Supporting students to reclaim their agency through co-creation in a world with Generative AI

18 September 2026

Many discussions about Generative AI (GenAI) and chatbots powered by Large Language Models (LLMs) have gone hand-in-hand with discussions about prompt engineering, which became a popular term soon after ChatGPT was unleashed on the world. I have written about prompt engineering in the context of higher education before and while there are a lot of different prompting techniques for different contexts, prompt engineering is essentially about writing well-structured prompts to increase the probability of useful LLM outputs. Or, simply put, being specific.

Generative AI prompting and specificity  

But it turns out that LLM chatbots are quite good at dealing with ambiguity and they will helpfully (or infuriatingly) fill in any gaps if you’re not sure what you want or if you don’t articulate it particularly clearly. This capability can be leveraged to refine and even write whole prompts for you. This can be useful in some cases but problematic in others. A long LLM-generated prompt may give the illusion of accuracy because of its specificity, but it may not actually be aligned with your intent. 

As an example, try asking your favourite chatbot “give me a random number between 1 and 10”. It will probably return “7”, because “7” shows up more often in the training data in contexts related to random numbers between 1-10. Not so random after all. This is what we might call a false affordance. LLM chatbots are rife with false affordances. The chatbot seems to be usable for a certain thing, it says it can do the thing, but it doesn’t actually do the thing, despite reporting back that it has done the thing. 

Now try asking “programmatically generate a random number between 1 and 10”. Many contemporary chatbots will treat this prompt entirely differently by generating and executing actual code to generate a random number. The word “programmatically” has made all the difference here. 

While this is a relatively trivial example, it highlights the importance of domain knowledge, clarity of intent, and that prompt engineering is still a thing, even if some of us may have grown weary of it as a term. Any critical AI literacy efforts need to help students and staff navigate these false affordances of GenAI and explore potentially useful but hidden affordances accessible through different prompting techniques. It should also help them be able to decide proactively when not to use GenAI. 

Uncovering hidden affordances of Generative AI 

LLM chatbots are typically designed to answer questions in an engaging manner, reduce friction and increase immediate user satisfaction, often at the expense of accurate information. Just as social media recommender algorithms are designed to optimise for engagement, commercial LLM chatbots are also typically designed for engagement, with user agency and empowerment a secondary concern. 

Asking chatbots factual questions usually returns accurate answers, but problems arise when they generate content which is inaccurate and the user doesn’t notice because they have grown accustomed to trusting the seemingly all-knowing bot (see automation bias). 

So, are there different ways of using LLM chatbots which mitigate the risk of developing over-reliance on them as questioning answering machines and help students develop rather than relinquish their agency? Plenty. One approach is flipped interaction prompting which instructs the chatbot to ask the user questions, instead of the other way round. 

One potentially compelling use case for this approach is to aid reflection. A structured coaching prompt could be used to support a student to work through the Gibbs reflective cycle in order to attend to their feelings about a specific experience, analyse what they have learned from it and what lessons they might take forward. 

The case for Small AI 

Interestingly, you don’t need to use the latest and most powerful chatbot platforms for this type of reflective exercise to be effective. In fact, smaller language models like Ministral 3 3B (from Mistral AI) can work just as well. They use less energy (or “compute”) than the frontier models typically available through tools like ChatGPT, Claude or Gemini. 

Because of this, these smaller models can be run comfortably on many laptops and smartphones without any connection to the internet required. This can afford total privacy (ie your chat doesn’t leave your device) and can mitigate environmental impact due to significantly less energy consumption. 

At Queen Mary University of London (QMUL), we are taking the case for Small AI seriously, in line with our organisational values, with small language models (SLMs) being just one example. 

Co-creation as a means of supporting student agency 

As part of ongoing research at QMUL which was initiated as a Learner Interns Programme (LIP) project with student co-creation as a key focus, the development of a student guide for using Generative AI to support reflection is underway based on the findings from several student-centred focus groups. Some guiding principles already surfaced through the focus groups include: 

• abstaining from using GenAI to aid reflection is a valid position as part of critical AI literacy 

• reflection should be emphasised as a personal process and not something to be offloaded 

• GenAI should be positioned as a temporary supporting scaffold, with ‘off ramping’ as the ultimate goal 

• Practical prompt templates should be provided to mitigate the limitations and maximise affordances of GenAI tools used to support reflection and encourage customisation by students 

• examples of alternative ways of engaging with GenAI (eg small language models, locally hosted models) should be provided which may mitigate ethical concerns related to privacy, environmental impact, or sources of training data. 

These all link back strongly to several of QMUL’s graduate attributes which include: be AI and digitally literate, promote sustainability, take responsibility, act ethically, engage critically  

But developing a co-created student guide is really just a first step. The focus groups led to discussions around what needs to happen in the classroom and how educators have a responsibility to help students navigate GenAI and the implications of its use on learning, reflection and critical thinking. 

This may be a daunting task for many educators, especially for those who lack the confidence or experience in using digital technologies such as GenAI within their discipline or are understandably concerned over the impact of these technologies on learning. Perhaps these concerns can themselves fuel meaningful classroom discussions, laying the foundations for transformational learning that helps students grow as individuals, develop their own agency, and learn to positively influence the world around them. 

For educators to become trusted partners in learning alongside their students, has there ever been a stronger case for collaborative learning, student-staff partnerships and co-creation? 

This article was originally published by AdvanceHE on 15 September, 2026. 

https://advance-he.ac.uk/news-and-views/supporting-students-to-reclaim-their-agency-through-co-creation-in-a-world-with-generative-ai/  

Jonathan Jackson: IoT Lead, School of Physical and Chemical Sciences

https://www.qmul.ac.uk/spcs/staff/academics/profiles/jjackson.html