Can ChatGPT pass a physics degree? Making a case for reformation of assessment of undergraduate degrees

Abstract

The emergence of conversational natural language processing models presents a significant challenge for Higher Education. In this work, we use the entirety of a UK physics undergraduate (BSc with Honours) degree including all examinations and coursework to test if ChatGPT (GPT-4) can pass a degree. We adopt a "maximal cheating" approach wherein we permit ourselves to modify questions for clarity, split questions up into smaller sub-components, expand on answers given - especially for long form written responses, obtaining references, and use of advanced coaching, plug-ins and custom instructions to optimize outputs. In general, there are only certain parts of the degree in question where GPT-4 fails. Explicitly these include compulsory laboratory elements, and the final project which is assessed by a viva. If these were no issue, then GPT-4 would pass with a grade of an upper second class overall. In general, coding tasks are performed exceptionally well, along with simple single-step solution problems. Multiple step problems and longer prose are generally poorer along with interdisciplinary problems. We strongly suggest that there is now a necessity to urgently re-think and revise assessment practice in physics - and other disciplines - due to the existence of AI such as GPT-4. We recommend close scrutiny of assessment tasks: only invigilated in-person examinations, vivas, laboratory skills testing (or "performances" in other disciplines), and presentations are not vulnerable to GPT-4, and urge consideration of how AI can be embedded within the disciplinary context.