Educational Technology: From Teaching Machines to AI Tutors

Abstract

Every generation of computing has been offered to schools as the thing that would finally individualise teaching. Sidney Pressey built a mechanical “teaching machine” in the 1920s and sold about 160. B. F. Skinner built another after watching his daughter’s arithmetic lesson in 1953 and set off a programmed-instruction boom. Patrick Suppes at Stanford put children in front of computer terminals from 1966, PLATO built a teaching network in Illinois, and three student teachers in Minnesota wrote The Oregon Trail in 1971. Seymour Papert argued the opposite: that children should program the computer, not be programmed by it. Then came Apple IIe donations to every California school, laptops for every child, Khan Academy, MOOCs and chatbot tutors. The historian Larry Cuban described the recurring pattern in 1986: exhilarated promises, research showing modest effects, disappointment, and blame for teachers. The OECD found in 2015 that students who used computers very frequently at school did worse.

The Automatic Teacher

Sidney L. Pressey, a psychologist at Ohio State University, built the first teaching machine in the mid-1920s. It presented multiple-choice questions; the student pressed a key for an answer, and the machine could be set to advance only after the correct one, so that it taught as well as tested. Pressey described it in 1926 in School and Society as “a simple apparatus which gives tests and scores, and teaches.”

He could not sell it. Pressey approached at least six manufacturers in 1926 without success, and in 1929 signed with the W. M. Welch Manufacturing Company of Chicago, a maker of scientific instruments. Welch built about 250 “Automatic Teachers” and sold about 160 before shelving the project. The Depression arrived as the machine did, and teachers facing unemployment had little reason to welcome a device sold as a labour-saver. In 1932 Pressey still predicted an “industrial revolution in education.”

Skinner’s Box for Children

On 11 November 1953 the behaviourist B. F. Skinner visited his younger daughter’s fourth-grade arithmetic class at Shady Hill, a private school in Cambridge, Massachusetts, on a fathers’ visiting day. What he saw contradicted everything his laboratory work on reinforcement had taught him: all the children worked at the same pace, and they learned whether their answers were right only the next day, when the graded papers came back. Within days he built a prototype, a wooden box with sliders on which a child set a two-digit answer and a light that confirmed a correct one. He presented the idea at the University of Pittsburgh in March 1954, and in the same year published “The Science of Learning and the Art of Teaching” in the Harvard Educational Review.

Skinner’s version, programmed instruction, broke a subject into small steps, each requiring an active answer and giving immediate feedback, so that the student was reinforced dozens of times an hour and progressed at their own pace. Norman Crowder, who had trained Air Force technicians, devised a branching variant around 1959 to 1960, in which a wrong answer sent the student to a remedial frame instead of simply repeating the question. Around 1960 programmed instruction became a business, with programmed textbooks and machines sold to schools, and an “Autotutor” was demonstrated at the 1964 New York World’s Fair. The ideas fed into the computer-assisted instruction that followed.

Computers in the Classroom

The computer offered a teaching machine that could branch, keep records for every student and be reprogrammed. At Stanford the philosopher Patrick Suppes began research on computer-assisted instruction in January 1963, with a grant from the Carnegie Corporation. With funding from the U.S. Office of Education he set up a laboratory at Brentwood Elementary School in East Palo Alto, where the first pupils used an IBM 1500 instructional system on 27 October 1966; other schools received drill-and-practice lessons in arithmetic and spelling over teletype lines. In 1967 Suppes founded the Computer Curriculum Corporation to sell the lessons, which it did for decades.

At the University of Illinois, PLATO began in 1960 as a teaching system on the ILLIAC computer and became, by the 1970s, a network of thousands of terminals with a programming language for teachers, TUTOR, and an online community that its designers had not planned (see PLATO and the Birth of Online Community). Control Data Corporation spent heavily to sell PLATO as a commercial education product and never made it pay.

The most remembered educational program of the period was written for one history class. In 1971 Don Rawitsch, a student teacher from Carleton College in Minnesota, wanted his eighth-graders to experience the hardships of the westward migration. With two fellow student teachers, Bill Heinemann and Paul Dillenberger, he wrote The Oregon Trail in BASIC on an HP 2100 minicomputer, played on a teletype that printed each turn on paper; it debuted on 3 December 1971. The Minnesota Educational Computing Consortium (MECC), a state agency that ran a time-sharing network for the state’s schools, added an improved version to its system in 1975, and later versions for the Apple II, one of them redesigned by R. Philip Bouchard in 1985, carried it into schools across the United States (see The Home Computer Boom).

The Child Programs the Computer

Seymour Papert, a mathematician who had worked with Jean Piaget in Geneva and then at the MIT AI Laboratory, rejected the whole tradition. In the teaching-machine model the computer programmed the child. In his, the child programmed the computer: with the Logo language and its turtle, children would learn mathematics by building things in it, as they learned French by living in France. His Mindstorms (1980) became the manifesto of educational computing, and his The Children’s Machine (1993) asked why schools had nevertheless absorbed the computer into their existing routines instead of being changed by it (see Seymour Papert and Computer Science Education).

Papert’s students and colleagues carried the idea into simulation. Mitchel Resnick’s StarLogo let students program many “turtles” at once. Uri Wilensky released NetLogo in 1999 at Northwestern University’s Center for Connected Learning and Computer-Based Modeling: a multi-agent modelling environment in which students give instructions to hundreds or thousands of independent agents and watch how patterns at the level of the whole system emerge from individual behaviour. NetLogo, the next generation of the StarLogo line, is used for simulating natural and social phenomena (see Cellular Automata and Artificial Life).

A Computer in Every School

The hardware came to schools through marketing as much as pedagogy. In 1982 Steve Jobs lobbied Congress for a bill, introduced by Representative Pete Stark, giving corporations a tax deduction for donating computers to schools. It passed the House 323 to 62 and stalled in the Senate. California passed its own version, AB 3194, signed by Governor Jerry Brown in September 1982, with a 25 percent state tax credit. Apple’s response, Kids Can’t Wait, offered an Apple IIe with monitor, floppy drive, Apple Logo and software coupons to each of about 9,000 eligible California schools in 1983. Creative Computing estimated the retail value at about $21 million and the net cost to Apple, after the credit, at about $1 million. A generation of Californian schoolchildren learned on Apples, and Apple led the education market for years. In Britain the BBC’s Computer Literacy Project brought the BBC Micro, launched in December 1981, into schools (see The UK Computing Industry).

Twenty years later the same ambition went global. One Laptop per Child, launched by Nicholas Negroponte in 2005, promised a $100 laptop for every child in the developing world; the machines cost around twice that, and studies of the largest deployments found little effect on test scores (see Nicholas Negroponte).

Video, MOOCs and Tutors

The internet revived the teaching machine in new forms. In 2004 Sal Khan, a hedge-fund analyst, began tutoring a young cousin in mathematics remotely, recorded short explanatory videos for her and posted them online, and in 2008 founded Khan Academy, a non-profit with thousands of free videos and exercises, funded by philanthropy including the Gates Foundation and Google. Its practice exercises gave immediate feedback and let students advance at their own pace, as Skinner had proposed.

In 2011 and 2012 Stanford and MIT professors launched massive open online courses (MOOCs) through Coursera, Udacity and edX, with enrolments in the hundreds of thousands per course, and the press called 2012 “the year of the MOOC” (see Andrew Ng and AI Education). Completion was the weak point. A 2015 analysis of 221 MOOCs by Katy Jordan found a median completion rate of 12.6 percent, and a 2019 study of Harvard and MIT courses on edX by Justin Reich and José Ruipérez-Valiente in Science found that low completion rates had not improved in six years.

The COVID-19 pandemic closed schools for more than a billion students in spring 2020, according to UNESCO, and turned remote teaching from an experiment into the default for months. Students without devices or connections fell furthest behind. From 2023 the large language models produced the newest version of the individual tutor: Khan Academy piloted an AI tutor and teaching assistant, Khanmigo, and students worldwide used general chatbots for homework, whether their schools allowed it or not.

Dead End: The Machine That Replaces the Teacher

In Teachers and Machines: The Classroom Use of Technology Since 1920 (1986), the Stanford historian and former teacher Larry Cuban traced film, radio, television and early computers through the same cycle: promises that the new medium would transform schooling, research that found modest or no gains, disappointment, and then blame directed at teachers for failing to use it. His Oversold and Underused (2001), a study of Silicon Valley schools with abundant equipment, found that teachers and students used computers far less in class than at home.

The OECD’s 2015 report Students, Computers and Learning, based on the 2012 PISA tests, gave the numbers. Countries that had invested heavily in computers for schools showed no noticeable improvement in reading, mathematics or science. Students who used computers moderately at school did slightly better than those who used them rarely, while students who used them very frequently did a lot worse, even after accounting for social background. The OECD did not conclude that technology caused the lower results, only that it had not delivered what had been claimed for it.

What the teaching-machine tradition kept getting right was feedback: immediate correction and self-paced practice help, from Pressey’s box to Khan Academy’s exercises. What it kept getting wrong was the claim that the machine could replace the teacher. Each generation’s tools ended up absorbed into ordinary classrooms as aids, which is what Papert’s The Children’s Machine had predicted and regretted.