owing to immense quantities of training data and complex architecture. The most common model architecture for LLMs is the 'transformer', which forms the backbone of modern models such as BERT, GPT, Bard and LaMDA. These LLMs are essentially massive systems of nonlinear regression equations (specifically, neural network machine learning models). These models often have millions of parameters that were calculated by taking sentences as predictors (an X in a regression equation) and masked-out words or the next sentence as an outcome (Y). The prediction error from one fit is then used to update the models' parameters (via backpropagation), and the process is repeated Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.