Charles Darwin University study finds 20 Australian universities classify identical student generative AI uses inconsistently

A preprint posted on arXiv, first submitted on 30 August 2026 under the name Biranchi Poudyal, examines how Australian universities govern student use of generative AI. It asks whether the overlapping policies, procedures, guidance and assessment instructions at an institution lead to the same classification when the same student conduct is described. The findings indicate that they often do not.

The researchers built 15 standardized scenarios and applied them to the public policy materials of 20 universities, producing 300 classifications. Binding instruments were analyzed first on their own and then together with the full official environment, including guidance. Each case received one of five labels, running from clearly permitted to clearly prohibited, with indeterminate as a catch-all. Divergence between universities was measured with normalized Shannon entropy.

Of the 300 classifications, 120 (40.0%) were clearly prohibited, 97 (32.3%) were potential breaches, 27 (9.0%) were allowed subject to conditions and 56 (18.7%) were indeterminate. None met the strict test for clearly permitted. Universities disagreed most over a scenario with disclosed language rewriting and one with a disclosed AI-drafted paragraph, while an explicit assessment ban produced complete agreement. Binding instruments said nothing about generative AI in 100 combinations, and guidance resolved 88. The authors conclude that written rules address keeping AI-generated text more clearly than they address help limited to process.

For reliability, the lead researcher coded with AI assistance and checked every queued row by hand, and an independent human coder without AI assistance coded a blind sample of 75 rows. Agreement was 57.3%, with an unweighted Cohen's kappa of 0.395. The authors add that this testing does not predict misconduct or enforcement decisions.