Could B.F. Skinner—the archetypical behaviorist psychologist—have been right…or at least half-right?
© Howard Gardner 2026
B. F. Skinner with a pigeon in an experimental laboratory, demonstrating his work on operant conditioning.
When a young scholar first learns about a discipline or “field,” that scholar is likely to adopt the judgments—and (not infrequently) the prejudices—of leading figures in that specialty. And in most cases, the novice (unless a habitual contrarian) is likely to echo these judgments, perhaps even uncritically. For example, if historians-in-training work primarily with Marxists—or for that matter, critics of Marxism—those trainees are likely to be sympathetic to the pervasive perspective. Or, to switch fields, if a budding psychiatrist is trained primarily by Freudians, that trainee is likely to be sympathetic to psychoanalytic perspectives, concepts, and practices.
As a student learning about psychology in the 1960s—particularly its cognitive and developmental strands—I was no exception. I was swept up by the developmental perspective of Swiss psychologist Jean Piaget as well as the cognitive focus of Jerome Bruner, my teacher and mentor at Harvard. Also, at the time, as a student living in Cambridge, Massachusetts, I was strongly influenced by the perspective of linguist Noam Chomsky at nearby MIT—then emerging not only as the leader of a “structuralist” approach to the study of language, but also as a severe critic of the dominant “behaviorist” school of psychology. When in 1971, Chomsky published brutal attacks on Skinnerian approaches to language (see here), I cheered from the sidelines; and, when in 1976, Jerome Bruner took on the entire Skinnerian enterprise in psychology (see here), I applauded him as well.
What was the contrasting, pure Skinnerian view? As a dyed-in-the wool behaviorist, Skinner believed that the psychological actions exhibited by rats, pigeons, dogs and other animals were adequate to explain the behaviors of human beings (as well as other primates). If you could describe, understand, and model the behaviors or actions of the range of species, there was no need to develop or deploy other “special” or “dedicated” psychological mechanisms to explain how human beings behaved or acted or thought. Indeed, even the concepts of “thinking,” “reasoning,” “problem-solving,” (let alone “problem finding” or “creativity”) were mistaken and unnecessary concepts in accounting for the accomplishments of the human mind. Indeed, Skinner would probably have objected to the concept of “mind”—and here I am, sixty years later, having authored over a dozen books featuring the word “mind” in their title!
(Not that this critique was personal. Skinner was a neighbor and we were cordial to one another. He sometimes accompanied me when I walked through the neighborhood with our young son Benjamin. The by-then-elderly Skinner would tease me by saying: “One day, you will see through all this cognitive baloney.” Perhaps he was prescient! He was certainly a true believer in his psychology and his philosophy!)
For decades, I retained my views—and my prejudices. Indeed, over a decade ago, I even engaged in a debate with Henry D. Schlinger—a scholar of strict Skinnerian leanings. I parroted back some of the arguments I had gleaned from my mentors—Piaget, Bruner, and Chomsky. (See the blog I posted afterward here.)
But now, with the advent of artificial intelligence (AI), and the possibility of artificial general intelligence (AGI), it’s high time to re-examine my loyalties…and, to the extent that seems warranted, to embrace aspects of a behaviorist or Skinnerian perspective. I hope that this conclusion is a reasonable one—not one engendered by creeping senility or by a late-in-life desire to heal all intellectual wounds no matter what.
To begin with, while avoiding technicalities, I should be more explicit about my earlier views—which may in part have been prejudices. Echoing Chomsky, I thought of human linguistic capacity (and performance) as the necessary outcome of a special “mental organ”—one predisposed by biology and neurology to learn language—and particularly its syntactical features—quickly as a result of extensive genetic pre-programming, sometimes nicknamed a “Language Acquisition Device.” (For the most elaborated version of this position, see Lenneberg, 1967.)
From Piaget, I thought of human development as consisting of a number of discrete stages—which cut across the gamut of human cognitive capacities. A child at the level of “concrete operations” would bring essentially the same analytic equipment to bear on a range of contents—from spatial reasoning to numerical computation to moral judgments…and with the transition in adolescence to the stage of “formal operations,” a more sophisticated toolkit would be brought to bear on the aforementioned range of contents (Gardner, 1973).
What’s prompted this re-examination of my views, or indeed, my prejudices?
On the one hand, it’s grown out of my own research-based belief in the plurality of human intelligences. Piaget (and his students and followers) saw various human capacities as essentially “of a piece”—one overarching intelligence. (It’s perhaps worth noting that Piaget himself became interested in human cognition as a result of a stint working in the Parisian laboratory of Théodore Simon, where IQ tests had been initially devised.) But, as a result of my own extensive research, I’ve come to believe that human beings harbor a number of relatively separate intellectual “faculties”—which I’ve dubbed the “multiple intelligences.” One may well be at a concrete level with one faculty (say, numerical), at a formal level with a second faculty (say, spatial), and somewhere in between with a third faculty (say, understanding of other persons).
A far greater influence on my recent thinking has been the emergence of powerful forms of AI. AI does not need to be “prepared” by biological and chemical processes that evolved over thousands of years, nor does it need to draw on pre-existing or prepotent mental faculties. It merely needs to draw on numerous—indeed thousands, or more typically, millions of— examples that are similar in one or more respects. And in an overwhelming number of cases, the AI will arrive essentially at the same kinds of categorizations and reach the same sorts of conclusions as do well-informed and well-educated human beings—in a fraction of the time, and with a “hit rate” that often exceeds that of humans and sometimes approaches 100%. (Though, to be sure, there are occasional hallucinations—sometimes quite bizarre ones.) That’s why the chatbots with which many of us interact regularly seem very sharp—often anticipating our arguments and counterarguments.
The Chomskyan in me reacts: But AI needs millions of examples! Humans reach that level with far fewer examples. But then, my contemporary self replies—To be sure, but the linguistic examples encountered by the young child are heavily contoured and contextualized—if the young child were surrounded and raised strictly by inflexible, pre-programmed automata, linguistic development would take far longer. To borrow the lingo, kids learn language readily because they are guided by “motherese” (simplified and accentuated conversational speech) and highly supportive contexts (see Berko Gleason, 1985).
A young girl using a Skinnerian teaching machine in the early 1960s
My Piagetian superego responds: Sure, there can be some differences across subject matter and disciplines when children in middle childhood reach the age of concrete operations—we have long recognized this finding with our term “horizontal decalage.” But Gardner 2026 replies: Yes, but as a Piagetian, you have been unaware of the nature and scope of different intelligences. Your analysis applies primarily to the realm of logic and mathematics; it has little explanatory power when one considers the understanding of music, of other persons, of the natural world, or of oneself—the foci of other intelligences that I have identified.
So far, despite my literary efforts, my review may seem like “inside baseball.” Chomsky and anti-Chomskyans can debate; Piagetians and anti-Piagetians can dispute; but they still take a fundamentally cognitive and developmental approach to the analysis of human thought.
But what’s really raised my analytic blood pressure has been the educational implications of such theoretical perspectives. When Skinner introduced the idea of teaching machines well over half a century ago (1968), I found them repellant. I did not respond well to the idea of technology which “knew” the right answer and which—through strategically situated rewards and punishments (positive and negative reinforcement)—"formed” or “molded” the student so that she could parrot back that answer and then move to the next carefully-sequenced issue, subject matter, or discipline. It seems like training pigeons or seals—not in the least like the child-centered strands of progressive education for which I have long advocated.
To the extent that AI approaches to education adopt that pre-packaged perspective, I continue to have little sympathy for them…and certainly don’t want them used with my five grandchildren!
A computer screenshot from a conversation with Therabot, a generative A.I. therapist developed by Dartmouth researchers (NYTimes)
But as I now envision them, AI tutors are increasingly coming to resemble the most skilled instructors in well-supported educational settings. Far from simply guiding the student efficiently to “emit’ the correct response, AI tutoring systems can engage the student in dialogue, bring up new questions and fresh perspectives, revisit misconceptions, challenge them deftly, and seek to resolve or dissolve them. Far from just withholding the right answer till its pre-programmed appearance in the “schedule of reinforcement,” a well-designed AI tutorial system co-constructs a fresh perspective…and then moves on to a next possible slant, topic, answer, or question…one which can be stimulated by the student’s specific questions, answers, confusions, or anticipations.
Indeed, that’s the way that I use Claude and other chatbots—and I hope I’ll never to have to abandon them.
A comparison with AI therapists may be appropriate. A therapist in the tradition of a Skinnerian box would be unimpressive—if not frankly insulting (Weizenbaum, 1966). But an AI therapist that engages in conversation, that reflects along with the patient, that presents various alternative patterns of behavior (and streams of thoughts) and helps the patient arrive at the one most suitable for the occasion would—dare I say it—be more effective than most human therapists…though I’d still rather be seen by the psychoanalyst Erik Erikson, who was my undergraduate tutor as well as the teacher of Sherry Turkle (2025)—our most effective critic of AI-conducted psychotherapy.
A Moral
No simple blog of a few thousand words should have any pretense of changing lifelong attitudes and convictions. At the same time, it’s inappropriate (and shoddy scholarship) to adhere to a perspective—be it cognitive, structural, developmental, or clinical—just because it’s familiar and has been regnant (in one’s own mind) for a considerable time. The advent of AI and AGI has stimulated me to reconsider some long-held beliefs—and prejudices—and I’m glad that has occurred. And if, at some point, my “synthesizing mind” can help to integrate the perspectives associated with Chomsky, Piaget, Bruner, and Skinner—a shout-out to all of them!
To conclude, a thoughtful response from Steven Pinker
“With Skinner, I think one has to distinguish three strands of his thought (if you’ll pardon the expression): his behaviorism (the philosophy of psychology that eschewed mentalistic explanation), his associationist learning theory (Pavlovian and operant conditioning), and his scientism (teaching machines, Beyond Freedom and Dignity, and the rest). There was something to his associationism, as we see in LLMs, but (as I point out in a recent essay) LLMs are certainly not stimulus-response machines. Their hidden layers, embeddings, and attention heads are all cognitive, perhaps even concessions to Chomsky’s original points about language.
Likewise, I agree that today’s AI tutors are awe-inspiring, but that’s precisely because their own understanding, and their understanding of the tutee’s understanding, are so deep. Very much unlike the superficial “computer-instructed education” of the 1980s (where the computer just turned the pages), let alone Skinner’s teaching machines, which by design diced the content into tiny passages prompting one-word responses. One can endorse the very idea of teaching machines (and even, pace Sherry, automated therapists) without endorsing Skinner’s rationale for them, namely behaviorist response-reinforcers.
Hard to say whether Joe Weizenbaum had the last laugh or that the last laugh was on him. LLMs, in a sense, grew out of ELIZA (though, as I argue in my essay, with an infusion of Chomkyesque accommodations).
I doubt that motherese closes the gap between the 10-trillion-word input to LLMs and children’s input. Training an LLM with a child-size corpus of motherese would come nowhere near to making it smart. And we have reason to believe that kids don’t need most of the features of motherese—many cultures don’t have them, and many kids learn language from peers.
ACKNOWLEDGMENTS
For their helpful comments on earlier drafts, I thank Kirsten McHugh, Annie Stachura, and Ellen Winner.
REFERENCES
Gleason, J. B. (1985). The development of language. Longman Higher Education.
Bruner, J. S., Goodnow, J. J., & Austin, G. A. (1956). A study of thinking. John Wiley and Sons.
Bruner, J. S., Olver, R. R., & Greenfield, P. M., et al. (1966). Studies in cognitive growth. Wiley.
Bruner, J.S. (1976). Psychology and the Image of Mind. Times Literary Supplement.
Chomsky, N. (1957). Syntactic structures. Mouton & Co in The Hague.
Chomsky, N. (1971). The case against B.F. Skinner. The New York Review of Books.
Gardner, H. (1973). The quest for mind: Piaget, Levi-Strauss, and the structuralist movement. Knopf.
Lenneberg, E.H. (1967). Biological Foundations of Language. Wiley.
Piaget, J. (1970). Piaget’s Theory (G. Gellerier & J. Langer, Trans.). In: P.H. Mussen (Ed.), Carmichael’s Manual of Child Psychology (3rd Edition, Vol. 1). Wiley.
Skinner, B.F. (1968). The technology of teaching. Appleton-Century-Crofts.
Turkle, S. (2025). Reclaiming Conversation in the Age of AI. https://www.afterbabel.com/p/reclaiming-conversation-age-of-ai
Weizenbaum, J. (1966). ELIZA—A computer program for the study of natural language communication between man and machine. Communications of the ACM, 9(1), 36–45.
