Background
Artificial Intelligence (AI) is a disruptive technology. When disruptive technology emerges, older ways of work become obsolete whilst new opportunities arise. If organisations fail to adopt these new technologies, they cease to be cost-effective or efficient. Effective AI is undoubtedly the largest technological change we have undergone since the advent of the Internet. Potentially replacing huge numbers of jobs with automated solutions that can not only do the jobs for a fraction of the cost of employing a human; sometimes more accurately and reliably, and considerably faster. AI is no different; it is expected that AI in U.S. education will grow by 47.5% from 2017-2021 according to the Artificial Intelligence Market in the US Education Sector report. This blog will focus on the following:
- How is AI applied to learning and development?
- AI in education accessibility
- The role of AI in student support
So, what do we mean by AI?
Traditional software uses algorithms to make decisions. These decisions often resemble a flow chart, where experts have decided which outcomes are appropriate for given inputs, often based on a number of intermediate steps. These algorithms are often very efficient at processing small numbers of inputs. Such software responsible for automating tasks within educational establishments has been around for a long time – in such diverse areas as lesson scheduling, attendance monitoring, or presentation / whiteboard software – but this isn’t (necessarily) AI. However, with a hand-crafted algorithm, it’s possible to follow all the decisions made to work out why an output was chosen for a given input. Engineers can then modify the algorithm to iteratively improve its performance.
AI, on the other hand, can make decisions without the need for experts to explicitly build a decision tree or algorithm. The training process is modeled on how the human brain works. None of us are born knowing what a car looks like, but by being shown a few examples, we can then identify new instances on our own. Once trained, an AI with hundreds, thousands or even millions of inputs will know how each input relates to each other, and can determine, statistically, the most likely outcome for any given input set.
AI can take millions of pixels in a digital X-ray and tell you that it shows a patient with a rare disorder, out-performing some of the best radiographers in the world. It can take the chemical makeup of a protein and tell you the most likely shape it forms (something called ‘protein folding’) – impossible a decade ago. AI can identify objects in images, spoken text in audio, patterns in complex multi-dimensional data, can understand relationships between words, translate text, convincingly speak and do myriad other things besides – and we are at the very beginning of this journey. More specifically, this type of AI is called “Machine Learning” (ML) and is different from the generalized AI that we often see in science fiction. There is no danger of Machine Learning becoming self-aware, determining its own destiny and sending a robot back in time to kill John Connor. There are many other discussions about Generalized AI, but for the purposes of this article, we won’t discuss it here.
why don’t all algorithms now use AI then?
AI is opaque – it is never clear why, for a given input, a specific output was selected. It is also challenging and sometimes not obvious how to improve the performance of a particular model. It is easy to trick by providing inputs that are unusual. Machine Learning contains no ‘common sense’, just statistical likelihood based on the training it’s received.
Training a model requires huge datasets and a large amounts of computing resource. Google use billions of images to train their image-detection AI. Even where it makes sense to use Machine Learning, the outputs of such a system are usually fed into more traditional software algorithms to further process the data and do something useful with it. More often AI/ML forms a small component of a system, rather than its whole.
So how does AI apply to education?
No more ‘one size fits all’
In the past, unless a student was lucky enough to afford 1-to-1 tuition, “One size fits all” was the only practical approach to education. Students all go through a similar programme of lectures, practice, tests and exams until graduation. However, this ignores the fact that students start with different levels of knowledge, they respond in different ways to various learning styles, learn at different rates, and often have other commitments.
Intelligent tutoring systems (ITS) are an example of rapidly developing education-based AI. There is evidence to suggest that ITS systems perform as well, if not better, than individual human tutors for many students. An example is Carnegie Learning’s “Mika” software, which uses cognitive science coupled with ML to provide personalized tuition and real-time feedback for post-secondary education students, particularly first-year college students who would otherwise need additional academic support. Carnegie estimates that the annual cost of additional academic support could be as high as $6.7 billion and low success rates (33%) with maths students. Whereas, ITS provides a new less expensive, flexible avenue for students to receive personalized education over a long-term basis.
Duolingo, a popular language-learning application, also supports flexible teaching methodology. Duolingo uses Machine Learning to predict the probability of remembering certain words and concepts and then uses this information to repeat concepts with a frequency according to each students’ retention and capability. Educational establishments can use feedback from automatically-set multiple-choice tests and automatically repeat and re-enforce specific topics on a per-student basis, especially when the content is delivered digitally, rather than requiring a lecture hall.
Combing the above examples, AI could monitor a student’s performance in specific skills across a wide variety of subjects each year, whilst automatically providing new content or specified learning parameters which could help meet students’ need for continual, targeted practice and feedback. With this information, teachers could better understand student performance and enact more effective personalized learning strategies. This approach is also applicable to corporate organisations such as Invensis which could tailor training plans to company-clients as determined by the AI analytics.
Marking and Setting Tests
Many of the issues surrounding education are due to availability of educators – there often aren’t enough hours in the day for all students to be mentored and have their questions answered – especially when they are struggling.
With this in mind, routine tasks such as marking tests are obvious candidates for automation – leaving staff additional time to produce lessons and offer student support.
Multiple-choice tests have long been processed automatically. Without any AI, it’s trivial for a computer to see whether you’ve selected the right answer when it’s one of a discreet set of values. But what about essays? Well, much as it sounds like science fiction, AI is now being used to mark essays at college level in Utah and Ohio.
Although the technology is primarily spotting what looks like a good essay structure based on millions of human-graded examples, it will almost certainly improve to the point where it can understand themes and more nuanced content. However, there is a long way to go for any AI model to appreciate creative uses of language, and using a “robo-grader” comes with a large number of caveats, as demonstrated by Les Perelman’s Babel generator, which can fool the system with gibberish specifically designed to get a great robo-grader score.
AI isn’t just useful for marking essays; it can be used to collate the answers from multiple-choice tests. It can spot patterns in results, weak spots in individual student’s cognition, or even topic areas that are testing badly across the board. Educators can then improve their course and feasibly automatically generate new material and tests based on these weak areas.
Accessibility and Online Study
The traditional reach of a university is those who can attend its physical location, can speak its language and who can afford its fees – but those times are changing. Online courses such as those offered by Coursera (set up by Stanford University), Udemy, Khan Academy, and more traditional institutions such as The Open University, educate students from all around the world with many different backgrounds and abilities.
Their reach can be enhanced by AI systems which can automatically translate course materials into a wide variety of languages. As these systems improve, they will amplify the reach of good teachers to a much wider audience.
These online courses are driving other innovations, which often use AI. If a course has thousands of students, systems which create well-matched study groups also become helpful. These systems can ensure the skills within each group are balanced, so that students can help each other.
Student Support
Chatbots
Students can now interact with chatbots to answer common questions about their course, finance, societies, timetables and many other topics. Beacon, in use at Staffordshire University, is one such bot; Jill Watson is another, in use at Georgia Tech. Only if a question can’t be answered automatically (within a chosen degree of certainty) will a human be connected to the conversation. By doing this, students still get the help they need, but common questions and answers don’t unnecessarily overload the faculty staff.
Cheating
Students who cheat are a blight on educational establishments. They undermine the purpose of education and devalue the efforts of honest students. AI has been successfully used to mitigate these issues.
Plagiarism detection
Traditional mechanisms of spotting plagiarism within submitted essays, dissertations and other academic papers have long been broken. It’s no longer good enough to simply look for similar text in previously-submitted works, as there are too many ways to subtly alter text and pass the test.
To combat this problem, AI has been successfully deployed to spot plagiarism more effectively. Natural language processing and other techniques can spot wider similarities between documents, even when words or sentences have been changed substantially. It’s even possible to use AI to spot whether two completely different pieces of text are likely authored by the same person.
Proctoring
When a student attends an exam in real life, their identity can be verified by an invigilator. However, if a test is taken online, how can we tell for certain who is there? Once again, AI provides some solutions. There are AI systems that can spot copy-pasted text whilst facial recognition and monitoring can spot unusual patterns in those taking online tests.
Conclusion
I’ve covered many areas where AI will improve the efficacy of education. Rather than the common cry “Robots are taking our jobs!” the trend in AI is towards efficiency and allowing good teachers to reach a wider audience, whilst reducing their administrative headaches.
Many of the technologies I’ve listed in this article are in their infancy, and will continue to improve dramatically over the coming years. And whilst it’s becoming easier to see a future where an AI could teach a class in isolation; education should be about far more than just absorbing knowledge – true insight comes from the collaborations we have with other students and teachers.