While teaching the MSc-level blockchain module at Imperial College London as lecturers during and after our Computer Science PhDs, our team found ourselves in a strange position. We were simultaneously witnessing some of the most powerful educational tools ever created and some of the most worrying changes in student learning we had encountered.
Almost overnight, tools like OpenAI's ChatGPT became deeply embedded into how students approached university work. Students were using AI constantly: to understand lecture material, work through tutorials, debug code, summarise papers, generate explanations, and explore ideas conversationally. In many cases, the results initially looked extremely positive. Students were progressing through exercises faster, asking more ambitious questions, and often appearing significantly more productive than previous cohorts.
For a while, it genuinely seemed as though AI might dramatically improve technical education. However, as the months went on, we began noticing something more complicated emerging beneath the surface.
The students using AI most heavily were often the same students who appeared to struggle the most when they were no longer able to access the tools. During tutorials, they could often arrive at workable answers quickly. Yet in examinations - where deep understanding, independent reasoning, and conceptual clarity become much harder to mask - performance began deteriorating in ways that concerned us.
What we observed was not simply "students cheating with AI". That framing is too simplistic and, we think, fundamentally misunderstands the issue.
Instead, we began seeing a growing divide between:
- students using AI to amplify their learning, and
- students gradually outsourcing the learning process itself.
The Difference Is Harder To See Than You Think
On the surface, those two behaviours can look almost identical. Both students may submit polished work, complete tutorials successfully, appear highly engaged and progress quickly through coursework. Yet underneath, the cognitive processes are entirely different.
One student is using AI as a tool for exploration: interrogating concepts, testing understanding, compressing feedback loops, and accelerating genuine comprehension. The other is slowly becoming dependent on the system to perform the difficult cognitive work that education is actually meant to develop. That distinction matters enormously because real expertise - especially in technical disciplines like computer science or mathematics - is not simply the ability to produce answers.
Real expertise is the ability to reason through uncertainty independently, hold complex systems in your head, debug flawed assumptions and construct understanding from incomplete information.
Those abilities are difficult to build. More importantly, they often require cognitive friction.
Struggle Is Not A Bug In Learning
One of the most obvious things about learning is that struggle is the mechanism by which learning occurs.
The process of sitting with confusion, attempting a derivation yourself, debugging your own reasoning or wrestling with an unfamiliar concept is precisely what creates durable understanding.
But modern AI systems can accidentally short-circuit that process. If every moment of uncertainty is instantly resolved by an external system, students can successfully complete tasks without ever developing the internal reasoning pathways those tasks were designed to cultivate. The result is a strange phenomenon where students may appear increasingly capable during the learning process itself while simultaneously developing a shallower grasp of the underlying material.
In technical subjects, this becomes especially dangerous because outputs can look deceptively convincing.
A student can produce functioning code and sophisticated technical explanations while still lacking deep, or indeed any, understanding of the underlying concepts.
The Problem Is Not AI
Importantly, we do not believe the solution is banning AI.
That approach is neither realistic nor desirable because the utility of these systems is simply too large, and in many respects they can genuinely improve education when used properly.
The problem is not AI itself. The problem is that universities currently lack the infrastructure layer needed to integrate AI into education responsibly.
General-purpose AI systems were not designed around pedagogy, curriculum structure, assessment design or institutional governance. They do not understand intended learning outcomes, distinguish between productive struggle and unnecessary frustration, or align themselves to how a specific university teaches a specific course. Critically, universities themselves typically have very little visibility or control over how students are interacting with these systems.
That creates a remarkable situation in which, for many students, the primary interface through which they now engage with knowledge is increasingly external to the institution itself.
Why We Created FactBeat
The idea behind FactBeat emerged directly from these experiences in the classroom. We became increasingly convinced that if AI is going to become a permanent part of education - and it almost certainly will - then universities need infrastructure that allows them to participate meaningfully in that process rather than simply reacting to it from the outside.
Educational AI should not simply optimise for convenience or answer generation. It should optimise for learning outcomes.
In some situations, the best response is not to immediately provide the answer at all. Sometimes the correct educational intervention is a hint, a guiding question, a reference back to lecture material or encouragement for the student to reason through the next step independently.
Achieving that requires systems designed around pedagogy rather than general purpose AI systems.
FactBeat is designed to help build that missing institutional layer.
The goal is not to replace the AI tools students already use. Rather, it is to allow universities to integrate their own:
- course materials,
- governance models,
- educational objectives,
- and institutional controls
into the AI learning process itself.
Institutions should be able to control what content AI systems can access, shape how that content is surfaced, understand where students are struggling, and ensure that AI reinforces learning rather than gradually replacing it.
The Next Educational Layer
More broadly, we believe higher education is entering a transition comparable to the early internet era.
Initially, universities viewed the internet as something external to education. Eventually, it became inseparable from education itself.
AI appears to be following the same trajectory, only much faster.
The central question is therefore no longer whether students will use AI, because they already are. The real question is whether universities will help shape how AI integrates into learning or whether they will simply adapt reactively to systems built elsewhere by organisations optimising for entirely different objectives.
The future of education will almost certainly involve AI deeply.
Whether that future produces deeper understanding or merely more efficient superficiality remains an open question.
FactBeat was created because we believe universities should have the tools to help answer that question properly.