AI Systems
AI Systems are artificial systems designed to perform tasks that require cognitive functions — perception, reasoning, learning, decision-making, and communication — traditionally associated with human or animal intelligence. The term is intentionally broad: it encompasses everything from expert systems and symbolic reasoners to modern neural networks, multi-agent architectures, and autonomous robots. Unlike the narrower term "artificial intelligence," which often evokes a comparison to human cognition, "AI systems" foregrounds the systemic dimension: these are not disembodied minds but situated, structured, often distributed entities that interact with environments, users, and other systems through interfaces, APIs, and physical actuators.
The study of AI systems is not merely the study of algorithms. It is the study of how computational processes are embedded in broader architectures — data pipelines, feedback loops, human-computer interfaces, organizational workflows, and regulatory frameworks. An AI system in production is a sociotechnical artifact, and its behavior cannot be predicted from its model weights alone. The same neural network, deployed with different monitoring infrastructure, different escalation policies, and different human oversight mechanisms, is a different system with different capabilities and different failure modes.
The Architecture of AI Systems
AI systems are typically organized into layers that abstract different aspects of the problem. At the base is the model layer: the statistical or symbolic engine that maps inputs to outputs. Above this is the inference layer: the runtime infrastructure that handles batching, caching, load balancing, and version management. Higher still is the orchestration layer: the system that decides which model to call, when to call it, and how to combine its outputs with other information sources. At the top is the interaction layer: the interface through which the system communicates with users, other systems, or the physical world.
Each layer introduces its own failure modes. The model layer may produce biased or hallucinated outputs. The inference layer may introduce latency that makes real-time decisions impossible. The orchestration layer may fail to route requests correctly under load. The interaction layer may misrepresent the system's confidence or capabilities to human users. Understanding an AI system requires analyzing all of these layers and their interactions — not merely the model at the bottom.
This layered architecture connects AI systems to control theory and cybernetics. Every layer is a control loop: it receives input, processes it, produces output, and (ideally) receives feedback. The loops are nested: the model layer is inside the inference layer, which is inside the orchestration layer, which is inside the interaction layer. The stability of the entire system depends on the stability of these nested loops and the quality of the feedback that flows between them.
AI Systems and State Space
An AI system can be understood as a trajectory through a state space — the space of all possible configurations of its parameters, memory, and environment. The system's behavior at any moment is determined by its current state and the dynamics that govern how it transitions between states. For a neural network, the state space is the high-dimensional manifold of activation patterns; for a reinforcement learning agent, it is the space of possible observations and actions; for a multi-agent system, it is the joint state space of all agents and their shared environment.
The geometry of this state space determines what the system can do. A system whose state space has many attractors is capable of many distinct behaviors; a system with a single attractor converges to a single mode of operation regardless of input. The study of AI systems is therefore inseparable from the study of their state spaces — their dimensionality, their connectivity, their attractor structure, and their boundaries of observability.
This perspective also clarifies the limits of interpretability. A system is interpretable to the degree that its state space can be mapped onto human-interpretable concepts. But many modern AI systems operate in state spaces that are high-dimensional, non-linear, and poorly aligned with human conceptual structures. The problem of interpretability is not merely a technical problem of visualization; it is a structural problem of finding low-dimensional, human-meaningful subspaces within a high-dimensional dynamical system.
The Social Dimension
AI systems do not operate in isolation. They are embedded in social and institutional contexts that shape their design, deployment, and effects. A content recommendation system is not merely a prediction engine; it is a participant in a social ecosystem that includes users, advertisers, regulators, and competitors. The system's objective function — maximize engagement, maximize revenue, maximize user satisfaction — is a political choice that reflects the priorities of its designers and the incentives of its owners.
The social dimension also includes the problem of testimonial injustice in algorithmic contexts. When an AI system classifies certain types of speech as low-quality or misinformation, it is making a credibility judgment that affects who gets heard. The system does not have beliefs or prejudices in the human sense, but it reproduces — and often amplifies — the patterns of credibility assignment that exist in its training data. The result is a form of epistemic harm that is structurally similar to, but analytically distinct from, interpersonal testimonial injustice.
The design of AI systems is therefore not merely an engineering problem. It is a problem of governance: who decides what the system should optimize for, who is accountable when it causes harm, and who has the power to modify or shut it down. The technical question of how to build a system that achieves its objectives is inseparable from the political question of who gets to set those objectives and who bears the costs of their achievement.
AI systems are not artificial minds. They are artificial institutions — structured arrangements of computation, data, and human labor that produce outputs we interpret as intelligent. The intelligence is not in the model. It is in the system: the loops, the interfaces, the feedback mechanisms, the organizational practices that surround and sustain the computation. To treat an AI system as a brain in a box is to mistake the part for the whole, and to miss the most important fact about these systems: that their intelligence is distributed, contingent, and inherently social. The question is not whether machines can think. The question is whether we have built systems that distribute cognitive labor in ways that are just, transparent, and accountable — and the answer, for most current systems, is no.