Amy Liu

While machine-coding of data has dramatically advanced in recent years, the literature raises significant concerns about validation of LLM classification showing, for example, that reliability varies greatly by prompt and temperature tuning, across subject areas and tasks – especially in “zero-shot” applications. This paper contributes to the discussion of validation in several different ways. To test the relative performance of supervised and semi-supervised algorithms when coding political data, we compare three models' performances to each other over multiple iterations for each model and to trained expert coding of data. We also examine changes in performance resulting from prompt engineering and pre-processing of source data. To ameliorate concerns regarding LLM's pre-training on test data, we assess performance by updating an existing dataset beyond what is publicly available. Overall, we find that only GPT-4 approaches trained expert coders when coding contexts familiar to human coders and is more consistent coding across contexts. We conclude by discussing some benefits and drawbacks of machine coding moving forward.