Risk is a key component of student behaviour, as I learned from a recent experiment. Confoundingly, it leads to underuse and overuse of artificial intelligence tools. I think managing risk from a student’s perspective could be one of the more powerful levers we have.
In their eyes, AI creates risk aversion all the way down, from test to results. Students want to do well. They use AI to avoid failing. They don’t use AI to avoid malpractice cases.
So, what was my experiment? First, I should say that it was flawed and lacked a control group, but offers interesting stuff nevertheless. I tried to de-risk one of my assessments. It was an online test worth 20 per cent of my students’ overall mark. It would have been a prime candidate for AI use.
The assessment was an unfinished Excel sheet mapping employee hours and sales. Students had to work out the formulas to calculate wages, commission and bonuses. I used AI to generate five versions of the test in the time it would have taken to produce just one. The students know this. I’m transparent with them whenever I use AI tools.
- Spotlight guide: AI and assessment in higher education
- How should assessors use AI for marking and feedback?
- (Re)learning critical reading in the age of GenAI
When I designed the test, I ran it through ChatGPT, Claude and Gemini to see how well they could do. This is my standard practice in assessment design. ChatGPT managed 14 per cent, Claude passed with just over 50 per cent, and Gemini refused to help until I explained that I was the professor testing the test! The combination of complex input and maths stumped all three GenAI models. Using Copilot built into Excel and asking the right questions, I got up to 60 per cent. But I had to do a lot of hand-holding.
So, students probably could use AI. But would they?
Well, if you only get one shot at an assessment, and it’s part of your degree, you should logically use anything you can to get a better score. But because I had multiple versions of the test, I gave the students five swings at it over five weeks. I told them they could try the test as many times as they liked, and they’d keep the best mark they got.
The results were interesting. Weaker students tried the test early, mostly failed, got feedback, and tried again and again. Soon they had blasted past the 50 to 60 per cent that AI could manage and were scoring more than 80 per cent.
When I asked students about AI use, they hardly used it at all, mostly just using ChatGPT like a supercharged stack overflow when they got stuck on a formula. When I pressed them with follow-up questions, it seemed like they would have used AI in a one-off test, because they didn’t want to risk failing. But because they knew they had multiple attempts, they were confident to see how well they did without.
However, I also got feedback from students who said they would usually not use AI but that they had been confident to try it out in this case because: a) I’d told them they could; and b) they were interested to see how much it could help them improve. A lot of students tried once without AI and then used AI for their second or third attempt. Others said their biggest reason not to use AI was fear they would be reprimanded for cheating.
This pattern matches what Tricia Bertram Gallant, academic integrity director at the University of California, San Diego, and psychology professor David A. Rettinger, from the University of Tulsa, note in The Opposite of Cheating: Teaching for Integrity in the Age of AI (University of Oklahoma Press, 2025): that stakes matter but in complex ways. This iterative test was high stakes overall (20 per cent of the final mark) but each attempt wasn’t. So, students cared about the results but were given enough of a safety net to experiment with both their own abilities and AI tools.
The key insight: most students want to learn and want to test their own abilities. But they don’t want to look like an idiot when their friends pass using AI while they toil away for hours for similar marks. Multiple attempts solved this. They could keep trying until they beat the best score an AI could manage, learning deeply in the process.
Risk drives student behaviour around AI in both directions. They use it when they’re scared of failing, and they avoid it when they’re scared of getting caught or becoming dependent.
So, give them a safe space to experiment – where failure isn’t final and AI use isn’t forbidden – and they’ll often choose the harder path of learning first, then use AI strategically to push beyond what they could achieve alone.
This combination of high-stakes outcomes with low-stakes individual attempts could be a valuable tool for dealing with AI in higher education. It pushes students beyond AI capabilities while letting them experiment safely. The average mark at the end of term was above 75 per cent, with most students making at least two attempts.
Of course, this approach requires significantly more work for educators: multiple versions, multiple marking sessions, more feedback. It’s only possible because AI tools can help increase our productivity. Which raises an interesting paradox: we may find that most AI productivity gains in education are consumed by the higher level of work we need to do in a post-AI world.
Chris Jones is an automation (AI) consultant/programmer and visiting lecturer at Regent’s University London, Kingston University, Met Film School and Ithaca College.
If you would like advice and insight from academics and university staff delivered direct to your inbox each week, sign up for the Campus newsletter.
comment