Can machine tacit knowledge spread across organizations unlike human tacit knowledge?

The paper argues that AI introduces "tacit machine knowledge" into organizations — a new form of knowledge that, like human tacit knowledge, is hard to articulate ("machines now also 'know more than they can tell'") but, unlike human tacit knowledge, "need not remain local: once embodied in a model, it can be made available at scale through software." This portability is called the "key property of this new type of tacit knowledge," and the paper's claim is that Ikujiro Nonaka's 1991 theory of the knowledge-creating company still holds "mutatis mutandis" once this new tacit machine knowledge is added to the spiral between tacit and explicit knowledge that drives organizational innovation.

The reasoning runs through NotCo, the plant-based food company whose AI system Giuseppe searched ingredient and sensory data at a scale "no food scientist could match" to propose formulas for products like NotMilk. An early proposal containing dill came out green — Giuseppe had found a chemically plausible combination but, as the paper puts it, "it cannot taste the product, smell it, or understand everything people expect from a glass of milk." NotCo's fix was to build a continuous spiral: chefs and sensory panels tasted and judged each prototype, their judgments went back to Giuseppe, and Giuseppe generated new proposals in light of the feedback — eventually arriving at the "improbable but winning combination" of pineapple and cabbage. The paper reads this through Nonaka's concept of ba, a shared context where redundant, overlapping judgments from the machine, the chefs, and the food scientists let knowledge "acquire meaning through interaction." The organizational prescription follows directly: "access to tacit machine knowledge alone — no matter how impressive — does not create organizational knowledge... The imported knowledge becomes organizational knowledge only after it enters the human-machine knowledge spiral."

This gives "tacit knowledge" a second, organizational-theory sense alongside the mechanistic one in Do language models possess tacit knowledge in Davies' sense?, which locates tacit knowledge inside a model's weights as causal structure it does not explicitly represent; this paper instead treats tacit machine knowledge as a black box whose outputs must be externally judged by humans before they count as organizational knowledge at all. The NotCo loop of machine proposal, human judgment, and revised proposal also parallels the pattern in How should AI agents and humans divide research tasks?.

The excerpt is a conceptual, theory-building paper, not a measured study: NotCo and the ChatGPT training example are illustrations chosen to fit Nonaka's framework, not evidence that the framework was tested against alternatives or that it predicts anything they don't. The implication the paper draws — that the knowledge-creating company's job is unchanged in kind, only in what kinds of tacit knowledge it must now circulate — rests on the strength of the analogy rather than on an independent test of it.