Abstract
(from article abstract) To inform decision making for the sustainable use of plastic, policymakers require
accounts of all short-term and long-term use of plastic at a national level. However, the
underlying data used for plastics accounting requires extensive time and effort for collation or
estimation. Generative AI has recently demonstrated promise in research tools due to its use of
extensive knowledge bases. In this study, we tested whether GPT, the model behind Open
Artificial Intelligence’s ChatGPT, could produce estimates for plastic accounting tasks that are
comparable to experts. Fine-tuned GPT-3.5 turbo models were used for two tasks; i) estimating
the material composition of imported products by plastic type, and ii) estimating the volumes
of eight polymers in packaging in use in the Norwegian economy between 1951-2020. We
tested whether the training dataset size and modifying model hyperparameters improved
performance. Increasing the training dataset size was found to improve the performance of
models trained for the first task, while increasing the number of training rounds improved the
precision and accuracy of models trained for the second task. However, in both tasks models
did not provide estimates comparable to those provided by experts. Although future generative
AI models that are specifically trained for accounting tasks may increasingly become useful
tools for reducing manual work, currently they do not provide reliable estimates.