Expanding the methodological toolbox: Machine-based item desirability ratings as an alternative to human-based ratings

The accuracy of self-reported data in the social and behavioral sciences may be compromised by response biases such as socially desirable responding. Researchers and scale developers therefore obtain item desirability ratings, in order to maintain item neutrality, and parity with alternative options when creating forced-choice items. Gathering item desirability ratings from human judges can be time-consuming and costly, with no consistent guidelines with regard to required sample size and composition. However, recent advancements in natural language processing have yielded large language models (LLMs) with exceptional abilities to identify abstract semantic attributes in text. The presented research highlights the potential application of LLMs to estimate the desirability of items, as evidenced by the re-analysis of data from 14 distinct studies. Findings indicate a significant and strong correlation between human- and machine-rated item desirability of .80, across 521 items. Results furthermore showed that the proposed fine-tuning approach of LLMs results in predictions that explained 19 % more variance beyond that of sentiment analysis. These results demonstrate the feasibility of relying on machine-based item desirability ratings as a viable alternative to human-based ratings and contribute to the field of personality psychology by expanding the methodological toolbox available to researchers, scale developers, and practitioners.