Systematic review of research on artificial intelligence applications in higher education – where are the educators?

Artificial intelligence in education has been an active research field for about thirty years. And yet, educators still don't know how to use it well in their classrooms. That gap is the starting point for a systematic review by Zawacki-Richter and colleagues, who screened two thousand six hundred fifty-six publications from 2007 to 2018 and synthesized one hundred forty-six peer-reviewed articles to map what we actually know about AI in higher education. First, some vocabulary. AIEd, which stands for artificial intelligence in education, covers systems that perform cognitive tasks usually associated with human minds. The main subfields are intelligent tutoring systems, which simulate one-on-one tutoring; learning analytics, which mine student data for patterns; and adaptive learning systems, which tailor content to individual learners. Machine learning sits underneath most of these — software that finds patterns and makes predictions. Now, who is building these systems? Mostly computer scientists. Zawacki-Richter and colleagues found that sixty-one of one hundred forty-six first authors came from computer science departments, another twenty-nine from STEM fields, and just nine from education. Half of all articles came from four countries: the United States, China, Taiwan, and Turkey. The methods match the demographics. Seventy-three percent of empirical studies used quantitative approaches, one study was qualitative, and only eight used mixed methods. The review's title asks the obvious question: where are the educators? The research that does exist falls into four application areas. In profiling and prediction, every study used machine learning, including neural networks, support vector machines, random forests, and decision trees. In every comparison, machine learning outperformed logistic regression. Acikkar and Akay's support vector machine predicted student admissions with an accuracy of ninety-three point eight percent. In assessment and evaluation, automated essay scoring agreed with human raters between ninety-four point six and ninety-eight point two percent of the time, according to Gierl and colleagues. Adaptive systems mostly operate at the course level. Schiaffino and colleagues' eTeacher profiles student behavior and recommends personalized reading and exercises. Intelligent tutoring systems like MetaTutor teach both content and self-regulation strategies. A meta-analysis by Steenbergen-Hu and Cooper found they produce moderate learning gains, outperforming most instruction modes but falling short of human tutoring. Those are the technical wins. Here's what the field is getting wrong. Only five of one hundred forty-six articles explicitly defined artificial intelligence. Only two critically reflected on ethical implications. Li flagged privacy and discrimination risks. Welham raised costs and implementation time. Selwyn warned about surveillance, including face-recognition systems. More than forty percent of studies had no theoretical grounding in pedagogy at all. So, what does good practice look like? The review points to interdisciplinary teams that include educators, explicit ethical frameworks — Prinsloo's ethics of care gets a mention — mixed methods that capture student and teacher voice, and design-based research approaches. Dikli's interviews with English as a Second Language students about automated versus human feedback are held up as a model. The technical infrastructure for AI in higher education is genuinely impressive. Building it without educators in the room is how you end up with thirty years of research and still no clear answer on how to teach with it. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.