Free Film "The History of Applied Behavior Analysis: Part 1" Now Available!
Store My Library About Us Blog Login

What a 1988 Critique of AI Gets Right Today

aba ai literacy artificial intelligence behavior analysis clinical leadership history of psychology n. h. pronko philosophy of science reading the past testing the present Oct 30, 2026

What N. H. Pronko's From AI to Zeitgeist can teach us about psychology, scientific explanation, and the questions we are still asking nearly four decades later.

Reading the Past, Testing the Present | Essay No. 1

In 1988, psychologist Nicholas Henry Pronko published a book that, nearly four decades later, remains remarkably relevant. From AI to Zeitgeist: A Philosophical Guide for the Skeptical Psychologist examines the philosophical assumptions underlying psychology, the theories scientists construct, and the consequences of confusing those theories with the phenomena they are intended to explain.

The first chapter, titled "AI (Artificial Intelligence)," begins with what was then an extraordinary list of technological accomplishments. Computers could play chess, process financial transactions, perform medical analyses, operate industrial equipment, and recognize patterns in biological information. Pronko acknowledged these achievements while expressing considerable skepticism about the theoretical interpretations being attached to them.

His concern extended beyond whether computers could perform increasingly sophisticated tasks. He questioned whether their performance justified a particular understanding of human behavior. More specifically, he challenged the increasingly popular assumption that understanding computers would tell us how the human brain operates, and that understanding the brain as a computer would eventually explain psychological events.

In the decades since, artificial intelligence has developed capabilities that Pronko and his contemporaries could scarcely have anticipated. Still, many of the underlying questions remain. In some respects, our technological accomplishments have made them more urgent.

I'm beginning a series of essays examining this book, chapter by chapter, because I believe it offers something our profession increasingly needs: an opportunity to reconsider familiar problems without assuming they have been resolved simply because the terminology, technology, or prevailing scientific consensus has changed.

Behavior analysis has a rich intellectual history, but I don't think we always make sufficient use of it. Important debates about explanation, causation, learning, measurement, and the relationship between organisms and their environments have unfolded repeatedly throughout psychology's development. Sometimes those debates have produced meaningful advancements. Other times, we have exchanged one vocabulary for another without substantially changing the assumptions underneath.

I want to examine what Pronko argued, what subsequent scientific developments support or challenge, and what those differences mean for behavior analysts today. Some of his arguments will hold up, and others will inevitably require us to reconsider what he believed to be true. Revisiting older scholarship gives us an opportunity to recognize that previous generations of scientists were frequently asking questions we have yet to finish answering.

A book about psychology's assumptions, not just artificial intelligence

Before examining Pronko's arguments, it is worth explaining why I think From AI to Zeitgeist deserves a place on the reading list of anyone interested in the theoretical foundations of psychology.

Published by Greenwood Press in 1988, the book is a sustained examination of psychology's philosophical commitments. Across its pages, Pronko challenges readers to consider how their assumptions about the world influence the kinds of questions they ask, the methods they select, and the conclusions they consider legitimate.

Scientific methods do not operate independently of philosophical assumptions. Decisions about what constitutes an appropriate unit of analysis, what qualifies as evidence, and what it means to explain an event are influenced by assumptions that may rarely be made explicit. Behavior analysts encounter these issues whenever we debate the distinction between description and explanation, the legitimacy of hypothetical constructs, or whether a functional relation is sufficient to account for behavior.

The same concerns arise when researchers use computational models to explain cognition, when clinicians interpret assessment results, and when technology companies claim that an algorithm can identify behavioral functions or recommend effective interventions. Pronko's book encourages us to examine those claims more thoroughly, including the assumptions that influence how we interpret a procedure's usefulness.

A model can be useful without being a complete or accurate representation of the phenomenon it models. Likewise, a technology can demonstrate impressive performance without necessarily validating the theoretical framework used to describe that performance. That tension runs throughout the opening chapter, and I think it provides a useful starting point for this series.

When a metaphor becomes an explanation

Pronko identifies a reciprocal assumption that was becoming increasingly influential in both psychology and artificial intelligence: "the brain is a computer and the computer is a brain" (p. 1). He presents this formulation as a prevailing doctrine, one he believes deserves considerable scrutiny.

The problem is relatively straightforward. If we begin by assuming that the brain operates like a computer, then construct computers based on our understanding of brain processes, we risk creating a circular system of explanation. The computer becomes evidence for our theory of the brain, while the brain becomes the justification for how we designed the computer. Neither necessarily establishes that the analogy accurately describes the processes responsible for psychological events.

Consider a familiar example from contemporary behavior analysis. A clinician observes a learner successfully completing a task, then proposes that the learner possesses a particular cognitive capacity responsible for that performance. If the capacity is inferred exclusively from the observed performance, using it to explain the same performance adds little without independent evidence or additional predictive value. We have described the observation differently, but have we explained it?

The computational analogy can create a similar problem when researchers move from observing that computers and humans perform comparable tasks to concluding that both must operate through equivalent processes. Computational models, of course, can contribute substantially to scientific understanding. Computational approaches have generated valuable predictions, supported experimental investigations, and provided increasingly sophisticated tools for studying biological and behavioral processes. What matters is the purpose of the model and the conclusions we believe its performance justifies.

A computational model might predict the probability of a particular response under specified conditions. It might provide a mathematical description of changes across repeated learning opportunities. It might even identify relationships that would be difficult to detect using conventional statistical methods. Those accomplishments are potentially valuable. However, accurately predicting an outcome does not, by itself, establish that the model has identified the causal processes responsible for it. The implications become clearer when considering artificial intelligence in applied behavior analysis. A machine learning system might predict that a client is more likely to engage in a particular response during certain instructional activities. It may identify correlations involving time of day, staffing, environmental conditions, or previous behavior. Those predictions could help a clinician determine where additional assessment would be useful. But prediction is not the same as experimental demonstration of a functional relation.

A system that successfully anticipates behavior has not necessarily established why that behavior occurs. Nor has it demonstrated that changing the variables contributing to its predictions will produce the desired clinical outcomes. For behavior analysts, this should be familiar. It is one of the reasons our field has historically emphasized experimental analysis rather than relying exclusively on descriptive observation. Pronko's argument serves as a useful reminder that technological sophistication does not eliminate the need to distinguish among description, prediction, and explanation.

The nervous system is necessary, but is it sufficient?

One of the chapter's most important discussions concerns the nervous system's role in psychological events.

Pronko describes an intact nervous system as "a necessary but not sufficient condition" for the psychological functioning of an organism (p. 2). I think this statement deserves careful consideration because it reflects a position that remains central to many contextual and behavior analytic accounts. The nervous system is unquestionably important. Neurological functioning enables and constrains countless aspects of behavior. Injury, disease, pharmacological manipulation, and biological development can dramatically alter an organism's behavioral repertoire.

However, identifying the biological structures involved in behavior does not automatically explain how particular responses were established, maintained, or altered through an organism's interactions with its environment. A neuroscientist might identify patterns of neural activity associated with reading. Those findings could reveal important information about the biological processes participating in reading behavior. But they would not independently explain why one individual learned to read fluently, another developed persistent difficulties, and a third acquired different reading repertoires under different instructional conditions. Those explanations may require consideration of instructional histories, stimulus control, reinforcement contingencies, prerequisite repertoires, and opportunities for practice. The biological and behavioral levels of analysis need not compete. They address different, sometimes overlapping, questions.

Pronko's broader philosophical orientation becomes particularly interesting here. He emphasizes the organism's history and the circumstances in which behavior occurs, rather than treating psychological events as products of an isolated internal mechanism. From this perspective, understanding behavior requires analyzing the organism's interactions with its environment across time, including the biological conditions that participate in those interactions. Importantly, the history of an organism may be essential to identifying the behavioral phenomenon we are attempting to explain in the first place. Simply incorporating history as another variable within a neurological model may leave important questions unanswered.

Modern neuroscience has advanced enormously since Pronko wrote these passages. Researchers can investigate neural activity, connectivity, plasticity, and the biological effects of learning with methods unavailable in 1988. Computational neuroscience has likewise become an increasingly productive scientific enterprise. These developments complicate any sharp separation between neurological and behavioral explanation. Learning histories produce biological changes, and biological conditions influence the effects of environmental events.

Yet they do not make an organism's history irrelevant.

If anything, they provide additional opportunities to investigate how biological and environmental processes interact. The contemporary challenge involves determining which questions can be answered at different levels of analysis, how those explanations relate, and whether our methods provide evidence appropriate to our conclusions.

Where Pronko underestimated artificial intelligence

While I find Pronko's philosophical skepticism valuable, a serious examination of the chapter should also acknowledge where subsequent scientific developments challenge his expectations. Pronko describes the accomplishments of artificial intelligence as products of human creativity and engineering, while rejecting grander predictions about machines surpassing human intellectual capabilities.

He writes, "Whatever they do is the result of human creativity, design, and manufacture" (p. 3).

The statement remains defensible as an observation about the human origins of artificial technologies. But it does not establish an upper limit on what those technologies can accomplish. After all, the fact that human beings designed an artificial system does not mean its performance must remain inferior to human performance across every domain. Nor does it mean that all the system's specific outputs were directly anticipated or individually programmed by its creators. By 2026, we have substantial evidence that artificial intelligence can achieve levels of performance that would have been extraordinary, and in some cases implausible, in 1988.

The 2026 Stanford AI Index documents systems meeting or exceeding established human baselines across several demanding academic and technical benchmarks. Advances in language modeling, computer vision, mathematical problem solving, and software engineering have dramatically expanded the range of tasks machines can perform. Research applications are equally significant. AlphaFold 3, described in a 2024 Nature publication, demonstrated substantial advances in predicting the structures of interacting biological molecules, including proteins, nucleic acids, and smaller molecular compounds. These accomplishments deserve serious recognition, regardless of whether the systems responsible are artificial.

Pronko was correct to question whether sophisticated machine performance establishes equivalence with human psychological processes. However, any broader assumption that machines could not substantially exceed human performance in complicated intellectual tasks has become increasingly difficult to defend. I think we have to evaluate those claims separately. We can recognize extraordinary technological accomplishments without concluding that machines possess human psychological histories, biological characteristics, or subjective experiences. Likewise, we should remain open to the possibility that systems without those characteristics can develop capabilities that equal or exceed human performance in particular domains. A philosophical position becomes scientifically vulnerable when it is used to impose limits on empirical possibilities without sufficient evidence. Behavior analysts should be particularly sensitive to that problem. Our philosophical commitments should guide inquiry while remaining responsive to evidence that might contradict our expectations.

The uneven development of artificial intelligence

Modern AI also illustrates why the concept of intelligence itself deserves scrutiny. The 2026 Stanford AI Index describes a strikingly uneven pattern of technological competence. Systems capable of extraordinary mathematical or scientific performance can still struggle with apparently simple perceptual tasks. AI agents have made substantial progress completing real computer-based activities, yet failures remain frequent. Robotic systems that perform well in controlled environments often struggle when introduced to the variability of ordinary household settings. These difficulties raise questions about how we define, measure, and interpret intelligence.

If a system outperforms highly trained professionals on one task but fails a comparatively simple task under different conditions, what does it mean to describe that system as intelligent? The answer depends partly on the construct being measured. A general label such as intelligence can conceal substantial variation in the repertoires, conditions, and performance criteria involved. Behavior analysis offers a useful perspective here because our explanations have traditionally emphasized relations among behavior, environmental circumstances, and learning histories rather than relying exclusively on global descriptions of ability. A learner's successful performance under one arrangement of stimuli does not automatically demonstrate generalization to different materials, instructions, people, or settings.

Similarly, an AI model's performance on a benchmark does not necessarily establish equivalent performance in unfamiliar environments. This becomes especially relevant when comparing the behavior of artificial systems with human behavior. A language model might produce a convincing explanation of a complicated social interaction. That response does not establish that the model acquired the relevant repertoire through the same processes as a person who experienced comparable interactions. The observable products may resemble one another while the histories and mechanisms responsible differ substantially. The machine's performance can still be practically valuable, even if the similarities between the resulting products tell us relatively little about the processes that produced them. One lesson from modern AI research is that highly capable systems can also be remarkably inconsistent. Another is that our methods for evaluating them are frequently overtaken by improvements in the systems themselves. Both lessons should encourage methodological humility.

Computers can produce answers. Can they produce explanations?

One of Pronko's concerns was the tendency to mistake computational output for psychological understanding. Today, we encounter this issue in a different form.

Large Language Models can generate coherent explanations of psychological theories, describe behavioral principles, recommend interventions, and produce extensive clinical documentation. Their outputs can be sufficiently fluent that readers have difficulty identifying when an explanation is incomplete, inaccurate, or unsupported. The concerns extend beyond occasional factual errors. Generative systems can construct responses that appear logically developed while incorporating false premises, unreliable citations, or unjustified conclusions. These risks are documented in the National Institute of Standards and Technology's Generative Artificial Intelligence Profile and remain an active concern in contemporary AI evaluation. Of course, these limitations do not make every explanation generated by an AI system worthless. A well-supported explanation remains useful regardless of who or what produced it. An explanation should be evaluated according to its content, supporting evidence, assumptions, and the predictions it permits. The challenge, then, concerns how we evaluate these explanations.

Suppose a behavior analyst asks an AI system to explain why a student's disruptive behavior increases during mathematics instruction. The system might produce an interpretation involving task difficulty, escape-maintained responding, insufficient prerequisite skills, or the history of reinforcement associated with the instructional context. Any of those possibilities might warrant investigation. But without sufficiently detailed observations, appropriate assessment, and relevant experimental evidence, the explanation remains a hypothesis. Its professional vocabulary does not change its evidentiary status. What distinguishes a scientifically useful account is the extent to which it identifies relevant variables, generates testable predictions, and survives contact with evidence. A plausible explanation may be a reasonable starting point, but we should be careful about how much confidence we place in it.

I believe behavior analysts have an important contribution to make to contemporary discussions of artificial intelligence for precisely this reason. Our field has spent decades examining the conditions under which observed relations justify particular interpretations. We have also developed methods for evaluating whether changes in environmental variables produce changes in behavior. Those commitments are directly relevant to the evaluation of AI-generated claims. We should apply the same skepticism to human explanations, of course. A credentialed practitioner can produce an unsupported interpretation just as readily as an artificial system. The standard should remain the quality of the evidence.

What this means for behavior analysts and ABA organizations

For practitioners and ABA company owners, these philosophical questions have immediate operational implications. Artificial intelligence is already being incorporated into scheduling, documentation, clinical data systems, staff training, administrative workflows, and decision-support applications. Some uses may improve efficiency and reduce unnecessary administrative work. Others could introduce errors, inappropriate recommendations, privacy risks, or overreliance on systems whose limitations are not adequately understood. As these technologies become more common, we need to establish whether particular applications can demonstrate meaningful value under conditions appropriate to their intended use.

Consider an AI system designed to assist clinicians with session documentation. Its performance might be evaluated using several dimensions: factual accuracy, completeness, consistency with recorded events, time required for clinician review, and the frequency of corrections before submission. If the system consistently reduces documentation time without compromising accuracy, that would provide evidence supporting its usefulness for the assessed workflow. But it would not establish that the same system can independently develop treatment plans, identify behavioral functions, or make clinically appropriate decisions. Each proposed application introduces a different set of performance requirements. A tool that summarizes existing clinical information is different from a tool that generates new clinical recommendations. The latter requires additional evidence concerning the validity and safety of its conclusions.

Likewise, an algorithm that identifies statistical patterns in behavior data is not necessarily capable of conducting a functional assessment. It may generate useful hypotheses. Whether those hypotheses identify functional relations is a separate empirical question. These considerations become especially important when software companies describe their products using terms such as intelligent, predictive, personalized, or clinically informed. These descriptions may be reasonable, but they should not substitute for clearly defined performance criteria.

For organizations purchasing or developing AI tools, the practical questions become increasingly specific. What outcome is the system intended to improve? What evidence supports its performance? Under what conditions was that evidence collected? What types of errors occur, and how frequently? What happens when the system encounters circumstances that differ from its evaluation environment? It is also necessary to evaluate whether the system performs comparably across relevant populations, contexts, and types of clinical work.

These questions should be familiar to behavior analysts. They extend the same scientific reasoning we already apply when evaluating assessments, interventions, and other technologies. Professional guidance has begun to address these issues. The Behavior Analyst Certification Board discussed considerations for AI in ABA practice in its July 2024 newsletter, highlighting risks involving privacy, inaccurate information, and professional responsibilities. The Council of Autism Service Providers subsequently developed Practice Parameters for Artificial Intelligence Use in Applied Behavior Analysis, addressing organizational oversight, selection, deployment, monitoring, and auditing. These developments are encouraging, although professional guidance will need to evolve alongside the technology and the evidence supporting its applications. There are also legal and technical obligations that cannot be reduced to professional judgment alone. When vendors create, receive, maintain, or transmit protected health information on behalf of HIPAA-regulated organizations, applicable business associate requirements and safeguards must be addressed. The U.S. Department of Health and Human Services provides guidance on these obligations.

For organizational leaders, responsible adoption therefore requires more than purchasing a tool and informing employees that it is available. It requires the development of procedures for evaluating the tool, preparing staff, defining appropriate uses, protecting sensitive information, monitoring performance, and responding when errors occur. The fact that a technology is capable of doing something does not establish that it should be used for that purpose.

From artificial intelligence to philosophy of science

There is another implication of Pronko's chapter that I think deserves more attention than the technological debate itself. The history of psychology demonstrates that our preferred explanatory language changes over time. At different points, investigators have described psychological events using the language of association, reflexes, drives, information processing, cognitive representations, neural networks, and increasingly sophisticated computational processes. Some of these developments have produced substantial scientific progress. Others have introduced new descriptive systems without sufficiently resolving the explanatory problems that preceded them. Scientific terminology can create the appearance of precision even when the relationships under consideration remain poorly understood. Pronko asks readers to examine the assumptions that make particular explanations seem reasonable. I think that remains a valuable exercise.

Consider how frequently behavior analysts debate terms such as function, motivation, cognition, reinforcement, and learning. These discussions sometimes reveal genuine disagreements about observable processes. At other times, they expose differences in philosophical orientation or the levels of analysis being employed. Those differences cannot always be resolved through additional data alone because participants may disagree about what the data are supposed to explain. This is one of the reasons philosophy of science remains important. It helps us distinguish an empirical disagreement from a conceptual one. It asks whether a construct identifies a phenomenon requiring investigation or merely renames the phenomenon already observed. It also encourages us to examine whether our theoretical commitments unnecessarily restrict the kinds of evidence we are willing to consider.

For behavior analysts, these considerations should encourage thoughtful engagement with neuroscience, cognitive science, and artificial intelligence, even when their conceptual frameworks differ from our own. Scientific progress requires the possibility that other approaches will discover relationships our preferred methods have not adequately addressed. At the same time, methodological sophistication and technological novelty do not automatically resolve conceptual problems. An increasingly accurate mathematical model is an accomplishment. Determining what that model explains, and what it leaves unexplained, remains a separate responsibility. I think this captures much of what makes Pronko worth reading today.

What has changed, and what remains unresolved?

Nearly forty years after publication, I would characterize the opening chapter's legacy in three ways.

  1. Pronko's skepticism toward treating computational analogies as established psychological explanations remains valuable. A machine's ability to perform tasks associated with human intelligence does not automatically demonstrate that its operations are equivalent to the biological, behavioral, or experiential processes involved in human performance.
  2. His expectations concerning the limits of artificial systems require substantial revision. Modern AI has demonstrated capabilities beyond those available in 1988, and future progress should be evaluated empirically rather than restricted by assumptions about what machines must be incapable of achieving.
  3. His insistence that psychology consider organisms in relation to their histories and environments remains an important corrective to explanations that locate psychological events exclusively inside the brain.

I think these conclusions leave plenty of room for both technological optimism and scientific skepticism. We can recognize that artificial intelligence has become a powerful scientific and practical tool while continuing to question the assumptions used to interpret its accomplishments. We can also acknowledge that contemporary biological and computational research contributes meaningfully to psychological science while recognizing the continued importance of behavioral levels of analysis. The more productive question concerns how different forms of inquiry can contribute to a more comprehensive and empirically accountable understanding of behavior. It is a question psychology was wrestling with in 1988, and one we are still working through in 2026.

Why I am revisiting this book, one chapter at a time

This essay begins what I intend to be a sustained examination of From AI to Zeitgeist. Each installment will consider a chapter's major arguments, place them within their historical and philosophical context, examine relevant scientific developments since publication, and consider their implications for contemporary psychology and behavior analysis. Some of Pronko's arguments will hold up remarkably well. Others will require qualifications or substantial revision. Still others may prove most valuable because they reveal how our contemporary assumptions differ from those that influenced earlier generations of psychologists.

I don't think we should approach this work searching for confirmation that behavior analysis has always been correct. Likewise, the fact that a scientific idea is newer hardly establishes that it is an improvement over what came before. A discipline develops through the accumulation of evidence, conceptual clarification, methodological innovation, and the revision of conclusions that no longer withstand scrutiny. That process depends partly on remembering what questions previous generations asked and understanding why they considered those questions important. It also requires distinguishing scientific progress from changes in intellectual fashion.

This is especially relevant in a period when artificial intelligence is influencing how research is conducted, how clinical work is documented, how professional education is delivered, and how information is produced and evaluated.

For readers who are particularly interested in these questions, Project Chiron is another place where Dr. David J. Cox and I are examining the practical, scientific, and ethical dimensions of artificial intelligence for behavior analytic professionals. That work is more directly focused on AI literacy and responsible applications, while this series will take a broader historical and philosophical perspective. The two share an underlying commitment: understanding the assumptions, evidence, and consequences of the tools and explanations we use.

Returning to the questions that matter

What I appreciate most about Pronko's opening chapter is that it gives us an opportunity to consider what scientific advancement actually means. The computers he described in 1988 and the artificial intelligence systems we use today are separated by extraordinary developments in computing, mathematics, engineering, and scientific research. Those developments have not eliminated the need to ask what constitutes an explanation, how we distinguish models from the phenomena they represent, or what evidence is required to justify claims about psychological processes. If anything, they have made those questions more consequential.

For behavior analysts, I think the challenge is to remain appropriately skeptical while being open to advancements outside our traditional areas of research. We should be enthusiastic about scientific progress while continuing to examine the conceptual commitments that guide our investigations. That requires understanding what a technology actually demonstrates, what remains uncertain, and how its use changes the conditions under which people behave. I hope these essays encourage readers to seek out Pronko's book and engage directly with his arguments. The brief passages considered here cannot substitute for the intellectual context provided by the complete work. There is something especially valuable about encountering a question as it was formulated decades before our contemporary debates. It gives us an opportunity to determine whether we have genuinely advanced our understanding or simply developed more sophisticated ways of discussing familiar problems. Pronko offers plenty to disagree with, but even more to think about. 

And that's precisely what makes this book worth revisiting.

References and further reading

 

Close

50% Complete

Two Step

Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.