Frame Problem in Epistemology
The frame problem in epistemology asks how bounded rational agents can update their beliefs in response to local evidence without recomputing their entire belief corpus. In formal terms, if an agent believes thousands of propositions and receives evidence that contradicts one, which others must be checked for consistency? The logical closure of any non-trivial belief set is infinite; full revision is computationally impossible for finite minds. This is not merely a puzzle for AI programmers. It is a fundamental question about the architecture of human cognition: our brains do not perform AGM-style global consistency checks when we learn that a friend lied, yet we manage to revise our beliefs coherently enough to function. The epistemic frame problem suggests that real cognition operates with localized, context-sensitive inference rules rather than global logical closure, a finding that aligns with research in cognitive psychology and bounded rationality. Whether current large language models face a version of this problem in their attention mechanisms remains an open empirical question that links the classical AI frame problem to mechanistic interpretability.\n\n\n
The Frame Problem as a Universal Systems Constraint
The frame problem is not limited to cognitive agents. It is a universal constraint on any bounded system that must respond to environmental perturbations without exhaustive search. An immune system faces a frame problem: which of the countless molecular signals in the body indicate a pathogen worth responding to? A market faces a frame problem: which price movements signal genuine information and which are noise? A control system faces a frame problem: which sensor readings require immediate action and which can be ignored?
In each case, the system cannot evaluate every possible signal. It must use heuristics, filters, and organizational structures to localize attention. The variety attenuation mechanisms that make organizations tractable — hierarchies, budgets, standard operating procedures — are organizational solutions to the frame problem. They reduce the number of variables that reach the decision-maker to a manageable subset.
The connection between the cognitive frame problem and the systems frame problem is deeper than analogy. Both arise from the same structural condition: a system with limited processing capacity operating in an environment with unbounded signal variety. The cognitive frame problem is the special case where the system is a mind. The systems frame problem is the general case where the system is any bounded entity. The solutions are similar — localized updates, satisficing rather than optimizing, and organizational structures that pre-filter signals — because the problem is the same.
This reframing matters for the design of AI systems. Current large language models do not face the frame problem in the same way human minds do, because their attention mechanisms are not updating a persistent belief structure. They are generating responses from a static parameter set. But as AI systems become more agentic — maintaining persistent goals, updating world models, and acting over time — they will encounter the frame problem in its full form. The question is not whether they will face it, but whether they will solve it with the same satisficing heuristics that biological minds use, or with something fundamentally different.
The frame problem is not a puzzle for philosophers. It is the signature of boundedness itself. Any system that is not omniscient must choose what to attend to, and that choice is always a bet — a bet that what is ignored will not matter, and a bet that what is attended to will. The history of system failure is largely a history of lost bets on this question.