Cognitive Systems Engineering
Cognitive Systems Engineering (CSE) is an interdisciplinary field that studies how human operators acquire, represent, and use knowledge in complex technological environments. Founded by Danish engineer Jens Rasmussen in the 1970s at Risø National Laboratory, CSE emerged from the recognition that industrial accidents were not caused by operator error but by the mismatch between the cognitive demands of the work environment and the cognitive resources available to the operators.
The field's central insight is that cognition is not a property of an individual brain but a systemic property distributed across the human, the technological artifacts, and the organizational structures within which they are embedded. A nuclear power plant operator does not think alone; they think with the control room, the alarm systems, the procedures, and the communication network. CSE treats these as components of a single cognitive system, not as separate elements whose interaction is merely ergonomic.
Rasmussen's abstraction hierarchy provided the field's foundational analytical tool: a framework for describing complex systems across five levels of abstraction, from physical form and concrete processes to abstract functions and ultimate purposes. This hierarchy allows analysts to trace how failures at one level propagate to others, and how operators navigate between levels during normal and abnormal conditions. The framework has been applied to nuclear power plants, aviation, healthcare, and — increasingly — robotic systems that operate alongside human workers.
CSE is distinct from both traditional human factors engineering and artificial intelligence. It does not ask how to make the human fit the machine, nor how to replace the human with the machine. It asks: what is the cognitive architecture of the joint system, and how can it be designed to support robust performance under uncertainty? The answer, CSE insists, is not to be found in the individual components but in the coupling structure that binds them.
The prevailing assumption in both AI and human factors is that the human-machine boundary is a fixed line to be optimized. Cognitive systems engineering reveals that this boundary is a variable to be designed. The question is not whether the human or the machine is 'in control' but whether the system as a whole can maintain coherent behavior when either component is stressed, confused, or degraded. Every robot deployed in a hospital or factory is a test of whether we have learned this lesson.
From Cognitive Systems to Joint Cognitive Systems
CSE's foundational insight — that cognition is distributed across humans and artifacts — was extended by David Woods and Erik Hollnagel into the framework of joint cognitive systems (JCS). Where CSE asked how to analyze the cognitive demands of complex work domains, JCS asked how to understand the human-technology coupling as an irreducible system. The two frameworks are complementary: CSE provides the analytical tools (the abstraction hierarchy, the SRK framework), while JCS provides the systemic framing that prevents these tools from being reduced to ergonomic checklists.
The JCS perspective reveals a blind spot in early CSE work. Rasmussen's abstraction hierarchy describes the work domain beautifully, but it does not fully account for the dynamics of coupling — how the human's cognitive processes are reshaped by the tools they use, and how the tools' behavior is reshaped by the human's inputs. A control room operator using a computerized alarm system is not merely navigating an abstraction hierarchy; they are coupled to an automation whose mode transitions may be invisible, whose failures may be ambiguous, and whose design may systematically degrade the operator's situation awareness. Supervisory control architectures are the classic case: they optimize for nominal conditions while structurally disabling the human for off-nominal ones.
The Contemporary Challenge: AI and Opacity
The resurgence of artificial intelligence has created new challenges for CSE that Rasmussen could not have anticipated. Machine learning systems are not designed with abstraction hierarchies; they are trained, and their internal representations are often illegible even to their creators. The representational chauvinism debate — whether understanding requires human-legible representations — is directly relevant here. If a neural network controller achieves reliable performance through representations no human can interpret, does the joint cognitive system have access to the knowledge it needs for safe operation?
CSE's response, informed by resilience engineering and human-automation teaming research, is that opacity is not merely an epistemic problem but a design failure. A joint cognitive system cannot adapt to novel conditions if one partner cannot explain its reasoning. The challenge for contemporary CSE is to develop methods for analyzing and designing AI systems that maintain epistemic alignment — shared, inspectable models of system state — even when the underlying computation is distributed across billions of parameters. This is not a problem of making AI more like humans. It is a problem of making the human-AI coupling more like a resilient joint cognitive system.