There is a common assumption circulating in educational circles, in staff rooms, at conferences, and across social media, that artificial intelligence arrived in education with the launch of ChatGPT in November 2022. For many teachers and trainers, that moment felt like a sudden rupture, a before and after. The reality, however, is considerably more complicated, and considerably more interesting. AI has been present in education in various forms for well over sixty years, and the current fixation on generative AI risks obscuring a much richer and more instructive history.

A recent piece on Substack by Nick Potkalitsky, titled "If Testing Companies Use AI to Grade, Why Can't We?", makes this point with particular clarity [1]. Potkalitsky describes a workshop in which teachers were alarmed to learn that Ohio's standardised tests were being graded by AI. The assumption in the room was that this meant something like ChatGPT was reading and evaluating student essays. The reality was quite different. Ohio uses a system called Autoscore, developed by Cambium Assessment, which relies on a technique called Latent Semantic Analysis (LSA) - a method that dates back to the 1990s [1]. This is not the same technology that writes essays on demand. It is a tool that classifies and scores existing text, a fundamentally different operation. The confusion in that workshop is, unfortunately, entirely typical of the current moment.

The technical distinction at the heart of this confusion is the difference between discriminative AIandgenerative AI. These are not simply two points on a spectrum; they represent different approaches to what AI is asked to do.

Generative AI, the category that includes ChatGPT, Claude, and similar tools, learns patterns from enormous datasets and uses those patterns to produce new content. Given a prompt, it generates text, images, or other outputs. It is, in essence, a sophisticated content creator. Discriminative AI, by contrast, takes existing data and classifies it. It learns to distinguish between categories and then applies that knowledge to new examples. Spam filters, fraud detection systems, and automated essay scoring tools are all examples of discriminative AI at work. The Ohio system that caused such alarm in Potkalitsky's workshop does not write anything; it reads student essays and assigns a score, much as a human marker would, but using statistical patterns learned from thousands of previously scored examples.

The table below sets out the key differences between these two categories, along with examples relevant to education:

Dimension	Generative AI	Discriminative AI
Core function	Produces new content from patterns in training data	Classifies or scores existing data
Familiar examples	ChatGPT, Copilot, Gemini	Spam filters, plagiarism detectors, automated essay scoring
Educational applications	Drafting lesson plans, generating practice questions, tutoring dialogue	Grading essays, identifying at-risk learners, detecting plagiarism
Main concern in education	Academic integrity, deskilling, over-reliance	Bias in scoring, lack of transparency, fairness for diverse learners

Understanding this distinction matters not just as a point of technical accuracy, but because the two types of AI raise very different questions for educators and policy makers. The concerns about generative AI - that students might submit AI-written work as their own, or that the ability to write might atrophy through disuse - are real and worth taking seriously. But they are not the same concerns as those raised by discriminative AI, which centre on whether automated scoring systems treat all students fairly, particularly those from non-English-speaking backgrounds. Potkalitsky cites research showing that AI scoring models trained primarily on native-speaker writing can systematically undervalue essays by English language learners, assigning scores up to 10.3% lower than human raters would give for work of equivalent quality [1]. These are different problems requiring different responses, and conflating the two types of AI makes it harder to address either.

The history of AI in education is a long one. The first computer-aided instruction systems appeared in the 1960s. The PLATO system (Programmed Logic for Automatic Teaching Operations), launched at the University of Illinois in 1960, delivered interactive lessons on mainframe computers and introduced features that would later become standard in online learning: instant feedback, online quizzes, and even early forms of discussion forums [2]. Around the same time, the behavioural psychologist B.F. Skinner was developing his "teaching machines," devices designed to automate and individualise instruction so that students could progress at their own pace and receive immediate feedback on their answers. These were not AI in the modern sense, but they embodied the same underlying aspiration: to use technology to personalise learning.

The 1970s and 1980s saw the development of Intelligent Tutoring Systems (ITS), which brought more sophisticated AI techniques to bear on the problem of personalised instruction. The LISP Tutor, created at Carnegie Mellon University in 1983, could identify student errors in real time and provide targeted feedback as learners worked through programming exercises [3]. It built a model of each individual student's knowledge and used that model to decide what to teach next, an approach that anticipates the adaptive learning platforms of today. These systems used simple machine learning to construct what we might now call a "digital profile" of each learner, tracking their progress and predicting where they were likely to struggle.

The 1990s brought two developments of particular relevance to the current debate. First, Automated Essay Scoring systems began to be used in large-scale assessments. The Intelligent Essay Assessor, which used Latent Semantic Analysis to evaluate the conceptual content of student writing, was among the first [4]. Educational Testing Service (ETS) introduced its e-rater system in the late 1990s, which has been used as a co-rater in standardised writing assessments ever since. These are the direct ancestors of the Ohio system that caused such consternation in Potkalitsky's workshop. Second, the first Learning Management Systems appeared. Blackboard launched in 1997 and the Open Source Moodle platform followed in 2002 [3]. These platforms incorporated basic AI functions from the outset, including automated grading of multiple-choice questions and rule-based alerts when student engagement dropped below a threshold.

The 2000s and 2010s were characterised by the growth of adaptive learning platforms and the emergence of learning analytics as a recognised field. Adaptive learning systems such as Knewton and DreamBox used AI to create genuinely personalised learning paths, adjusting the difficulty and sequencing of content in response to each student's performance [5]. Learning analytics, meanwhile, applied data mining and machine learning to the vast quantities of data generated by digital learning environments, with a particular focus on identifying students at risk of disengaging or dropping out [6]. Early warning systems of this kind are now widely used across higher education and, increasingly, in vocational training.

For those working in vocational education and training, this historical perspective is more than an academic curiosity. VET has its own particular relationship with technology, and the sector has been using various forms of AI-adjacent tools for longer than is often acknowledged. Competency-based assessment, which is central to many VET qualifications, lends itself naturally to automated scoring approaches: the criteria are explicit, the performance indicators are defined, and the task of a discriminative AI system is to determine whether a piece of evidence meets those criteria. This is a very different challenge from assessing the quality of an argumentative essay, and it is one that AI tools are, in principle, well suited to support.

Intelligent tutoring systems are claimed to be relevant to VET through providing targeted guidance to for example apprentices learning a complex technical process and who can benefit from a system that tracks their progress through each step, identifies where they are making errors, without requiring a human supervisor to be present at every moment. This is seen not be a replacement for skilled workplace mentoring but a supplement of the kind that has been available in various forms since the 1980s.

The broader point is that the arrival of generative AI, for all its genuine novelty, does not represent the beginning of AI in education. It represents a new chapter in a story that has been unfolding for decades. Treating it as though it were the whole story leads to poor policy, poor pedagogy, and a failure to learn from what has already been tried, tested, and, in some cases, found wanting. The questions that Potkalitsky urges us to ask of any AI system - what kind of AI is this, what was it trained on, what validation has been done, and who might be disadvantaged by it - are not new questions. They are the questions that researchers and practitioners have been grappling with since the first automated scoring systems appeared in the 1990s. The current moment of heightened attention to AI in education is an opportunity to engage with those questions more seriously.

References

[1] Potkalitsky, N. (2026, February 19). If Testing Companies Use AI to Grade, Why Can't We? Substack. https://nickpotkalitsky.substack.com/p/if-testing-companies-use-ai-to-grade

[2] Paccone, P. (2025, July 3). The History of AI in Education. Medium. https://ppaccone.medium.com/the-history-of-ai-in-education-7305b6f30a39

[3] eLearning Industry. (2025, July 28). The History Of AI: eLearning Edition. https://elearningindustry.com/the-history-of-ai-elearning-edition

[4] Foltz, P. W., & Landauer, T. K. (1999). The Intelligent Essay Assessor. Interactive Multimedia Electronic Journal of Computer-Enhanced Learning, 2(2). http://imej.wfu.edu/articles/1999/2/04/printver.asp

[5] University of Illinois College of Education. (2024, November 11). Traditional AI vs. Generative AI: What's the Difference? https://education.illinois.edu/about/news-events/news/article/2024/11/11/what-is-generative-ai-vs-ai

[6] Mustofa, S., et al. (2025). A novel AI-driven model for student dropout risk analysis. Computers and Education: Artificial Intelligence. https://www.sciencedirect.com/science/article/pii/S2666920X24001553

About the image

This digital collage modifies Tyrone Comfort’s 1934 painting “Gold is where you find it”. It connects mining with large-scale data extraction. A miner working underground blends into an endless grid of images scraped from the internet to train artificial intelligence without consent. The image links past and present forms of extraction, showing how data, memory and culture are treated as raw material for profit.