What is tacit knowledge, and why can't AI just learn it?
Tacit knowledge is the know-how you pick up from experience — riding a bike, sensing when a negotiation is going sour, or knowing which perfume notes blend well without consulting a formula. It's personal, context-dependent, and often so automatic that experts can't fully explain it. AI, by contrast, works with explicit, codified knowledge: rules, data, patterns it can extract from examples. The two are fundamentally different forms of cognition, creating what researchers call a 'cognitive asymmetry' [2]. In a study of professional perfume creation, researchers found that human perfumers rely on embodied, tacit knowledge — the feel of a scent — while generative AI works from codified chemical data. This gap means AI can suggest novel combinations, but it can't judge whether a scent 'works' in the way a human nose can [2]. Similarly, in recruitment, AI can screen resumes for keywords, but it can't assess cultural fit or read a candidate's unspoken cues — that requires the recruiter's tacit understanding of their organization's social dynamics [5].
The limits of AI's ability to capture tacit knowledge are stark. In a pipe corrosion inspection system, engineers interviewed experts to convert their tacit knowledge into explicit rules for a machine learning model. Even after this careful process, the AI only reached 70% accuracy — meaning human experts still had to review nearly a third of the cases [9]. A literature review on AI and tacit knowledge concluded flatly that 'humans remain irreplaceable in capturing tacit knowledge' [6]. The reason is simple: tacit knowledge is often unconscious and situational. An expert may not know why they spot a flaw in a weld or a subtle shift in a patient's condition — they just do. AI can't learn what isn't explicitly taught or labeled.
How do humans and AI actually collaborate in a workflow?
The most effective human-AI workflows don't try to replace humans or hand everything to AI. Instead, they split tasks based on what each does best. AI handles high-volume, pattern-based work — scanning thousands of satellite images for potential archaeological sites, for instance — while humans apply judgment, context, and tacit knowledge to interpret the results. In archaeological site detection, an AI model achieved about 80% accuracy, but researchers emphasized that even inaccurate predictions were useful when 'interpreted by a trained archaeologist' [1]. The human expert's role was to refine the dataset, correct errors, and decide which predictions warranted an onsite survey. This is a classic human-in-the-loop design: AI proposes, human disposes.
This collaborative pattern appears across domains. In predictive maintenance for factories, an AI-based system achieved 82% accuracy, but the framework explicitly included a 'human-in-the-loop mechanism' to incorporate expert tacit knowledge for final decisions [4]. In COBOL legacy system modernization, researchers found that AI could help with automation and knowledge abstraction, but that a 'hybrid human–AI model' was essential to manage tacit knowledge transfer from retiring experts [7]. The key insight from a study of 523 employees was that AI collaboration actually amplified the benefits of human mentorship: when employees worked alongside AI, they were better able to acquire tacit knowledge from their mentors, boosting creativity [8]. AI doesn't replace the human knowledge transfer — it enhances it.
What goes wrong when you remove the human from the loop?
The COVID-19 pandemic provided an accidental experiment in what happens when human interaction — and the tacit knowledge sharing it enables — is disrupted. A study of AI practitioners found that remote work, even for people already used to it, led to a 'decrease in tacit knowledge sharing' and a consequent drop in the diversity of AI project outputs [3]. The researchers noted that AI practitioners' interactions are 'partly embedded in AI tools and partly in human exchange' — and when the human exchange was reduced, the quality of work suffered. This wasn't about losing access to data or computing power; it was about losing the informal, face-to-face conversations where tacit knowledge flows.
The same principle applies to explainable AI. Even when AI models can generate explanations for their decisions, those explanations may not be understandable from a human perspective without human involvement. Researchers argue that 'human-in-the-loop methods have been widely employed to enhance and/or evaluate explanations of machine learning models' because purely automated explanations often miss the mark [10]. Without human judgment, AI can produce technically correct but practically useless or even misleading outputs. In the perfume creation study, researchers found that human experts had to 'steer' generative AI outputs as their own interpretations evolved — the AI couldn't self-correct without human guidance [2]. The bottom line: AI can process vast amounts of data, but it lacks the contextual awareness, ethical judgment, and embodied experience that make knowledge useful in the real world.
About These Sources
This answer is built on 10 peer-reviewed studies — published from 2021 to 2026, 5 from 2024 or later, 4 in Q1 journals, collectively cited 180 times — selected as the most relevant from 12 studies that passed quality screening, drawn from 62 papers retrieved from a database of over 500 million.
Sources used in this answer
A human–AI collaboration workflow for archaeological sites detection
A deep learning model for archaeological site detection reached ~80% accuracy, but human experts were essential to interpret predictions, refine datasets, and decide which sites to survey — even inaccurate predictions were useful when interpreted by a trained archaeologist.
Organizing across cognitive asymmetry in human–AI collaboration: A study of perfume creation
In a qualitative study of perfume creation, a 'cognitive asymmetry' between humans' tacit, embodied knowledge and AI's codified knowledge created a representational gap that required 'representational integration' — allocating tasks, converting knowledge forms, and steering AI outputs as problem-solving evolved.
The effects of the COVID-19 pandemic for artificial intelligence practitioners: the decrease in tacit knowledge sharing
Based on 57 interviews before and during COVID-19, remote work reduced tacit knowledge sharing among AI practitioners, leading to decreased diversity in AI project outputs, even though practitioners adapted well to routine tasks.
Artificial intelligence-based human-centric decision support framework: an application to predictive maintenance in asset management under pandemic environments
An AI-based human-centric decision support framework for predictive maintenance achieved 82% accuracy in a real-world case study by incorporating a human-in-the-loop mechanism that exploited expert tacit knowledge for final decisions.
Considerations on Human-AI Collaboration in Knowledge Work – Recruitment Experts’ Needs and Expectations
In interviews with 15 recruitment experts, AI was seen as a complementary information source that augmented — not replaced — human expertise; experts would evaluate AI outputs using their tacit knowledge and situational judgment.
Integration of AI in Capturing Tacit Knowledge of Employees Leading to Innovation in Organizational Learning: A Literature Review
A literature review concluded that 'humans remain irreplaceable in capturing tacit knowledge' and that human-AI collaboration is essential for embedding continuous learning and innovation in organizations.
Human–AI Collaboration in the Modernization of COBOL-Based Legacy Systems: The Case of the Department of Government Efficiency (DOGE)
In modernizing COBOL-based legacy systems, AI supported interoperability and automation but introduced risks around workforce disruption and knowledge retention; a hybrid human-AI model was deemed essential for managing tacit knowledge transfer.
Employee–AI collaboration empowers mentor networks to enhance employee creativity: a knowledge-management perspective
A survey of 523 employees found that mentor network strength boosted creativity through tacit knowledge acquisition, and that employee-AI collaboration strengthened this effect — AI amplified the benefits of human mentorship.
Pipe Corrosion Inspection System based on Human-in-the-Loop Machine Learning
An AI-based pipe corrosion inspection system reached 70% accuracy after converting expert tacit knowledge into explicit rules; the human-in-the-loop architecture retrained the model using manual results, with human experts supporting the AI during development.
The Role of Human Knowledge in Explainable AI
A literature review on explainable AI found that human-in-the-loop methods are widely used to collect human knowledge for improving and evaluating AI explanations, because automated explanations alone may not be understandable from a human perspective.