Stress-testing assessments against AI

British universities have been warned to “stress-test” all assessments after new research revealed “almost all” undergraduates are using generative artificial intelligence (genAI) in their studies.

Weale, S. (2025). UK universities warned to “stress-test” assessments as 92% of students use AI. The Guardian

Stress-testing against AI is problematic for several reasons:

• The rapidly evolving nature of AI models means any testing quickly becomes outdated
• The sheer number of available AI tools makes comprehensive testing impossible
• Different prompt engineering techniques can yield vastly different results from the same model
• Access to paid vs. free models creates equity issues in testing
• The resource investment required for continuous testing is unsustainable for most institutions
• Testing can’t account for human creativity in using these tools in unexpected ways

Philip Dawson suggests that there are 3 paths that unsupervised assessments might follow (and no, we’re not going to move to supervised assessments across the sector):

1. Massive grade inflation if we leave assessments as they are (because students will use AI in ways we cannot detect…and they will all use AI).
2. Move to norm-referenced assessment (i.e. grading to a curve), which isn’t ideal in professions that assess against standards e.g. regulated professions.
3. Significantly raise our expectations for what students can do (with the help of AI) i.e. standards creep.

I’ve been advocating for the 3rd option in our school; as AI increases in capabilities, so will student performance increase, so our assessments will need to take that into account. Or, as I’ve been saying for a while, if AI enables superhuman performance then our assessments need to evaluate superhuman outputs.

For example, instead of asking students to write a basic analysis of a case study, we might require them to:

• Generate multiple solution scenarios, complete with risk assessments and implementation strategies for each, comparing them using advanced metrics
• Develop comprehensive cross-disciplinary approaches that combine insights from multiple fields, with detailed justification for how these connections enhance the solution
• Create complete project proposals including detailed stakeholder analyses, resource allocation plans, and contingency strategies, rather than simple problem-solution papers

Unfortunately, this introduces it’s own problems, not least the fact that it forces students to use AI (although this report suggests that they’re using it anyway).

While there’s no perfect solution to the AI assessment challenge, raising our standards to match enhanced AI capabilities seems the most forward-thinking approach. This requires careful curriculum redesign, clear communication with students about AI use expectations, and ongoing faculty development to understand AI’s capabilities. Rather than fighting against AI adoption, we need to embrace it while ensuring our assessments truly measure students’ ability to leverage these tools professionally and ethically. What experiences have you had with adapting assessments for the AI era?