Machines 'know more than they can tell': economists add five AI moves to Nonaka's knowledge spiral

The Human-Machine Knowledge Spiral

Aaron Chatterji, Daniel Rock, Eduard Talamas

econ.GN, cs.CY

2026-06-28

Extends Nonaka's knowledge spiral to AI: machines now hold tacit knowledge, five new knowledge movements appear, and the firm's job is still to provide shared context.

What problem this solves

Nonaka's 1991 theory of the knowledge-creating company rests on a cycle: individuals articulate tacit knowledge into explicit form, others internalize it through use, and knowledge spirals up through the organization. The theory carries an unstated assumption: knowledge lives in humans, and only in humans.

AI breaks that assumption. Models learn from data rather than following rules people wrote, so what they know is also hard to articulate. Polanyi's line 'we know more than we can tell' now applies to machines. Three economists (Duke, University of Pennsylvania, IESE Business School) name this object tacit machine knowledge and ask whether Nonaka's theory survives it, and whether the firm's role changes.

Method

This is a conceptual paper: theory reconstruction plus case analysis. The move is to add five machine-involving knowledge movements to Nonaka's original four conversions (socialization, externalization, combination, internalization).

• Encoding: documents, formulas, and experimental records go into a learning system and become machine capability.
• Embedding: people transfer tacit judgment through demonstrations, corrections, and ratings; the paper calls this socialization of machines.
• Extraction: the machine surfaces what it learned as predictions, designs, or recipes that humans can make and test.
• Assimilation: people who repeatedly build and refine machine proposals absorb new possibilities into their own tacit knowledge.
• Machine transfer: machine to machine. Knowledge in weights can be copied, distilled, fine-tuned, or served over an API without ever being articulated as rules.

The decisive property is portability. Human tacit knowledge travels slowly through personal contact; machine tacit knowledge stays opaque while moving at software speed.

The case is the food company NotCo and Giuseppe, its formulation system. In an early attempt, Giuseppe put dill in a plant-milk formula and the milk came out green. Nothing in the data says milk is not green; anyone standing in the kitchen knows it. NotCo's response was to spin the spiral: by 2021, more than twenty chefs and food scientists were testing over a hundred recipes a month, with sensory judgments fed back into Giuseppe. A combination no human team would likely have tried, pineapple and cabbage, became NotMilk, which reached US Whole Foods stores nationwide in 2020; the 2021 Series D raised $235M at a $1.5B valuation. NotCo then turned Giuseppe itself into a B2B platform (a $70M raise in 2022) and in 2025 partnered with chocolate maker Barry Callebaut to move Giuseppe into chocolate recipe development. That last step is machine transfer: accumulated tacit knowledge enters a new domain without being spelled out.

The conclusion reads ChatGPT's RLHF training (Ouyang et al., 2022) as the same spiral at scale: vast text encoded into the model, human demonstrations and rankings embedding judgment, outputs extracted for inspection, builders assimilating lessons from unexpected successes and failures.

Results

No experiments and no baselines; this is a theory piece. The supporting material is one company's business numbers:

The theoretical conclusion: Nonaka's central claim holds. The firm's job is still to provide shared context where different kinds of knowledge meet around the same object. What changes is that machines join the spiral, and knowledge can now cross organizations without ever being made explicit.

Why it matters

For people deploying AI inside companies, the paper first shifts the frame: the value of human-machine interaction goes beyond task automation, with machines and people feeding each other and producing knowledge together. It then supplies organizational language for familiar failures.

• Dropping a model into a business unit does not create organizational knowledge. Imported knowledge arrives with someone else's context and none of your goals; it counts only after entering a local human-machine spiral.
• Multiple experts evaluating the same output looks like duplication from an information-processing view. From a knowledge-creation view, the overlap builds common cognitive ground; the paper borrows Nonaka's concept of redundancy for this.
• Failures are valuable. The green milk exposed the distance between molecular resemblance and what milk means, making tacit knowledge visible.
• Knowledge creation cannot be parked in an isolated AI unit; Nonaka never assigned it to a single department either.

It also names a business shape: selling an internally accumulated model plus its feedback loop to peers. NotCo's B2B business and AI labs' API businesses are structurally the same thing.

Limitations

The authors concede part of this themselves. A footnote grants that machines do not seek knowledge as 'justified true belief'; they store and act on information in ways that resemble human knowledge. On Matsushita's bread machine, they retreat to the weak claim that knowledge crystallizes in products and triggers further cycles, without asserting full capture of the baker's tacit skill.

The structural problems are larger. A single positive case, sourced from press coverage and patents rather than internal data. The five movements are a taxonomy with no testable predictions. NotCo is a survivor; companies that installed AI and never spun the spiral do not appear. The Barry Callebaut partnership was announced, not studied; the paper's own phrasing is 'if the process follows' the feedback logic. And the lead author is OpenAI's chief economist. The paper states the work was not supported by OpenAI and does not represent its views, but the narrative of lab spirals feeding everyone else's spirals runs in the same direction as the employer's interests; readers can discount accordingly.

Terms

• Tacit knowledge: knowledge gained through experience that cannot be fully written down as rules; Polanyi's phrasing is 'we know more than we can tell.'
• Knowledge spiral: Nonaka's 1991 model in which tacit and explicit knowledge convert into each other (socialization, externalization, combination, internalization) and climb through the organization.
• ba: Nonaka and Konno's term for the shared context in which knowledge acquires meaning, such as the prototype on the kitchen table.
• Tacit machine knowledge: capability a model learns from data that cannot be articulated as rules, yet can be deployed at scale through software.
• Machine transfer: moving knowledge held in weights to other systems by copying, distilling, fine-tuning, or serving it over an API, without articulation.

Source

• arXiv 2606.29227v1
• PDF
• Hugging Face

What people are saying

• Tacit Knowledge Paper Argues Human-Machine Friction Matters Beyond Automatable R&D Tasks sebkrier

Related papers

• Off-the-shelf ChatGPT lifts a quiz 7.2 percentage points; 3.9 remain a week later
• Bare LLMs speed up 15% of US tasks; tools on top could expose half the work
• Once execution is free, rents migrate to underwriters: MIT's economics of AGI
• Kaggle medals expire in a year; format exit, not AI, drives most of the collapse
• U.S. Chip Controls Backfired: Chinese Devs Forked Open LLM Repos at 12x the U.S. Rate
• AI must hit 5.6x today's productivity to fund broad payouts; 33% public capture halves the bar

All paper explainers