If a company just rents access to an AI model, does it actually absorb the model's hard-won expertise, or just borrow a tool?
Can API access to trained models transfer tacit knowledge to companies like NotCo?
This explores whether a company that rents a trained model through an API, such as the food-tech firm NotCo, actually gets the hard-to-articulate know-how built into that model, or only gets a tool it still has to make sense of itself.
This explores whether renting a model through an API passes on real tacit know-how, the kind of skill that is hard to write down, to an outside company like NotCo. Or does the company only get a capable tool it still has to learn to use? The corpus's answer is partly. The know-how does travel, but it doesn't become the company's own knowledge just because the company has access. One caveat: the corpus doesn't study NotCo directly. NotCo works here as an example of the kind of company that receives this knowledge.
The case for transfer is strong. Human tacit knowledge stays stuck in people and places: a master chef's feel for texture leaves the building when the chef does. Machine tacit knowledge is different. Once it is built into a model, it can be copied to any number of organizations at once Can machine tacit knowledge spread across organizations unlike human tacit knowledge?. That knowledge has a specific origin. In RLHF, thousands of human raters turn their gut judgments into rankings, and those rankings shape the model. This repeats inside the lab the cycle Nonaka describes, in which unspoken knowledge becomes explicit and then becomes unspoken skill again. The output then reaches downstream companies through products and APIs Where does ChatGPT's tacit machine knowledge actually originate?. Whether 'tacit knowledge' is the right term for what a model holds is still open. The best evidence that it fits comes from a single fact-editing experiment, which has been hard to replicate Do language models possess tacit knowledge in Davies' sense?.
The catch is that portable is not the same as absorbed. The same work argues that machine know-how becomes organizational knowledge only through repeated feedback loops in which people test outputs, build shared context and decide what results mean. There is also a hard ceiling. No prompt can supply domain knowledge the model never learned in training; prompting only reorganizes what is already there Can prompt optimization teach models knowledge they lack?. So a company working on something niche, like how plant proteins behave, gets the model's general reasoning for free and has to provide the specialist part itself. Renting inference instead of building your own understanding is attractive, but models have been shown to invent 35–49% of their claims about users, so rented insight still needs checking Can LLMs infer user needs better than owned behavioral data?.
The surprising point is that the most effective transfer in the corpus runs the other way. In one industrial case, a company wrote its own experts' rules and design principles into the scaffolding around an LLM. Output quality rose 206%, and non-experts produced work that experts rated at their own level Can codified expertise let non-experts match specialist output?. A related finding: a stronger model can nearly double a weaker model's performance by building a harness around it, without retraining Can a stronger model lift a weaker one at test time without retraining?. If a company does want to put its own knowledge into the model's weights, how it organizes that knowledge counts for more than how much it has. Arranging data into a domain taxonomy reached half of full performance using just 0.3% of the data Can organizing knowledge structures beat raw training data volume?.
So the API brings in the lab's distilled judgment, but the advantage goes to the company that wraps that judgment in its own codified expertise and feedback loops. There is also a social limit. Inside a field, expertise is confirmed by membership in a community and a track record, which a model can't build for itself Can AI ever gain expert community trust through participation?. The company's own people still have to vouch for what the machine knows.
Sources 9 notes
AI introduces tacit machine knowledge that, unlike human tacit knowledge, can be distributed at scale once embodied in models. However, it becomes organizational knowledge only through human-machine feedback loops that build shared context and meaning.
RLHF training replays Nonaka's tacit-explicit cycle within AI laboratories: human raters' tacit judgments become explicit rankings, which train models whose outputs reveal new capabilities. This lab-internal spiral produces portable tacit machine knowledge that feeds knowledge spirals across countless organizations via software products and APIs.
Transformer LLMs can meet Davies' criteria for tacit knowledge based on architectural features and causal tracing with ROME edits. However, evidence rests on a single fact-editing case, and replication challenges suggest the causal localization may not be as precise as initially claimed.
Prompting works entirely within a model's pre-existing training distribution and cannot supply domain knowledge absent from training data. This creates a hard ceiling: no prompt strategy can compensate for missing foundational knowledge, only reorganize what already exists.
Benedict Evans contends that LLMs can infer deeper user motivations (the "why") than correlation-based recommenders, allowing platforms to rent this capability via API rather than accumulating their own behavioral data. However, research shows LLMs fabricate 35–49% of user attribute claims, undermining confidence in their inferred understanding.
Show all 9 sources
An industrial case study embedding domain rules and design principles into an LLM agent's scaffolding achieved 206% output-quality improvement and expert-level ratings from non-experts, bypassing the need for specialist oversight. The capability gain came from externalizing tacit expertise into structured harness components, not from model scale.
A stronger model built inference-time harnesses that nearly doubled weaker model performance on Theory-of-Mind benchmarks without retraining, primarily by moving unstable reasoning into deterministic code and task-specific routing rather than encouraging extended reasoning.
StructTuning achieves 50% of full-corpus performance using only 0.3% of training data by organizing chunks into auto-generated domain taxonomies. The model learns knowledge position within conceptual structures rather than raw text patterns, matching how students learn from textbooks.
Expertise is validated through social participation and track record within expert communities, not individual accuracy alone. AI cannot enter this validation circle because it lacks social embeddedness, testable judgment history, and ability to participate in the consensus-building processes that define expert paradigms.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- What Do Large Language Models Know? Tacit Knowledge as a Potential Causal-Explanatory Structure
- Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills
- The Human-Machine Knowledge Spiral
- Injecting Domain-Specific Knowledge into Large Language Models: A Comprehensive Survey
- Eliciting Reasoning in Language Models with Cognitive Tools
- Educating LLMs like Human Students: Structure-aware Injection of Domain Knowledge
- How to Build AI Agents by Augmenting LLMs with Codified Human Expert Domain Knowledge? A Software Engineering Framework
- AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses