What Was Programmed Instruction, and Why Did It Disappear?

Programmed instruction decomposes a subject into thousands of single-step frames, requires a response at each frame, gives immediate feedback, and blocks the learner from advancing until the response is correct. The method was invented by Pressey (1926), formalized by Skinner (1954), branched by Crowder (1958), and computerized by Bitzer at Illinois (PLATO, 1960). Meta-analyses found positive effects on college-level achievement (Kulik, Cohen, and Ebeling, 1980) and stronger effects, around half a standard deviation, for mastery-learning variants (Kulik, Kulik, and Bangert-Drowns, 1990). The movement died by the mid-1970s because authoring a course-length program took a subject-matter expert a year or more of full-time work, and the result was fixed at publication. Large language models eliminate that constraint by generating frames, branches, and remediation in real time from each learner's specific response, and by updating instantly when underlying content changes. No current EA exam prep vendor ships a product built this way. We are bringing it back. Stay tuned.

In 1912, Edward L. Thorndike imagined a book that would reveal page two only to a reader who had done what page one directed. Much of what then required a live teacher, he argued, could be handled by print if print could be made conditional on reader response. The idea sat for over a decade.

Sidney L. Pressey, a psychology professor at Ohio State, took up the problem in the early 1920s. He built a small device out of typewriter parts that displayed multiple-choice questions one at a time and required the student to press a key to record an answer. Correct answers advanced the drum. Incorrect answers did not. The device kept a running count.

Pressey demonstrated the machine at the 1924 and 1925 meetings of the American Psychological Association and published the first paper on it in School and Society in 1926, under the title "A Simple Apparatus Which Gives Tests and Scores -- and Teaches." A second machine, published in 1927, dropped a question from the rotation once the student had answered it correctly twice in succession, anticipating by half a century the adaptive scheduling logic of modern spaced-repetition software.

In 1932 Pressey predicted his device would bring an industrial revolution to education. The Great Depression killed the commercial market, Pressey could not sustain investor interest, and the field went dormant for two decades.

In 1954, B. F. Skinner published "The Science of Learning and the Art of Teaching" in the Harvard Educational Review. The paper applied his work on operant conditioning -- behavior shaped by its consequences -- to classroom instruction. Skinner argued that learning happens when a learner produces a response and the response is immediately followed by reinforcement, that reinforcement loses power if it is delayed, and that traditional classroom teaching was structurally incapable of delivering enough immediate reinforcement to shape behavior in 30 students at once.

Skinner's solution was a teaching machine more sophisticated than Pressey's, and a method for writing curriculum to feed it. The method came to be called linear programmed instruction, and its unit of construction was the frame.

A frame presented a small piece of content, no more than a sentence or two, and required the student to produce a response, usually by filling in a blank. The next frame confirmed the correct answer and added a small piece of new content. Each frame moved the student forward by a tiny increment. The principle was that the steps had to be small enough that the student almost always answered correctly, because each correct answer was a reinforcement, and reinforcement is what shaped retention.

Skinner and James Holland produced a 1961 self-instruction text built this way, The Analysis of Behavior, that taught the principles of behavior analysis to thousands of Harvard undergraduates and later to a broader audience. The B. F. Skinner Foundation still distributes a version of it.

The construction principles were specific and demanding. A well-written linear program had a vanishingly low error rate -- Holland's quantitative measure of program quality treated error as evidence of bad writing, not bad students. Frames had to build on each other in a logical sequence with no leaps. New terms had to be defined before they were used. Responses had to require recall, not recognition. Every learning objective in the course had to be decomposed into a chain of micro-objectives, each one its own frame.

A few years after Skinner, Norman Crowder, an Air Force training researcher, developed an alternative he called intrinsic programming. The frame presented more content -- often a half page or full page -- followed by a multiple-choice question. A correct answer advanced the student to the next main frame. An incorrect answer routed the student to a remedial sequence that diagnosed the specific error before returning the student to the main path.

Crowder published his intrinsic programming work in 1959 and 1962, and Doubleday released his TutorText series the same year. Because the pages of a TutorText did not flow in order -- a correct answer might send the reader to page 47, an incorrect answer to page 112 -- the format was sometimes called a "scrambled book." Crowder's AutoTutor, a desktop electromechanical device released in the early 1960s, provided the same branching logic with a film reel and a few buttons.

Crowder rejected Skinner's behaviorist framing. He treated a program as communication between an author and a learner, with branching as the mechanism for adapting communication to the listener. He accepted errors as informative -- a 20 percent error rate was acceptable if the program used the errors to diagnose and remediate. Skinner, who treated error as evidence of frame failure, found this approach loose. The two camps argued for two decades. Both were nearly gone by the 1970s.

While Skinner and Crowder argued, Donald Bitzer at the University of Illinois was building a computer-based version. PLATO -- Programmed Logic for Automatic Teaching Operations -- launched on the ILLIAC I in 1960 and grew over the next two decades into the first general-purpose computer-assisted instruction system. By 1972, PLATO IV terminals featured a plasma display panel that Bitzer had co-invented for the purpose. By the late 1970s, PLATO had thousands of terminals worldwide and an online community that prefigured most of the social internet.

PLATO ran lessons across every level, from elementary arithmetic to nursing maternity care to advanced foreign languages. It was the closest thing the movement produced to a working proof of concept at scale, and it was extraordinarily expensive. The hardware cost a fortune. The authoring cost was worse.

In 1980, James Kulik, Peter Cohen, and Barbara Ebeling published a meta-analysis in Educational Evaluation and Policy Analysis pooling controlled comparisons of programmed instruction against conventional teaching in college courses. The average effect on achievement was small and positive. Programmed instruction also showed reductions in instructional time, in some studies substantial. The 1982 follow-up on secondary schools (Kulik, Schwalb, and Kulik) found effects close to zero.

The 1979 meta-analysis of Keller's Personalized System of Instruction -- a closely related mastery-learning approach that used programmed materials with a peer-proctor system -- showed larger and more consistent gains. Later meta-analyses placed the average mastery-learning effect at roughly half a standard deviation, with the strongest effects on weaker students.

In 1984, Benjamin Bloom published the 2-sigma problem paper in Educational Researcher, reporting that students tutored one-to-one with mastery learning scored roughly two standard deviations above conventionally taught students. Programmed instruction was Skinner's attempt at a mechanical proxy for one-to-one tutoring. The 1980 Kulik effect sizes suggest the proxy captured a small fraction of the tutoring effect. Bloom himself observed that one-to-one tutoring was too expensive to scale and challenged the field to find a method as effective as tutoring that could. The programmed instruction movement did not meet that challenge.

Three explanations are commonly offered.

The behaviorist framing fell out of academic fashion in the cognitive revolution of the late 1960s and early 1970s. Skinner himself believed this was the main cause. The cognitive revolution, however, did not disprove anything in the empirical record. It made the vocabulary unfashionable.

Students disliked the machines. Paul Saettler's history of educational technology reports that learners found programmed instruction tedious and sometimes broke the equipment. Linear programming in particular, with its incremental frames, produced a flat, repetitive experience.

The empirical results were less impressive than the founders had advertised. A small positive effect in higher education and a near-null effect in secondary education was not the industrial revolution Pressey and Skinner had predicted.

The binding constraint, examined in detail in Jason McDonald's 2003 Brigham Young University thesis and his 2005 paper with Stephen Yanchar and Russell Osguthorpe, was authoring cost. Writing a course-length linear program required decomposing a body of knowledge into thousands of micro-objectives, each one rendered into a frame that would elicit the right response from a learner who had seen only the previous frames. A typical program took a subject-matter expert a year or more of full-time work, paired with a programmer-of-frames. Branching programs were worse, because every branch had to be authored, every remedial path written, every error anticipated.

Once published, the program was fixed. A new tax year, a new exam blueprint, a corrected statute meant rewriting frames by hand. The unit economics worked only for subjects with stable content and very large audiences, and even there the margins were thin.

PLATO ran into the same wall. Programs ran on million-dollar mainframes and required custom authoring in TUTOR, a programming language built for the system. Control Data Corporation never made it profitable.

When the personal computer arrived in the early 1980s, the programmed instruction model could in principle have migrated to it. What migrated were the easier and cheaper things: drill-and-practice software, gradebooks, presentation tools. The serious frame-by-frame branching program never made the jump because nobody could afford to author it again from scratch.

The cognitive science of the last 30 years has largely vindicated the empirical instincts of the programmed instruction founders while abandoning their theoretical scaffolding.

Karpicke and Roediger's 2008 Science paper on the testing effect, which I covered in an earlier note, is a re-derivation of the central programmed instruction claim: producing a response strengthens memory in a way that reading does not. The 2013 Dunlosky review identified two study techniques out of ten with high empirical support; one was practice testing, which is what every frame in a programmed instruction text is. Bloom's 2-sigma result has not been replicated at full strength, but mastery learning consistently shows effects around half a standard deviation, and high-quality one-to-one tutoring approaches the Bloom number.

The core mechanism programmed instruction exploited was real. The reason the movement did not deliver on its promise was that authoring cost made the system rigid, expensive, and impossible to keep current.

That constraint has dissolved.

A large language model can generate a frame, a response prompt, a corrective explanation, and a follow-up frame conditioned on the student's specific error, in seconds. It can branch the way Crowder wanted, not from a precompiled tree of possible errors but in real time, from the actual response the student produced. It can rewrite a course when the underlying material changes. It can adjust the level of granularity to the individual learner: bigger frames for someone moving fast, smaller frames for someone struggling.

What an LLM cannot do, at least not yet, is design the pedagogy. It cannot decide on its own what objectives a course should cover, in what order, with what kinds of frame-level interactions, against what mastery criteria. That is human work, and it requires a subject-matter expert who understands the underlying body of knowledge and a designer who understands the architecture of programmed instruction.

Once that scaffolding is in place, the LLM can do the work that broke the original movement: the thousand-frames-an-hour authoring, the dynamic branching, the real-time remediation, the continuous updating. The unit economics that defeated Pressey, Skinner, Crowder, Bitzer, and the programmed instruction publishing industry have changed.

The Special Enrollment Examination is the kind of test programmed instruction was built for. The content is large but bounded. The objectives are public. The cognitive level is mostly recall and application, not synthesis. The dollar thresholds, depreciation lives, due dates, and code-section concepts are the kind of declarative knowledge that frame-by-frame programmed instruction handled well in the 1960s, when the limiting factor was authoring time and not pedagogical fit.

The current EA exam prep market -- Gleim, Surgent, Fast Forward Academy, Becker, and a long tail of smaller players -- does not, as far as I can tell, ship anything that meets the programmed instruction definition. The standard product is a textbook plus a question bank plus, at the higher tiers, video lectures and adaptive quizzing. The question bank performs some of the testing-effect work. The adaptive quizzing performs some of the branching work. Nothing in the market reconstructs Skinner's frame-level architecture, where the lesson itself is built out of small, sequenced response-eliciting steps and the student does not advance until they have produced the right response.

That is a gap.