If cheating is optimisation then assessment must not be pure

Effect tracking and assessment

Abstract

The emergence of generative AI has broadened the question of how to ensure academic integrity. Where, in the past, many tools sought to counter specific threats to integrity (for example, detecting plagiarism or outsourcing), we now need to consider integrity more fundamentally: What work can we be assured the student did? One way of viewing cheating is from the perspective of optimisation: producing a result without performing the work. This perspective creates an interesting parallel with computer programming. A programmer often wants their compiler to optimise their program, altering its instructions to make it run more efficiently, without affecting its observable behaviour. Some characteristics of programs make this easier or harder to do. For example, a function is said to be “pure” if it always produces the same outputs for the same inputs and has no other effects. If the result is known, then performing the function can be invisibly replaced by inserting its result. If, however, a function has an observable effect, then the work of performing the function cannot be optimised away. This paper explores academic integrity assurance through the lens of “effect types”. In functional languages, functions are often written in such a manner that the types of effects they have are explicit and tracked by the compiler. The paper likens this to tracking the types of academic integrity measures that are present in each assessment, how they compose across a course, and therefore what observable effects are present to assure each learning outcome.