The Unwritten Curriculum: How Informal Expertise Shapes AI Research and Who Gets Left Out

Every discipline has its unwritten rules—the practical wisdom that experienced practitioners carry but rarely articulate in formal terms. In most fields, this knowledge gap is acknowledged as a structural feature of professional development, something that apprenticeship, mentorship, and time are expected to bridge. In artificial intelligence research, however, the gap between what is published and what is known has grown to a point where it constitutes a meaningful barrier to scientific progress and a significant equity problem for the broader research community.

The phenomenon has a name in philosophy and organizational theory: tacit knowledge. The concept, developed by the philosopher Michael Polanyi, refers to the kind of knowing that resists explicit codification—the knowledge embedded in skilled practice rather than in propositional statements. A master craftsperson knows how to work with a particular material in ways that cannot be fully transmitted through a written manual. An experienced clinician recognizes patterns in patient presentations that they cannot always articulate in diagnostic criteria. In AI research, this same dynamic plays out in ways that have concrete consequences for who succeeds, which institutions dominate, and how efficiently the field as a whole advances.

The Gap Between the Paper and the Practice

Anyone who has attempted to reproduce a published AI result from scratch—following the methods section carefully, implementing the described architecture, using the specified hyperparameters—has likely encountered the frustrating experience of arriving at results that fall short of what was reported. Sometimes the discrepancy is attributable to the technical factors discussed elsewhere in reproducibility literature: hardware variance, uncontrolled randomization, incomplete documentation. But often, something else is at work.

Experienced practitioners in the field speak openly, in informal settings, about the decisions that never make it into papers. The choice to use a particular learning rate warmup schedule not because theory demands it but because it has worked reliably in past projects. The recognition that a certain type of loss curve behavior during early training is a signal to restart rather than continue. The judgment about when a validation metric plateau represents genuine convergence versus a saddle point that additional training will escape. These are not arbitrary preferences. They represent genuine knowledge, accumulated through extensive experimentation. They are also almost never documented in the published record.

The result is a published literature that describes what researchers did in idealized terms, stripped of the iterative, judgment-laden process through which the reported result was actually achieved. A graduate student or independent researcher attempting to build on that work is handed a map with significant portions left blank.

Mentorship as the Primary Transmission Mechanism

In practice, tacit AI knowledge is transmitted primarily through close mentorship relationships within laboratory environments. A doctoral student working alongside an experienced researcher absorbs, through observation and iteration, a range of practical competencies that are never formally taught. They learn which hyperparameter ranges are worth exploring for a given class of problem. They develop a feel for the debugging process—the systematic intuitions that allow an experienced practitioner to identify the source of a training instability faster than any formal diagnostic procedure would permit. They acquire, through proximity, a sense of which architectural choices are likely to be sensitive to dataset characteristics and which are relatively robust.

This mode of transmission is not inherently problematic. Mentorship is a legitimate and valuable form of professional development in virtually every field. The problem is what happens when mentorship is the primary mechanism for conveying knowledge that is genuinely important to research outcomes, and when access to that mentorship is highly concentrated.

In the United States, the landscape of elite AI research is well-documented in its geographic and institutional clustering. A small number of universities—MIT, Stanford, Carnegie Mellon, Berkeley, a handful of others—and an even smaller number of corporate research laboratories account for a disproportionate share of landmark publications, benchmark-setting results, and the mentors who carry the most accumulated tacit knowledge. Researchers trained at these institutions carry an advantage that extends well beyond the formal content of their education. They have been socialized into a set of practices, intuitions, and informal norms that their peers at less prominent institutions simply do not have equivalent access to.

The Equity Implications

The concentration of tacit knowledge at elite institutions has equity implications that the AI research community has been slow to fully reckon with. Researchers from historically underrepresented groups, who are disproportionately likely to pursue graduate training at institutions outside the top tier, face not only the well-documented structural barriers to entry in technical fields but also a knowledge deficit that is invisible in any formal sense. Their training programs may cover the same curricula, assign the same papers, and require the same formal competencies. What they do not provide is equivalent access to the informal expertise that makes a meaningful difference in research outcomes.

This dynamic also concentrates institutional power in ways that compound over time. Labs that produce successful results attract more funding, more talent, and more visibility, which in turn deepens the reservoir of tacit knowledge available to the next generation of researchers trained within them. The feedback loop is self-reinforcing and largely invisible to standard metrics of research quality.

Toward More Transferable Expertise

Addressing this problem requires deliberate effort on multiple fronts. Several directions merit serious consideration from the research community and from the organizations—including federal funding agencies and professional societies—that shape its incentive structures.

Structured documentation of experimental process. Funding agencies could require grant recipients to maintain and share process documentation that goes beyond the methods section: training logs, failed experimental branches, hyperparameter search records, and annotated notes on the decisions made during development. This is not a trivial burden, but it is a feasible one, and it would begin to externalize knowledge that currently evaporates when a project concludes.

Community-maintained knowledge repositories. Initiatives modeled on clinical practice guidelines—curated, periodically updated repositories of practical guidance for specific problem domains—could serve as a vehicle for codifying the kind of working knowledge that currently circulates only informally. These would need to be maintained by active practitioners, not extracted post hoc from publications.

Mentorship programs with explicit reach beyond elite institutions. Professional organizations in the AI field, including national and international research associations, have an opportunity to create structured mentorship programs that connect researchers at under-resourced institutions with experienced practitioners who can provide the informal guidance that geography and institutional prestige currently gatekeep.

Incentivizing honest failure reporting. Much tacit knowledge is knowledge about what does not work—the experimental dead ends and debugging lessons that are systematically omitted from published accounts because they do not serve the narrative of a successful result. Creating venues and incentives for sharing this negative knowledge, including dedicated tracks at major conferences, would help redistribute it more equitably.

A Field That Knows More Than It Shares

The AI research community is, collectively, far more knowledgeable than its published literature suggests. The gap between what practitioners know and what they write down is not a minor inefficiency. It is a structural feature of how the field operates, with real consequences for scientific progress, institutional equity, and the accessibility of AI capability to researchers who lack the accident of affiliation with the right lab.

Making tacit knowledge more explicit, more documented, and more widely accessible is not a simple task. It requires changing incentive structures, investing in infrastructure, and asking researchers to spend time on activities that current reward systems do not value. But the alternative—a field in which a significant portion of what matters most is passed quietly from mentor to student within a small cluster of elite institutions—is neither scientifically efficient nor consistent with the values of an open and equitable research community.