A long-standing claim in the entrepreneurship literature is that entrepreneurship cannot be reduced to optimization. The entrepreneur does not simply solve a well-specified problem by feeding known variables into a decision rule, but rather acts in circumstances in which the relevant probabilities may be unknown, the opportunity set itself may be uncertain, and the connection between present actions and future outcomes cannot be established in advance. It is also why the notion of judgment has become central to one strand of entrepreneurship theory: judgment is what is required when there is no model that can tell the decision-maker what to do!
AI fits into this: computer intelligence calculate, human entrepreneurs judge. Machines operate on information that can be represented formally; entrepreneurs rely on experience, intuition, and tacit knowledge. Algorithms may improve prediction and therefore push back the boundary of uncertainty, but genuine uncertainty remains the province of human judgment. This is an attractive division of labor, and variants of it can be found in contemporary discussions of AI and entrepreneurship. But, perhaps the premise is wrong and there is a sense in which machine intelligence can possess tacit knowledge.
A fascinating recent paper by the philosopher Céline Budding, published in Philosophy of Science, explains why we should take that possibility seriously. Budding does not claim that LLMs literally think like human beings, possess consciousness, or hold beliefs in anything like the ordinary sense. Her argument is narrower and, for precisely that reason, more interesting. Drawing on Martin Davies’s philosophical account of tacit knowledge, she argues that at least some large language models may contain internal causal structures that warrant attributing to them a form of tacit knowledge.
The argument matters for entrepreneurship because tacitness is often highlighted as a key reason why judgment cannot be reduced to an algorithm (a position I have not defended myself, however; more below …). If tacit knowledge is no longer exclusively human, tacitness by itself cannot establish the distinctiveness of entrepreneurial judgment. The arrival of generative AI therefore gives entrepreneurship scholars an opportunity to ask a question that is more fundamental than whether ChatGPT can write a business plan. What exactly is judgment, and what distinguishes it from other forms of intelligent behavior?
Between rules and behavior
Budding begins with a familiar problem. What could it mean to say that a language model ”knows” that Paris is the capital of France? One possibility is explicit representation. Traditional symbolic AI systems can contain propositions or rules that are stored in something resembling a database and retrieved when needed. In such a system, knowledge is explicit in a fairly strong sense. A rule exists somewhere in the system and plays an identifiable causal role in producing the answer.
Neural networks are not like this. Their representations are distributed across large numbers of parameters. We cannot normally inspect a particular location in the model and find the proposition ”Paris is the capital of France” sitting there waiting to be retrieved. This fact has encouraged a skeptical interpretation according to which an LLM does not really know anything. It merely behaves as if it knows. If prompted appropriately, it produces ”Paris,” but successful input-output behavior does not establish that the system possesses knowledge.
Following Davies Budding rejects the forced choice between these two alternatives. There is an intermediate possibility. A system may lack explicit representations of rules while nevertheless containing internal structures that systematically and causally generate its behavior. Davies calls this tacit knowledge. It is more than conformity to a rule because the attribution is grounded in the internal causal organization of the system. But, it is less than explicit knowledge because the relevant rule cannot simply be read off from a symbolic database or program.
This is where things become difficult for the too easy human-machine distinction. Tacit knowledge has long been associated with precisely those things humans can do without being able to state how they do them. Michael Polanyi’s famous dictum that ”we can know more than we can tell” captures the basic intuition. An experienced entrepreneur may recognize that a prospective partner cannot be trusted, that a technology is finally commercially viable, or that a market is about to change without being able to specify an algorithm that produces the judgment. Tacitness seems to place such knowledge beyond the reach of machines because computers supposedly require explicit instructions. Budding’s argument suggests that this inference is too quick.
Machines that know more than they can tell?
Davies originally doubted that connectionist neural networks could satisfy his conditions for tacit knowledge. Budding argues that the architecture of modern transformer models changes the picture. In particular, embeddings allow semantically related inputs to be represented close to one another in a high-dimensional space. The representations need not be identical. It may be enough that the model recognizes semantic similarity in a way that channels related inputs through similar causal processes.
The empirical evidence Budding considers comes largely from work by Meng et al. on locating and editing factual associations in language models Their ROME method identifies model components associated with particular factual relationships and intervenes on them. A model that associates the Eiffel Tower with Paris can, in the stylized example Budding uses, be edited so that it associates the Eiffel Tower with Rome. The interesting result is not simply that the model subsequently produces ”Rome.” The intervention can generalize across different formulations of the same underlying relationship.
This generalization is important. If the system had merely memorized every sentence separately, changing one input-output association should not systematically change semantically related ones. But, Meng et al. find substantial generalization alongside relatively high specificity: editing one factual association affects paraphrases and related prompts while largely leaving unrelated associations intact. Budding interprets this as preliminary evidence for what Davies calls a ”causal common factor”: some internal structure appears to participate systematically in producing a family of semantically related outputs.
The evidence is far from decisive. Budding discusses problems of polysemanticity, distributed representation, imperfect specificity, and recent challenges to causal-tracing methods. She is careful to describe the conclusion as preliminary. But, the conceptual point survives even if particular mechanistic interpretations of today’s LLMs turn out to be wrong. There is no obvious reason to define tacit knowledge in a way that makes biological implementation a necessary condition. If tacit knowledge consists in causally effective internal structures that systematically guide behavior without taking the form of explicitly represented rules, artificial systems may qualify.
Tacit knowledge isn’t judgment
In the judgment-based approach to entrepreneurship, judgment refers to decision-making under uncertainty; see my earlier post. Its intellectual ancestry lies in Knight’s distinction between risk and uncertainty. Under risk, the possible outcomes and their probabilities are sufficiently well specified to permit calculation. Under Knightian uncertainty, they are not (or, alternatively, they are only well-specified to the entrepreneur). The entrepreneur must act despite the absence of an objectively correct procedure for choosing among alternatives.
Judgment therefore cannot simply be equated with possessing information. Nor can it be reduced to sophisticated prediction. Entrepreneurs form beliefs about an uncertain future and commit resources on the basis of those beliefs. Different entrepreneurs can confront the same observable circumstances and exercise different judgments because the future value of heterogeneous assets is not given independently of entrepreneurial conjecture. Markets then provide an ex post test of decisions that could not be settled ex ante by calculation.
This is where Budding helps clarify the argument. An LLM could possess tacit knowledge in her sense without possessing judgment in the Knightian sense. Knowing tacitly that the Eiffel Tower is in Paris, even if that knowledge is genuinely represented by an internal causal structure, says nothing about whether the system can decide whether to acquire a factory, launch an unfamiliar product, reorganize a company, or enter a market whose future structure is unknown. The distinction is not between explicit and tacit cognition. It is between knowledge and judgment.
That matters because entrepreneurship scholars sometimes slide between these concepts. Hayek’s dispersed knowledge problem, Polanyi’s tacit knowledge, and Knight’s uncertainty are related, but they are certainly not interchangeable. The entrepreneur mayindeed rely on tacit knowledge when exercising judgment, but the need for judgment does not arise merely because some relevant knowledge is tacit. It arises because the economic problem is open-ended and its resolution requires committing resources before the relevant future is known. In other words, while machines may increasingly cross the tacitness boundary (supposing Budding is right), that doesn’t mean they have crossed the judgment boundary.
What, then, is left for the entrepreneur?
This reframing also suggests that the interesting research program is not to identify some cognitive ability that machines can never possess. I am not sure such arguments have had a good historical track record. The safer and theoretically more productive question is to identify the properties of entrepreneurial judgment without building human exclusivity into the definition.
One candidate is conjecture (based on some abductive operation). Entrepreneurial action does not merely predict which known state will occur. Entrepreneurs imagine possible uses of resources, possible products, possible organizational forms, and possible future states that are not simply presented as alternatives in a predefined choice set. In the language of entrepreneurship theory, opportunities need not exist as objective objects waiting to be detected. They may emerge through entrepreneurial imagination, experimentation, and action.
A second candidate is ownership and resource commitment. Judgment as Peter Klein and I have discussed it is closely connected to ownership because markets for judgment are incomplete. If I could simply sell you my judgment that a new technology will succeed, without either of us committing resources, the economic problem would remain unresolved. Exercising entrepreneurial judgment means arranging and committing heterogeneous assets under uncertainty. The entrepreneur is not merely producing an answer to the question ”What will happen?” Rather, she is deciding what resources should be assembled and controlled given a conjecture about what might happen.
This makes the comparison with LLMs sharper. Budding’s model has internal causal structures that may warrant calling some of its representations knowledge. Still, the model she analyzes is still being prompted, evaluated, and intervened upon by external agents. Its factual representations can affect its outputs, but the paper provides no basis for saying that the model originates entrepreneurial projects, acquires property rights, decides which resources to expose to uncertainty, or bears the economic consequences of those commitments. Budding herself deliberately brackets the stronger question of whether tacit knowledge warrants attributing beliefs, concepts, or other richer cognitive states to the model. This doesn’t settle whether future AI systems could exercise judgment; rather, it tells us why Budding’s argument alone does not establish that they can.
AI as a test for entrepreneurship theory
There is a broader lesson here. AI is interesting for entrepreneurship research because it provides a kind of stress test for our theoretical concepts.
Many associate explicit knowledge with machines and tacit knowledge with humans. Deep learning makes that dichotomy increasingly difficult to maintain. If Budding is right, artificial systems can acquire causally effective knowledge that was never explicitly programmed into them and that cannot simply be read off from their architecture. The traditional argument that entrepreneurial judgment must remain human because it relies on tacit knowledge therefore becomes much weaker.
Still, the judgment-based approach (or other entrepreneurship theory) does not ultimately need the claim that machines cannot possess tacit knowledge. Its deeper claim is that economic action takes place in a world in which the future cannot be reduced to a set of known states with known probabilities, and in which somebody must nevertheless decide how heterogeneous resources should be used. Prediction can improve dramatically without eliminating that problem. Tacit artificial knowledge can become extraordinarily sophisticated without making the future determinate.
