Teachers are not ends-in-themselves, they are always a means to an end - improvements in the learner. Given this premise, could it be possible to eventually replace teachers with AI technology? This may not happen soon but let’s, as a thought experiment, ask whether it could.
Obvious points are that AI is 24/7, fast, scalable and cheaper. This gives it a head start. But could it teach? First, we need to break down the functions of teaching and learning. I have used a PGSE schema as my starting point, supplemented by other learning tasks.
1. Searching for answers
Google is nothing but AI and has already delivered a remarkable pedagogic shift with search. Let’s also reflect on something that has already happened. Since Google’s inception, the number of librarians, teachers of sorts, has steadily fallen. We have less need of book and Journal warehouses, now that most knowledge is online. Beyond this, OER resources, such as Wikipedia, YouTube and Khan Academy have transformed the landscape. All of this is available through AI enabled search. Beyond search we see apps, such as Photomaths, where you simply point your smartphone at a maths problem and it gives you the answer. More than this, it gives you the steps in between. That’s how powerful AI has become.
2. Student support
We have a good example of student support in the remarkable Georgia Tech bot, Jill Watson. This AI teaching assistant bot, trained to answer based on prior responses, was used with 300 AI students, without them knowing. They were more than fooled, they praised the teacher for its efficacy and speed. In fact, it was only because it was so fast in its replies that they spotted it was a bot. We can expect a lot more of this, as teacher support, gives way to intelligent AI agent support. Teachers, especially online teachers, spend a lot of time on admin and support, so this is one area that is ripe for AI automation.
3. Demonstrate good subject and curriculum knowledge
This is where AI has made good progress and could outperform teachers. At one level Google already provides access to ‘knowledge’ on every subject, in detail. On ‘knowledge’ Watson beat the Jeopardy Champions way back in 2011. It is now a web-based service with access to huge knowledge bases and AI. At the skills level, YouTube is already the search engine of choice for learning how to ‘do’ things. With 3D, virtual worlds one can see how learn by doing can be expanded, as it was with flight sims, through cheap consumer technology, high in AI. The Todai project in Japan has produced AI software that passed the Tokyo University entrance exam, in Maths, Physics, English and History with a score of 53.8% versus the average human score of human 43.8%. Why is this important? Well, if AI has graduate level skills at this stage, it is likely to replace graduate-level jobs.
4. Plan and teach well-structured lessons
Lessons or learning experiences can be idiosyncratic, even flawed. AI design offers, not only optimal design but also continuous improvement, as it uses individual and aggregated data to spot poor components in lessons. There are plenty of examples of this already, from MOOCs and other forms of online learning, where errors in content, bad designed questions, overlong videos and presentations, have all been caught by data produced by online learning systems that constantly look for improvement. ‘Structured’ is an interesting concept here and one could argue that learning needs to be far more structured than the one-size-fits-all lesson plan. Differentiation could be identified and handled by AI in a way that traditional teaching cannot. The promise is of learning experiences that are not only structured towards individual learners but also continuously improve as machine learning identifies and acts on identified weaknesses. AI may even automatically produce lessons and content. See WildFire.
5. Promote good progress and outcomes by pupils
Progress tracking is not easy, as it requires the simultaneous tracking of actual performance across many learners. This is notoriously difficult in teaching. AI, on the other hand, promises to do this across many learners in realtime, as it gathers evidence that no teacher can possibly hope to gather through traditional observation and testing. More than this, one could argue that AI has a lot to offer in being free from human biases, that sometimes inhibit learner progress. We have ample evidence of gender and socio-economic bias in teaching. AI can be free from bias on gender, race, accent and background. Again, AI promises a scalable solution to this problem that may be of higher quality than a teacher-delivered system.
6. Adapt teaching to respond to the strengths and needs of all pupils
One of AI’s first forays into teaching has been through adaptive systems. These are already at work and producing impressive results. They act like a Satnav, which constantly monitors the performance of individual learners and adjusts what they are asked to do next. This is done in realtime. Content is no longer a linear curriculum of flat resources but a network of learning experiences that can be dynamically delivered to individual learners, based on their precise needs at that precise time. The analytic and predictive strengths of AI may very well identify factors that both inhibit and enhance learning in any individual.
In addition, AI can monitor and deliver to the needs of all pupils, especially those with special educational needs; high ability, English as an additional language, disabilities, and be able to use and improve distinctive approaches. Technology has already made a big difference in SEN, AI will make an ever bigger difference. All of this may be way beyond the ability of any teacher to deliver on scale. It is here that the first successes in AI driven teaching are being achieved.
Chris Piech’s team at Stanford & Google, used 1.4m answers to problems set by Khan Academy to diagnose what any learner was likely to get right/wrong in their next question. The software had to know what went wrong and why. It is, currently, accurate 85% of the time. Why is this important? Because that automated skill is better than the average teacher.
7. Make accurate and productive use of assessment
If we break assessment down into formative and summative, both, it can be argued, could benefit from AI.
Formative assessment is difficult and often of poor quality in one to many classrooms. It is largely absent from the lecture hall. There are three ways that AI could improve formative assessment, in ways superior to current ‘teacher’ delivery. First the quantity; adaptive learning systems, could deliver more feedback than teachers. It is self-evident that AI is scalable in the way a teacher is not and can deliver millions of pieces of feedback to millions of learners in milliseconds – this is what Google does already, albeit at a basic level. Second, AI could deliver higher quality feedback, which can also be used to determine what is literally delivered next in an online lesson. Formative assessment is one area where AI already excels. Increasingly we will also see, through AI, immediate feedback, delivered verbally or in text, as AI driven speech recognition and delivery becomes commonplace. With speech we will move towards the sort of frictionless interface than enables good teaching and learning.
On summative assessment, AI can deliver adaptive questioning and, using Item Response Theory, deliver assessment that includes learner confidence and other data during the assessment that no teacher could gather. It can also deliver to whatever statutory assessment requirements are in place. Essay marking is reaching a level, where it can perform as good as an expert assessor. Automated marking is also becoming more common. Online proctoring uses AI in typing patterns to identify the examinees, as is face recognition for digital identity and realtime face recognition, as the learner takes the exam.
Assessment is clearly one area where AI has, and will continue to make inroads. If teachers no longer have to set, and mark, homework, assignments and exams, a huge portion of the teaching task will have been automated.
8. Set high expectations which inspire, motivate and challenge pupils
This is tricky, as it is difficult to measure. No one would argue that technology is not motivating, even compelling, especially for the young. Whether it is motivating and compelling for learning is another question. With the emergence of the Smartphone, gamification, AR, VR and frictionless speech recognition, we already see signs of technology that is both powerful in terms of learning and compelling. Gamification certainly tackles the motivational issue. AR offers layered experiences. VR offers inspiring, complete immersion and attention, emotional pull, learning by doing, contextualised learning and high retention (witness flight sims for pilots). It promises to democratise something teachers in classrooms cannot offer - experiences. AI brings all sorts of additional dimensions to learning. Many aspects of good ‘learning theory’ are now being embedded in AI and used for teaching. I argue elsewhere, that we in the learning game, have a lot to learn from AI, as it starts to distinguish what is good from bad in learning theory.
9. Manage behaviour effectively
We could envisage a future where adult supervision of learning does not need highly qualified, subject-trained teachers but facilitators. We can even imagine the transformation of schools, into places where one learns, not places where teachers teach. This is a radical shift but as teaching becomes automated so schools become places of learning, not teaching. In practice this pendulum swing has started. There are plenty of online learning courses and degrees out there and learners are starting to do it for themselves. This can only continue. One argument is that it will accelerate, now that AI is in the field, as it has in other areas of human endeavour.
10. High standards of personal and professional conduct
One can hardly ask this of AI but perhaps it becomes less of a problem. However, the modelling that teachers provide is certainly important to young people. Teachers provide values and models of conduct that one hopes are emulated. Again however, we may see the development of attitudinal learning, that was never adequately delivered in classrooms, with simulations and the ability to put yourself in the shoes of others. AI is already feeding the VR industry with intelligent avatars, which are commonly used in games but increasingly in attitudinal learning. You become the bullied person, the subject of racism, sexism or bigotry.
Few saw self-driving, autonomous cars coming. That happened because of AI. Few may also see self-driving, autonomous learners. That may also come through AI. Machine learning not only embodies learning, it learns about learners while they learn. It is like a fast learning teacher. There is every reason to suppose that teaching may, to some degree, even a large degree, be automated by AI. If this is possible, it will be a momentous breakthrough, as education at all levels, whether schools, colleges, universities or workplaces, will experience massive reductions in costs. The developing world, where teaching is a problem in terms of quality and supply, is also likely to receive a huge boost. At some point we may look back at teachers and classrooms as we look back on manual manufacturing in factories. I’m not suggesting that teachers are in any way not valuable or smart, just that AI technology may, as in many other areas, get more valuable and smarter.