How do Imperial educators use AI for assessment?

Artificial intelligence (AI) is increasingly becoming part of higher education. At Imperial, staff are being encouraged to explore how generative AI can support teaching. Yet unlike with student use of AI, which is strictly regulated by policies, there appears to be no firm guidance governing how or how much lecturers can use AI in teaching and assessment.  

The result is a patchwork of approaches across departments, ranging from automated problem sets to AI-assisted marking and feedback.  For some students, these developments are exciting. For others, they raise questions about transparency, accountability, and the role of human judgement in education.  

The AI “principles” 

Imperial’s official position on AI in teaching is non-prescriptive. Guidelines encourage experimentation, suggesting staff may use AI to generate lesson plans, refine marking rubrics, and support feedback processes.   

In support, the College asks staff to “recognise the limitations of generative AI tools” and to “adopt transparent processes in evaluating and sharing how AI tools are integrated.” But it stops short of mandating how those principles are applied in practice. There are no enforceable rules specifying when AI can be used in marking, how students must be informed, or what a standard of disclosure looks like.

That ambiguity has given lecturers wide latitude and has resulted in them using AI in different ways.   

Automated homework and feedback  

The College currently operates an automated non-large language model-based assessment system embedded in Lambda Feedback, an Imperial-developed platform that automates feedback on mathematical and technical problems.  

The platform is currently used across multiple departments, and has reached more than 1,000 students. Its long-term goal is to provide personalised feedback immediately after homework submission. Although the current version does not use generative AI, developers say future versions could incorporate generative models to explain mistakes and suggest next steps.   

Imperial argues that automation can solve a longstanding problem in higher education: providing meaningful feedback at scale. In Mechanical Engineering, automated problem sets were introduced partly because homework previously received little or no feedback. Instead of waiting days or weeks, students can receive responses instantly and use staff contact time for deeper discussions.  

Using AI to assign marks   

In some cases, automation moves beyond feedback and into summative assessments.  

In the Department of Computing, Dr Robert Chatley has experimented with using large language models (LLMs) to assess weekly coding exercises in Software Engineering Design, a compulsory second-year module with approximately 230 students.   

Unlike traditional automated tests, which simply check whether code produces correct output, the system evaluates whether students have applied design principles taught on the course. Using detailed rubrics, the AI reviews submissions and generates feedback. Its output then feeds into a non-AI algorithm that calculates the overall mark – which only accounts for a small percentage of the module credit.

According to Dr Chatley, the system produced results that were often more consistent than human marking.   

“With a sufficiently detailed rubric, LLMs were able to read the students’ code and check for things very accurately, and consistently.”  

Dr Chatley argues that AI can be particularly effective for repetitive tasks that require consistency across hundreds of submissions. Comparing human feedback from previous years’ submissions with output from the LLM tool, he noted that where there was significant disparity in marking, “in most of the cases it was the tool that was correct, and the human marker had missed something.”

The approach also dramatically reduced turnaround times, allowing feedback to be returned within three working days.

AI as an Assistant   

Prof Thrishantha Nanayakkara from the Dyson School of Design Engineering uses ChatGPT to convert spoken feedback into written comments for students. After reading a report and recording around ten minutes of verbal observations, he asks ChatGPT to organise the content into sections such as strengths and areas for improvement.  

Importantly, in this case the feedback itself originates from the academic. Prof Nanayakkara added that he reviews the output to ensure it accurately reflects his intended feedback before sending it to students.

Voices against AI use

Not all professors are choosing to explore AI tools. Dr Dave Clements from the Department of Physics is an outspoken critic of AI, which he called the “current fashionable Kool-Aid”.

He is wary of professors “outsourcing some cognitive processing” to large language models, and sees “de-skilling” as a long-term risk: “a professor who keeps on outsourcing their feedback to an LLM is going to go through a cognitive decline and become less good at giving feedback.”

Dr Clements also said he was “not sure” AI tools could be used ethically, pointing out that the models have been trained “largely on stolen copyright works”.

“The real problem we have is that the workloads of all teaching staff are so high that they don’t have the time to do the job properly, so they outsource it to machines,” he analysed.

Regarding student perception, Dr Chatley also sensed an “undercurrent” of opposition to AI use in marking. “I think there’s a natural emotional reaction from students that they are interacting with a machine, rather than a human,” he said.

Such examples reveal the striking reality that there is no single “Imperial approach” to AI in teaching. Across Imperial, academics are experimenting with different tools for different purposes, while others reject this possibility. College guidance encourages transparency and emphasises that AI should not replace human expertise where judgement is required. Yet exactly how that principle is interpreted remains largely at the hands of individual lecturers.