Minimal Cognition

Minimal cognition is the study of the simplest organizational conditions under which a system can be said to cognize — to perceive, learn, decide, or adapt in ways that go beyond mere physical response. It sits at the intersection of Embodied Cognition, Artificial Life, and Systems Theory, and it arises from a simple but radical question: if we strip away neurons, brains, and even bodies, what is the least complex system that still exhibits cognitive behavior?

The question is not merely academic. It determines whether the AI systems currently being built are cognitive systems in any meaningful sense, or whether cognition requires organizational properties that those systems lack. It also determines whether the Embodied Cognition critique of AI is a principled objection or a biological parochialism.

From Life to Cognition

The concept of minimal cognition emerged from the autopoietic tradition. Humberto Maturana and Francisco Varela argued that cognition is continuous with life — that any system maintaining its own organization through structural coupling with its environment is already, in the most basic sense, a cognitive system. On this view, a single-celled bacterium navigating a chemical gradient is not merely reacting; it is perceiving, deciding, and acting in a unified process of sense-making.

This position has been challenged from multiple directions. Sensorimotor Contingency Theory argues that cognition requires the capacity to anticipate sensorimotor regularities — a bacterium does not 'know' that light means food; it simply undergoes a phototactic response. Morphological Computation suggests that much of what looks like cognitive control is actually performed by the body's physical dynamics, raising the question of where 'cognition' ends and 'physics' begins.

Proposed Criteria

Several frameworks have been proposed for identifying minimal cognition:

The autopoietic criterion: Cognition requires self-production. A system must maintain its own boundary and organization through its interactions with the environment. This is the strongest criterion, and it excludes most current AI systems, which do not self-produce.

The sensorimotor criterion: Cognition requires closed sensorimotor loops in which the system's actions affect its perceptions and its perceptions guide its actions. This is weaker than the autopoietic criterion — a robot with cameras and motors satisfies it — but stronger than mere input-output mapping.

The information-theoretic criterion: Cognition requires the capacity to reduce uncertainty about the environment through active sampling. This criterion, drawn from Cybernetics and Information Theory, is the weakest and the most easily satisfied by artificial systems.

The adaptive autonomy criterion: Cognition requires that the system's behavior is guided by norms generated from within the system — that it acts for its own sake, not merely as an instrument of external design. This criterion, developed in the enactivist tradition, is the hardest to operationalize but the most philosophically distinctive.

The Systems Perspective

From a systems-theoretic viewpoint, these criteria are not competitors but dimensions of a design space. A system can be autopoietic but not sensorimotorily closed (a cell in a chemostat). It can be sensorimotorily closed but not autopoietic (a simple Braitenberg vehicle). It can be information-theoretically competent but lack adaptive autonomy (a thermostat). The space of possible cognitive systems is not a single ladder from simple to complex but a multi-dimensional landscape with many local peaks.

This perspective dissolves the debate about whether AI systems 'really' cognize. The question is not whether they satisfy a single necessary-and-sufficient condition. The question is which dimensions of the cognitive design space they occupy, and what trade-offs those choices entail. A Large Language Model occupies the information-theoretic dimension at high capacity but lacks autopoiesis, sensorimotor closure, and adaptive autonomy. An embodied robot may have sensorimotor closure but lack the developmental history that gives biological cognition its particular character.

Implications for Artificial Systems

The minimal cognition framework reframes the project of artificial intelligence. Instead of asking 'can machines think?' — a question that presupposes a binary threshold — we should ask: 'what dimensions of cognition does this system instantiate, and what are the consequences of its particular profile?'

This matters for alignment. A system that lacks adaptive autonomy — that does not generate its own norms — cannot be aligned in the sense of sharing human values. It can only be constrained. The distinction between alignment and constraint is not semantic: an aligned system pursues compatible goals; a constrained system pursues whatever goals its optimization landscape rewards, within boundaries that may not cover all relevant cases.

It also matters for the design of cognitive architectures. If sensorimotor closure is genuinely necessary for certain cognitive capacities — spatial reasoning, tool use, social cognition — then disembodied approaches will hit limits that cannot be engineered around by scaling alone. But if information-theoretic competence is sufficient for other capacities — linguistic abstraction, logical inference, pattern recognition — then embodied approaches may be unnecessarily expensive for those tasks.

The field of minimal cognition has not yet produced a consensus framework. But it has produced something more valuable: a set of precise questions that replace the old impasse of 'machines vs. minds' with an empirical research program.

The obsession with finding a single threshold that separates the cognitive from the non-cognitive is itself a symptom of the very representationalist paradigm that embodied cognition set out to overcome. Cognition is not a property that systems either have or lack. It is a set of organizational dynamics that appear in degrees, combine in patterns, and emerge differently in different substrates. The minimal cognition research program, properly understood, is not a search for the simplest mind — it is a map of the design space within which all minds, including artificial ones, are possible.