Knowledge as Capability: A Better Foundation for Human-AI Cooperation

An operational definition of knowledge as the ability to produce outcomes, and why that makes human-AI cooperation easier to reason about.

For organizations, I define knowledge as the capability to reliably produce outcomes. I use the two expressions interchangeably.

I use this as a practical definition for governing human-AI cooperation, without proposing a metaphysical account of what knowledge is. Whether AI has mental states remains largely unanswerable and secondary to organizations’ practical need to work with these systems.

For nearly a century, machines have been producing outcomes we’d call “knowledge work” if produced by humans. The Enigma machine embodied cryptographic knowledge in its rotor wiring. A TI-30 calculator embeds trigonometric knowledge you don’t need to possess to use. Expert systems of the 1980s were explicitly called “knowledge-based systems.” None of this was controversial.

Large language models extend the range of machine capability: they produce working code, accurate summaries, and valid analyses across more domains than earlier systems. Their usefulness has made the question of whether machines know things more urgent.

The definition lets us compare human and AI contributions within the same framework for knowledge work.

The Practical Problem

Epistemology is the branch of philosophy concerned with knowledge: what it is, how we acquire it, when we can legitimately claim to have it. Traditional epistemology, up to the 1960s at least, largely settled on a working definition: knowledge is justified true belief.

The three conditions are independent and all must hold. You cannot have knowledge that the earth is flat. It fails the truth condition regardless of how sincerely you believe it or how convinced you are of your reasons. Belief that the earth is an oblate spheroid because a turtle told you also isn’t knowledge. It happens to be true, but “a turtle told me” isn’t justification. Correctly guessing tomorrow’s lottery numbers isn’t knowledge either but a lucky guess.

Since the 1960s, epistemological debate has become more complex, but the frameworks still assume agents with mental states. Justification, reliable belief formation, intellectual virtue, and sensitivity to truth all concern the epistemic status of a believer.

When an AI system produces working code, the traditional framework has no place to put that capability. Does the AI “believe” its outputs? The question has no answer because it’s built on a category error. AI systems are not the kind of entity belief-language was designed to describe.

Yet AI demonstrably does useful work that, in humans, we’d call knowledge work. We face a choice: either insist on definitions that exclude AI from the category entirely, or find operational definitions that make human-AI cooperation coherent. I’m proposing the latter.

Unresolved Questions About Minds

The operational definition does not depend on resolving the philosophical questions in this section. For its practical application, see The Operational Definition.

Philosophy hasn’t solved the solipsism problem. I cannot verify that you have experiences, only that you behave as if you do. Neuroscience hasn’t solved the hard problem of consciousness. Psychology doesn’t have a settled account of what cognition is or how it produces understanding. The confident claim that humans possess knowledge in some robust sense while AI systems categorically cannot assumes we know what knowledge-constituting mental states are and how to detect them. We don’t. That knowledge is not established in any discipline.

We extend epistemic charity across the solipsism gap for other humans. I assume you have experiences because you behave as if you do, even though I cannot verify this. Reasonable, but a choice, not a proof. The same logic applies to AI systems. We don’t know how they achieve their results. We can’t rule out that something analogous to mental states emerges from their operation. Confidently denying AI mental states requires the same leap of faith as confidently affirming human mental states. We just happen to be more comfortable with one leap than the other.

We lack the theory to establish whether AI systems have consciousness or experiences. Confident affirmation and dismissal are equally unjustified. Dismissing the systems as “just pattern matching” assumes criteria for human understanding that we have not established.

“It’s just matrix multiplication” has the same logical force as “it’s just electrochemistry”: none. Describing a process mechanically tells us nothing about whether it constitutes understanding; human brains can also be described mechanically. These dismissals assume that understanding has settled criteria and that AI fails to meet them. No discipline has established those criteria.

Assess systems by what they can do, while acknowledging uncertainty about their mechanisms and mental states.

Embedded Knowledge Has a Long History

The idea that artifacts can embody knowledge is not new or controversial. It has a seventy-year intellectual lineage.

Cybernetics established the framework. W. Ross Ashby’s Law of Requisite Variety (1956) states that effective control requires a regulator as complex as the system being regulated. Conant and Ashby’s Good Regulator Theorem (1970) makes this explicit: “Every good regulator of a system must be a model of that system.”

The epistemological implication is direct. A thermostat must embody knowledge about temperature dynamics to regulate them. The knowledge is in the mechanism, not in any mental state. David Armstrong used thermometer readings as a model for externalist epistemology precisely because they track temperature without believing anything.

Expert systems in the 1980s were explicitly called “knowledge-based systems.” Edward Feigenbaum defined them as programs that use “knowledge and inference procedures to solve problems that are difficult enough to require significant human expertise.” DENDRAL, MYCIN, and their successors contained “knowledge bases.” The terminology was deliberate. The philosophical debate was about whether this knowledge was adequate, not whether calling it knowledge was appropriate.

Distributed cognition (Edwin Hutchins) showed that cognitive processes extend across human-artifact systems. Navigation on a ship isn’t done by any single mind but by a system including people, charts, compasses, and procedures. The knowledge is distributed across the network.

The Extended Mind thesis (Clark and Chalmers, 1998) argues that cognitive processes can extend beyond the brain. Their Parity Principle: if a process in the world functions identically to one we’d recognize as cognitive if it happened in the head, it’s part of the cognitive system. A notebook can be external memory in the same sense that biological memory is internal memory.

These traditions provide an established basis for assessing knowledge through system performance.

The Operational Definition

I equate “knowing how” with “being able to do reliably.” The test is whether an agent can produce the outcome; capability has the role that justified true belief has in traditional definitions.

The key terms:

Agent: Any system capable of action, including a human, AI, organization, or human-AI hybrid, regardless of its implementation.

Outcome: An observable, verifiable state of the world, such as working code, a healed patient, a profitable trade, or a valid proof.

Context: The conditions under which capability is claimed. Knowledge claims are always bounded. “Can diagnose appendicitis in routine presentations” is a different claim than “can diagnose any abdominal condition.”

Reliably: Consistently across instances. Every domain requires repeatability, although surgery demands a higher threshold than brainstorming.

This definition has several features:

It’s operational. You verify knowledge by observing performance, not by inspecting mental states.

It’s agent-relative. The same capability may be present in one agent and absent in another.

It’s measurable. Degrees of knowledge correspond to reliability of outcome production across context instances.

It handles AI naturally. If the system can do the task, it has the knowledge. Whether it “understands” in some deeper sense is a separate question we can’t currently answer.

Tacitness and Composite Knowledge

Tacitness Becomes Relational

Traditional discussions of tacit knowledge (Polanyi’s “we know more than we can tell”) treat tacitness as an intrinsic property of knowledge itself. Under an operational definition, tacitness becomes a relationship between knowledge and knower.

Knowledge is tacit for agent B when no representational path exists from the knower to B. The capability exists but cannot be transferred through any available encoding and decoding process.

This makes tacitness agent-relative rather than intrinsic. The same knowledge can be tacit between a machinist and a novice, explicit between the machinist and another machinist, and newly explicit between the machinist and an AI system that can decode the machinist’s demonstrations.

For a human doctor: “This patient presentation suggests condition X” may be tacit. She can’t fully articulate why.

For an AI trained on millions of cases: The same diagnostic conclusion may be entirely computable from patterns in the training data.

Belief-based frameworks make this transfer relationship difficult to express. Defining knowledge as capability makes it easier to describe and provides new possibilities for human-AI collaboration in knowledge transfer.

Composite Knowledge Becomes Coherent

Executable Knowledge Architecture implements assistant pattern (asking AI to execute rather than answer) for professional work: it asks AI to translate expert intent into executable code rather than to provide answers directly. This only makes sense under a capability definition.

The expert has knowledge: the capability to recognize correct outcomes, specify intent, and verify alignment. The AI has different knowledge: the capability to produce executable implementations from specifications.

Neither alone constitutes the knowledge needed to solve problems reliably. Together, through a verification loop, they produce validated solutions. The composite system has capability that neither component has alone.

A belief-based definition cannot identify who believes the composite solution. A capability definition identifies the human-AI system that reliably produces verified solutions.

Tacit Space Can Shrink

As AI capabilities expand, more knowledge becomes transferable that was previously locked in human expertise. An AI that can learn from video demonstrations can decode knowledge that was previously tacit because the only transfer path was embodied apprenticeship.

This “shrinkage of tacit space” is only coherent if tacitness is about transfer relationships, not intrinsic properties. And that’s only coherent if knowledge is capability, not belief.

Scope of the Operational Definition

Whether AI has consciousness, phenomenal experience, or human-like understanding remains open.

The operational definition leaves aside the questions of justification, mental states, and epistemic normativity addressed by traditional epistemology. It therefore captures only part of what philosophers mean by knowledge.

Knowledge management theorists primarily described human knowers with consciousness, embodiment, and social relationships. Some definitions, including Senge’s “capacity for effective action,” support this proposal. Others, such as Nonaka’s emphasis on human subjectivity in tacit knowledge, do not. I use the parts relevant to governing human-AI work without claiming to extend the discipline as a whole.

For governing human-AI cooperation, capability-based definitions work better than belief-based definitions. We can assess which tasks a system performs, how reliably, in which contexts, and with which failure modes.

The Intellectual Lineage

This operational approach draws on established traditions, even if those traditions don’t fully support all implications:

Pragmatic philosophy (Peirce, James, Dewey): Knowledge defined by practical consequences and action. Dewey’s instrumentalism treats ideas as tools for problem-solving, not representations of fixed truths. “Knowledge is an instrument of action.” Dewey’s naturalistic continuity, treating human cognition as continuous with biological adaptation, provides the closest classical support for extending knowledge vocabulary to non-human systems.

Procedural knowledge (Ryle, Polanyi): “Knowing how” cannot be reduced to “knowing that.” We know more than we can tell. Knowledge is about being able, not about holding propositions. Polanyi’s framework was built for embodied human knowers, but his emphasis on capability over articulation points in a useful direction.

Cybernetics (Ashby, Beer): Knowledge as variety management for effective regulation. A control system must be capable of producing the responses required to manage disturbances. The Good Regulator Theorem directly implies that effective regulators embody models of what they regulate.

Reliabilism in epistemology (Goldman, Armstrong): Knowledge as true belief produced by reliable processes. This externalist approach doesn’t require conscious access to justification. What matters is the reliability of the process, not the agent’s introspective awareness of that reliability.

Organizational capability (Senge): Peter Senge defined knowledge as “the capacity for effective action,” adding, “There is no capacity for effective action in a database.” A database passively stores information and returns it on query. A calculator computes sin(45°) with encoded algorithms rather than retrieving it from a table. Language models likewise generate outputs through weighted transformations. Senge’s exclusion applies to passive storage; his definition allows computational systems that perform effective action to embody knowledge.

Some of these thinkers would reject my proposal, although their work provides decades of material for developing capability-based accounts of knowledge.

For Practitioners

Apply the definition by examining capabilities alongside documented knowledge:

Alongside “what do we know?” ask “what can we do?” The former maps your documented knowledge; the latter maps your actual capabilities. Both matter.

Alongside propositional capture, use practice, pairing, and system-building. Documentation works for some knowledge; other methods work for others. The capability lens helps you choose.

Alongside documentation metrics, track outcomes. Completion rates, success rates, capability demonstrations. The goal is capability transfer; measure whether it’s happening.

Include AI in your capability assessments. AI capabilities are organizational capabilities. They can receive knowledge transfer (training, prompting) and contribute to knowledge work. The capability definition gives you vocabulary for this.

Accept that capability can exceed articulation. Experts who can do but cannot fully explain still have knowledge worth transferring. Work with them to find transfer paths, including AI-assisted articulation.

The Foundation

Several further claims depend on this definition.

Agent-relative tacitness only makes sense if knowledge is capability, because tacitness becomes about transfer of capability, not about hidden beliefs.

Tacit space shrinkage only makes sense if tacitness is agent-relative, because AI capabilities can open new transfer paths.

Executable Knowledge Architecture only makes sense if human-AI systems can have composite capability, distributed across agents.

The governance challenges I’ve written about, knowledge that is explicit-for-systems but opaque-for-humans, only make sense if explicitness is about transfer relationships, not about propositional content.

This framework does not replace epistemology or knowledge management. Those disciplines work at an academic pace. Organizations adopting AI cannot wait for them, and any organization that wants a future beyond Amish hand-crafted furniture needs to adopt AI.

Reliable performance is the criterion for knowledge in this framework.