Artificial intelligence-based (AI-based) language analysis has undergone a “paradigm shift” in changing how AI are developed (Bommasani, et al., 2021, p. 1). A few years ago, most AI systems were purpose-built –optimized for one specific task such that systems for answering natural language questions (i.e. question answering) used a different model than that for sentiment analysis (scoring the positivity or negativity of a text). Now, most AI language systems are based on a large language model – a “foundational” model. The state-of-the-art system for sentiment analysis, question answering, paraphrasing, and dozens of other language tasks are based on the same underlying statistical deep learning model, which only needs to be “fine-tuned’’ to perform particular tasks. In fact, this technology now touches the daily life of nearly everyone with a smartphone, as it has quickly become the basis for modern web search, digital assistants’ language (Alexa, Siri, etc..), and machine translation. 

The technology enabling this supposed paradigm shift is the transformer-based Large Language Model  (Devlin et al., 2019; Bommasani et al., 2021). Large language models largely owe their success to their ability to model words in the context that they are used. Bringing such context to psychological language analysis, large language models can more precisely quantify the specific meaning of language and yield a truer understanding of the person behind the words.

In our research we have shown that Large language models push the assessment accuracy of language assessment to a theoretical upper limit of predicting mental health rating scales (upwards of r = .85; Kjell et al., 2022). Hence, based on early empirical successes of using large language models for mental health assessment (e.g., Matero et al., 2019; Mohammadi et al., 2019; Zirikly et al., 2019; Kjell et al., 2022), we present evidence showing that this technique needs not only transform the field of AI, but it is the missing piece for a corresponding transformation in psychological mental health assessment. We propose that the technique has the potential to modernize assessment methods from the reliance on closed-ended rating scale responses to more accurate, fine-grained, and ecologically valid assessments of individuals’ state of mind. By fully leveraging individuals’ personal descriptions of their mental state in their own words, the technique has the potential to –not only improve current assessments incrementally– but transform the very nature of how individuals’ state-of-minds are both measured and described, ultimately increasing our understanding of mental health.

This presentation provides evidence showing (1) the intrinsic advantages of natural language in communicating mental health and showing how language has several favorable measurement characteristics, (2) how word context matters in mental health, and reviewing how the unique contributions of large language models may realize the measurement precision of language, and then (3) how the advantages of large language models can ground psychological mental health assessment in natural language a reality.