Get students on board with AI for marking and feedback
AI can potentially augment feedback and marking, but we need to trial it first. Here is a blueprint for using enhanced feedback generation systems and gaining trust
Higher education discussions on artificial intelligence (AI) often gravitate toward utopian ideals or dystopian fears. But instead of looking to science fiction, we might learn from the benefits of regulating previously unregulated markets when attempting to address the potential for unethical use of AI in marking and assessment. To foster trust, accountability and equity, educators should be supported in conducting transparent AI trials for marking and feedback, reducing the risk of covert use.
Low-stakes v high-stakes assessments
AI has proven its value in low-stakes formative feedback, where its rapid and personalised responses enhance learning. However, in high-stakes contexts where grades influence futures, autonomous AI marking introduces risks of bias and distrust. We therefore suggest that for high-stakes summative assessments, AI should be trialled in a supporting role, augmenting human-led processes.
In the first instance, it should only be used after educators have completed their usual marking and released their marks and feedback to students. In these AI-augmented trials, AI could provide additional complementary suggestions with human oversight and transparency. In these scenarios, students would be encouraged to “opt in” to provide feedback and “feed forward” on the process.
Enhanced AI feedback generation systems
Off-the-shelf generative AI tools such as ChatGPT, Copilot and Gemini often provide general or superficial feedback that lacks the depth, nuance and discipline-specific guidance needed for tertiary-level study. A more effective approach involves retrieval-augmented generation (Rag). Rag is a framework that combines a vector database with a generative AI model to retrieve relevant past information, such as documents (eg, past assignments) that have been uploaded into the vector database, for response generation. Rag requires access to an application programming interface (API) and setting it up typically involves coding. Numerous code examples and libraries are available from open-source platforms.
To further refine AI-generated feedback, institutions can additionally use fine-tuning, which customises the model by training it on subject-specific datasets or higher education-related data, enabling the model to produce even more relevant and contextually aware feedback. By combining Rag, which retrieves relevant past information for response generation, with fine-tuned AI models, which allow the model to learn from subject-specific data, educators can provide feedback that is not only more precise but also better aligned with course expectations. These AI-enhanced systems can highlight students’ skill gaps, suggest improvements, and encourage deeper reflection through tailored prompts. This approach does not replace educators; instead, it supports them by providing tailored feedback suggestions that educators can incorporate into their own feedback.
It’s easier than you think
In practice, implementing Rag and fine-tuned models is more straightforward than it may sound. Educators do not need to build AI models themselves – simply uploading proprietary course materials can significantly enhance AI-driven feedback. Even the technical setup is becoming increasingly accessible, with some educators able to integrate these tools independently rather than relying on central IT resources.
What about data privacy?
Concerns about data privacy in AI-driven feedback are valid, but Rag and fine-tuning do not increase these risks. Unlike general-purpose GenAI models that rely on broad, pre-trained datasets, Rag restricts AI’s access to specific, institution-approved sources, and fine-tuning embeds only vetted, proprietary knowledge.
If educators and institutions do not take the lead in shaping responsible AI feedback systems, commercial technology firms will fill the gap, potentially prioritising profit over pedagogy. By engaging proactively, the sector can retain control over ethical AI deployment, ensuring that student data is handled securely while maximising AI’s potential to support learning.
- Spotlight guide: Bringing GenAI into the university classroom
- Assessing the GenAI process, not the output
- Designing assessments with generative AI in mind
Addressing resistance
Resistance to AI in marking and summative feedback remains strong among educators and students. When one of us discussed with students recently in a creating digital futures course, their reaction to the idea of educators using AI in marking and feedback provision was overwhelmingly negative. As one student put it, “I would be really upset if assessors started to use AI. I know it’s coming, but I hope I’ll have finished my degree by then.”
This sentiment reflects broader concerns about fairness, transparency, potential dehumanisation and value for money, highlighting the need for thoughtful communication and ethical safeguards.
To navigate resistance we recommend the following principles:
- Include students and colleagues in discussions about AI’s potential role and limitations
- Emphasise the role of humans-in-the-loop (ie, with educators involved at every step) and human oversight (ie, with educators being ultimately responsible) to ensure fairness
- Begin with opt-in trials outside the traditional marking process and with research ethics approval in place to build trust and gather feedback.
A proposed pilot might proceed as follows:
- Traditional marking first: educators complete grading and provide feedback as usual
- Opt-in participation: students are invited to try an experiment exploring AI-augmented feedback, with no impact on their previous communicated grades
- AI feedback generation: AI supplements the original feedback, providing additional insights reviewed by the assessor for accuracy and relevance
- Combined and augmented report: students who opt in receive a combination of human and AI-generated feedback to compare and evaluate (without the previous summative feedback or mark being amended)
- Critical reflection and feedback: participants are invited to share their experiences to inform future iterations.
Rethinking assessments and building a shared vision
Beyond integration, we should also rethink assessments altogether. Peer evaluations, oral exams and synoptic assessments might offer more holistic methods of evaluating learning and knowledge. Furthermore, such assessment innovations might reduce over-reliance on rigid marking schemes, embrace diverse learning styles and potentially increase student-educator contact time.
Isabel Fischer is an associate professor (reader) of responsible digital innovation and education at Warwick Business School; Ashkan Eshghi is a Houlden fellow at Warwick Business School; Matt Lucas is a senior product manager at IBM and a teaching associate at the University of Warwick; Neha Gupta is an assistant professor in the Information Systems and Analytics Group at Warwick Business School.
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.
