The Sciences of the Artificial

Before the program ran, people became its subroutines

One evening in early January 1956, Herbert Simon brought his family into a Carnegie Tech classroom. He, Allen Newell, and Cliff Shaw had been developing what became the Logic Theorist, a program for searching out proofs in symbolic logic. Simon's wife, three children, Newell, and a graduate student each received an instruction card and enacted a part of the program. One child stood when a “not p” operation was needed; another served as theorem memory; another recorded branches on the blackboard; Simon's wife played the executive program. Kathie Simon Frank's family recollection preserves the scene from a participant's perspective.¹

The demonstration made an abstract claim tangible: a difficult intellectual task could be represented as symbols, rules, memories, and selective search. Newell and Simon's 1956 RAND report specified the Logic Theory Machine in a pseudocode suitable for eventual computer implementation. The paper explicitly set realization aside. After Cliff Shaw programmed the system, the team reported empirical runs in 1957. The hand simulation, formal specification, and later program showed that heuristic search could find proofs in a narrow logical domain. That achievement did not establish general machine intelligence, but it supplied one of the concrete research settings from which Simon's larger science of design emerged. ²

The Sciences of the Artificial was not written as a report on that program. Its 1969 first edition grew from Simon's 1968 Karl Taylor Compton lectures at MIT. The 1981 second edition interleaved those chapters with his 1980 H. Rowan Gaither lectures at Berkeley; the 1996 third edition expanded them again. “The Architecture of Complexity,” first published in 1962, became a separate chapter. Simon credits Newell, students, and colleagues throughout. He also records that some design research behind the book received ARPA support alongside health and private-foundation funding. The result is a layered argument assembled across computing, psychology, economics, engineering education, organization theory, and public planning—not the product of one canonical design episode.³

Artificial things can be studied because purpose meets environment

Simon calls a phenomenon artificial when goals or purposes mold a system to the environment in which it lives. A clock, computer program, organization, or city is contingent on what someone wants it to do, yet its fit with an environment can still be investigated. An artifact's behavior often depends on the relation between an inner environment—its organization—and an outer environment that supplies conditions, tasks, and disturbances.⁴

The Logic Theorist illustrates the composition. Its theorem memory and branching rules mattered because they met a particular logical problem. Simon's ant on a beach makes the inverse point: a complicated path may reflect the uneven beach more than a complicated ant. Both are authored examples used to redirect explanation. Neither establishes that inner structure is unimportant or that social environments can be reduced to task constraints.⁵

By the expanded editions, the book's stronger artificial-intelligence claim is the physical symbol system hypothesis: a physical symbol system has the necessary and sufficient means for general intelligent action. Simon labels it an empirical hypothesis and offers evidence from functioning programs and experiments on human problem solving. A theorem prover's performance is evidence that symbolic search can do some intelligent work; it is not by itself proof of necessity, sufficiency, or generality. The reach from a formal proof space to embodied life remained contested.⁶

Design is bounded search toward a preferred situation

Simon defines a designer broadly: anyone who devises “courses of action aimed at changing existing situations into preferred ones.” The formulation connects the bounded rationality of Administrative Behavior to engineering, medicine, management, and public policy. Because a designer cannot enumerate every alternative and consequence, design depends on representations, heuristics, selective search, and satisfactory solutions.⁷

The phrase “preferred ones” puts value before technique. A problem representation selects what counts as a state, a difference, an available move, and a result. It can make a vast search tractable; it can also exclude a person, purpose, or consequence before evaluation begins. Simon's proposed curriculum for a science of design teaches formal optimization, search, evaluation, and the representation of complex problems. Those methods can compare means. They do not authorize the preferred state or determine whose preference should govern. ⁸

Lucy Suchman's research on human-machine interaction later challenged plan-based accounts of action. From situated observations of people using machines, she argued that plans are resources produced and interpreted in ongoing activity, not complete specifications that cause action from outside it. Rodney Brooks's mobile-robot research offered a different technical challenge: robots coupled directly to a changing world could acquire useful behavior without a central explicit model doing all the work. These critiques do not erase Simon's account of bounded search. They dispute a smooth progression from symbolic program to a general science of intelligent, situated action.⁹

Hierarchy works only when the interactions are actually weak

In Simon's parable, two watchmakers each assemble a thousand-part watch. Tempus's unfinished watch falls entirely apart whenever the telephone rings. Hora builds stable subassemblies, so an interruption destroys only the most recent piece of work. The quantitative thought experiment explains why nested stable systems can be built and changed faster. It is a parable and model, not an observed history of watchmaking.¹⁰

Near-decomposability is the conditional claim beneath it. When interactions inside a subsystem are stronger or faster than interactions between subsystems, the parts can be analyzed somewhat independently in the short run while their aggregate relations matter over longer periods. Hierarchy here means nested subsystems, not necessarily a chain of human command. Simon explicitly requires the condition to hold; modularity is not a license to assume that it does. ¹¹

The Boeing 737 MAX offers a later comparison of what happens when a system's boundaries conceal powerful cross-effects. The U.S. House investigation attributes Boeing's choices to program pressures and governance failures rather than to Simon's framework. It documents an interacting failure: production pressure, the objective of preserving pilot-training commonality, assumptions about pilot response, delegated certification, omitted information, and MCAS's dependence on a single angle-of-attack sensor crossed the nominal boundaries between software, aircraft, crew, regulator, and airline. Two crashes killed 346 people.¹²

Read through Simon's terms, the case asks an empirical question before a modular one: were the interactions weak enough to treat those components separately? They were not. A local objective—avoid costly simulator training—affected system design and what pilots were told. A sensor failure propagated through automation into a human cockpit under severe time pressure. The official record, not the watchmaker story, establishes those facts; near-decomposability supplies a later language for seeing why local compliance could not guarantee system safety. The Boeing history follows the organizational choices in more detail.¹³

The boundary of the artifact is part of the design

Simon's framework remains powerful because it turns creation into something that can be examined: name the purpose, model the environment, represent alternatives, search within limits, and study the coupling among parts. It also reveals the questions technique cannot answer for itself. Who chose the preferred situation? Who can reopen the representation when a worker, user, community, or ecosystem falls outside it? Which slow or apparently weak relationship carries an irreversible effect?¹⁴

The classroom simulation composed a program from people, cards, chalk, and a formal problem. The 737 MAX composed a safety-critical system from airframe, software, sensors, pilots, training, firms, and regulators. Between them lies the book's most durable lesson: an artifact never ends at the line its designer first draws. Its purposes and consequences are produced at the interface with an environment whose affected beings may understand the system differently—and may bear the cost when the chosen decomposition is wrong.

Structured reading paths and evidence limits

The Herbert A. Simon path locates the work within Simon's research program. Administrative Behavior is both the formal author-progression dependency and a conceptual predecessor: bounded administrative choice becomes a broader science of design and complex systems. Decision making, judgment, and bounded rationality connects that inheritance to later research on search and judgment.

The paths to knowledge, expertise, and professional autonomy and work design, productivity, and automation expose whose judgment is formalized, displaced, or left able to contest a representation. Boeing provides a later institutional case where technical, organizational, regulatory, and human couplings crossed nominal module boundaries; it is an editorial application rather than a test Simon designed. Organizational intelligence connects representation and decomposition to institutional sensing and revision, while benefit for all life asks who may define the preferred state and which human or ecological consequences remain outside it.

No structured impacts or typed relations are asserted for this work. The one formal dependency and eight related paths establish a reading order and analytical comparisons, not shared empirical methods. The evidence package combines Simon's primary text and edition records, participant family memory, primary computing reports, later human-machine and robotics critiques, and official 737 MAX investigation and regulatory records. It does not include representative accounts from people whose work was automated or modularized, 737 MAX passengers or bereaved families, affected communities, or nonhuman and ecological subjects; those absences constrain claims about whose preferred state the design serves.¹⁴