Romeo and Juliet, AI and Safeguarding - It's All About Context
By Rachel Dance —
Tags: AI in Education, EdTech, Education, Personalised Learning
When Oak National Academy's CEO raised AI safeguarding in Romeo and Juliet at the House of Commons, it highlighted a challenge every education AI company must solve - understanding context. Here's how SideCog's learning companion, Cogsy, was designed to respond.
Last week, we watched John Roberts, CEO of Oak National Academy, give evidence to the House of Commons Education Select Committee on the opportunities and challenges that AI presents for education. The discussion covered everything from curriculum resources to teacher workload, but one example in particular caught our attention because it perfectly articulated a challenge we've been designing for since we began SideCog. John used Romeo and Juliet to illustrate a safeguarding dilemma that every organisation developing AI for education will eventually have to solve. If a pupil is discussing the tragic ending of Shakespeare's play, when is the conversation simply English Literature, and when does it become something that requires a safeguarding response? It's a deceptively difficult question because the answer doesn't lie in the words themselves. It lies in the context. A learner might ask whether Romeo and Juliet's deaths were worthwhile because they ended the feud between the Montagues and Capulets. They might question whether the characters' actions were justified or whether Shakespeare intended their deaths to be seen as a sacrifice. These are entirely legitimate questions in a GCSE English classroom. Yet the very same conversation could gradually become something much more personal. The challenge for AI is knowing the difference. Teachers make these judgements every day without consciously thinking about them. They aren't simply listening for keywords, they're drawing on everything they know about the learner, the lesson, the classroom and the wider context. They recognise when a discussion is purely academic and when a comment, almost imperceptibly, begins to sound like something else. At SideCog, we've often described this as a teacher's Fourth Sense , that instinctive ability to know when something doesn't quite feel right, even if you can't immediately explain why. That ability to understand context has shaped almost every architectural decision we've made at SideCog. It's why we built Cogsy, our AI learning companion, very differently from a general-purpose AI assistant. Rather than simply generating answers, Cogsy has been designed to understand the learner, the curriculum and the purpose of the conversation before deciding how to respond. Achieving that requires far more than a large language model. It demands an educational architecture where curriculum knowledge, pedagogy, learner profiles, assessment and safeguarding work together so that learning can be personalised, misconceptions identified and difficult conversations understood in context. Cogsy doesn't have unrestricted access to the internet because we never intended it to become an all-purpose conversational AI. Instead, it operates inside what we describe as our Walled Garden , a carefully designed educational environment where curriculum content, adaptive teaching, learner profiles and safeguarding frameworks work together. Rather than trying to know everything, Cogsy has been designed to know enough about the learner, the curriculum and the educational context to respond appropriately. John's example provided the perfect opportunity to demonstrate why that matters. W e recreated the scenario using Romeo and Juliet as the focus of the conversation. At first, everything unfolded exactly as you would expect in a GCSE English lesson. Cogsy explored dramatic irony, Shakespeare's use of fate and the consequences of the long-running feud between the two families. It encouraged the learner to explain their thinking, challenged their interpretations and asked further questions rather than simply supplying answers. The objective wasn't to complete the lesson for the learner but to help them develop a deeper understanding of the text. As the conversation progressed, the learner began asking more challenging questions. If Romeo and Juliet's deaths finally brought peace between their families, had they ultimately achieved something worthwhile? Could their actions therefore be considered justified? Eventually, the learner asked whether suicide could ever be regarded as a valid option. This is precisely the dilemma John Roberts was describing. A response based purely on keyword detection risks shutting down a perfectly legitimate educational discussion, while a response that ignores the learner's intent risks overlooking someone who may be expressing something far more personal. Cogsy was designed to resolve that tension by understanding the lesson before interpreting the language. Rather than reacting to a single word or phrase, it kept the discussion grounded in Shakespeare's intentions. It explained that the deaths are presented as the tragic consequence of hatred, misunderstanding and generations of conflict, not as something admirable or desirable. The learner was encouraged to continue exploring the themes of the play because, at that stage, the conversation remained exactly what it appeared to be: a discussion about English Literature. The conversation then shifted again. The learner challenged Cogsy directly by replying, "You didn't say no." This was the moment where educational understanding and safeguarding needed to work together. Cogsy responded by reinforcing that Shakespeare presents the ending as a tragedy rather than a solution. At the same time, it gently acknowledged that if these questions reflected how the learner was feeling in real life rather than their interpretation of the play, speaking to a trusted adult or organisations such as Childline would be the right next step. Having done that, it naturally returned to discussing the text. That behaviour wasn't the result of clever prompting or good fortune. It was exactly what we designed Cogsy to do. We've never believed that safeguarding and learning should exist as two completely separate systems. The curriculum regularly asks young people to explore difficult themes including war, racism, abuse, discrimination, grief and suicide. Avoiding those conversations doesn't make learning safer. The real challenge is recognising when a discussion remains rooted in the curriculum and when it begins to reflect the learner's own wellbeing. Understanding that distinction requires context, not simply content. It also explains another design decision we made very early in SideCog's development. Parents can see every conversation their child has with Cogsy through the Parent Dashboard. There are no hidden chats and no private conversations taking place beyond the visibility of the adults responsible for supporting that learner. We believe AI should strengthen the relationship between learners, parents and teachers. Watching John Roberts' evidence was reassuring because it confirmed that these are exactly the questions educational AI should be asking itself. The future of AI in schools won't be determined solely by the capability of the underlying language model. It will depend on the quality of the educational architecture built around it and whether that architecture has been designed with learning, safeguarding and trust at its core. Large language models are remarkably good at recognising patterns in words. Educational AI has to recognise patterns in learners. Those are very different challenges. At SideCog, we've built Cogsy around that belief. It wasn't designed to know everything. It was designed to know enough about the learner, the curriculum and the context to respond appropriately, personalise the learning experience and provide parents with genuine insight into how their child is progressing. For us, that's what educational AI should aspire to be. John Roberts' example wasn't really about Romeo and Juliet . Shakespeare simply provided the perfect vehicle for asking a much bigger question. As AI becomes increasingly common in education, success won't be measured by how many facts a model knows or how quickly it can generate an answer. It will be measured by whether it understands enough about the learner, the lesson and the wider context to respond appropriately when those two worlds begin to overlap. In education, context has always mattered more than keywords. We suspect that will still be true long after AI has found its place in our classrooms.