Artificial Intelligence in education, part 2

By | May 27, 2018

A few months ago I wrote about the potential impact of Artificial Intelligence in the schools sector from the perspective of the challenge it presents. I was prompted to revisit this theme when I attended a talk on AI given by Microsoft CEO Satya Nadella this week.

Leaders from other industries also presented on how AI is affecting their work and one which resonated strongly was Professor Neil Sebire, Chief Research Information officer at Great Ormond Street Hospital, whose message was one of bullish pragmatism – they are exploiting the analyses made possible through machine learning to directly inform their practice. In paraphrase: “We’re not interested in a being a bit better as doctors. We want to do things as doctors which were previously impossible” and “When we reviewed our digital strategy we wanted to move away from the concept that ‘This is medicine, it’s different, that wouldn’t work’ to instead consider the technology that is transforming other sectors”. There’s a stridency and confidence there that I don’t hear (credibly) evinced in education.

But that’s enough editorialising, onto the news: how might AI start to affect education in the near future?



This is the most obvious starting point, as there’s a clear line of sight from what other sectors are doing to a use case in education and ‘data analysis’ is something we’re all familiar with. Indeed, AI data crunching is an avenue of enquiry my trust is already pursuing with specialist partners in the data and technology worlds.

Everyone knows that schools gather lots of data about pupils – these data are probably best described as ‘wide’ rather than ‘deep’, as the column-to-row ratio is relatively high in a single school. There are also data sets which aren’t ostensibly about children but do describe schools in comparable ways – teacher turnover, engagement rates, parental feedback, what the school spends its money on, etc. Today, groups of schools that can aggregate these data in one place can acquire the depth needed to start to confidently identify the patterns which seem to underlie success.

Pretty soon you’re swimming in a lake of data that looks full of boundless possibilities – if only you had the capacity to construct interesting hypotheses about how these myriad factors may inter-relate and then to perform the necessary analyses. To an extent, a talented and experienced human can add value here, but it’s slow, laborious work bounded by our cognitive capacity. A well-constructed and trained AI could rip through that ceiling-high paper stack like a chainsaw, carrying out tens of thousands of calculations while the human analyst is still hanging up their coat and boiling the kettle. By the time the first cup of tea has been drunk, it has hunted down hundreds of unsuspected correlations and posited interesting suggestions for further, more detailed investigation by a human. In short, AI could rapidly help schools construct new knowledge about the aspects of their operation that seem to lead to specific outcomes .

However, Garbage In Garbage Out remains true of any data system, regardless of the intelligence of the analysis. To truly understand how a school’s activity relates to outcomes, we would need a much more reliable and low-effort way of recording the detailed experience of each pupil. Which leads me to my next point…



We worry, as is natural, about the erosion of our professional status – and even our livelihoods – by clever robots. Like most fears, this isn’t grounded in reality. The oft-cited ‘Any teacher who could be replaced by a robot should be’ underscores the point that learning is fundamentally a social activity in which the quality of the (human) relationship between teacher and pupil is pivotal.

That doesn’t mean that there aren’t realistic benefits to be reaped from putting AI in the harness and letting it pull the cart up the hillier bits of the workload landscape. As Magnus Revang, Research Director of Gartner put it at the Microsoft event, “What we replace with AI are tasks not jobs”. The question we should all be asking is ‘With the strides made in Artificial Intelligence in the last 5 years, which tasks currently carried out in schools by humans could relatively unproblematically be offloaded onto AI?

One relevant example is found in EmPower MD, an intelligent scribe which is designed to work alongside doctors to automatically make contextually aware and coded notes during conversations with patients and consultations with colleagues, massively reducing the time the human doctors spend writing up, categorising and cross-referencing patient notes. This obviously frees them for the more meaningful task of doctoring. The video below shows this in action.


It’s not that hard to see how a similar ‘AI assistant’ for school staff could simply remove from teachers the need to record data. Of any kind. Working within a structure pre-loaded with contextual knowledge (edu-jargon, basically) and feeding into the MIS, any relevant conversation with a pupil (about next steps in learning), their parent (about targets or progress) or a colleague (about proposed interventions, for example) would be understood, codified and recorded against that pupil’s record.

This would extend to marking (which would become a spoken process, almost certainly done during the lesson). The workload impact would be dramatic, but that’s nothing compared to the knowledge we’d gain. Pretty soon the unprecedented depth and quality of what the system understands about the school experience of every pupil would make it possible for educational data systems to move beyond mere prediction to the era of prescription. That’s more than just snappy alliteration – prescription in our context means ‘We understand what has resulted in optimal outcomes for prior learners very similar to you, so at this stage in your education, your diet needs to look like this…’

With an automated way of recording specific interventions we might finally understand, to coin a phrase, what works and be able to apply that knowledge at an individual level.


Learning & Teaching

The same AI that codified pupil data in the previous paragraph could also provide transcription and translation of lessons on the fly, capturing every explanation (indexed and searchable, naturally) and allowing pupils to access teaching in their native language as the lesson goes on. That would be helpful to everyone in that classroom, live and long after the lesson has taken place.

[Unthinkable controversy alert!] Coding of these data – how something was taught, the language and methods used – could be factored into the AI ‘what works’ analysis to identify the most routinely successful ways of teaching something. Adequately anonymised, this would guide ITT and CPD, a revolution in our understanding of effective teaching. Hey, it might even bring to an end the trad-prog war that’s been stinking up our Twitter feeds for what feels like a decade…

A bot capable of holding human-level conversations (such as Microsoft’s Xiaoice) could take up the slack once the school day ends, providing support after-hours for students who need help understanding how to do something. This would be about prompting, demonstrating and placing students in position to make learning gains, rather than telling them an answer. The AI would need to be fed with and trained on specific domains of knowledge (for example, the manipulation of algebraic expressions), so that it could respond to specific questions about a part of the process, give similar examples, demonstrate alternative methods, provide encouragement and praise, etc, all while the pupil’s maths teacher is out at kickboxing class. Think of this as the best text book of all time, but one that’s able to hold a conversation with you about the stuff you don’t get.


All of the examples given above feel like tangible and genuine gains which Artificial Intelligence could deliver for education within the coming decade, given adequate appetite from the tech titans to make something special happen and enough enthusiasm from within our sector.  The technology is capable of the functions described above already, but there is much work to be done to implement it effectively into systems that would work with children. I think we should make a start immediately.

Leave a Reply

Your email address will not be published. Required fields are marked *