Metarationality

Working Definition

Metarationality is the skill of working not merely inside a formal system, but above and around it: judging which system (formal logic, statistics, an economic model, an engineering procedure) fits a given situation, how to adapt it to concrete realities, when to repair it, and when to set it aside altogether. It rests on the premise that no formal system contains within itself the rules for its own application, which means that the decision to deploy a model is always of a different kind than reasoning conducted inside its boundaries. In its technical formulation (David Chapman), it describes the practical craft of seasoned experts who treat formalisms as bounded tools rather than mirrors of reality; in its developmental formulation (Robert Kegan), it marks a stage of adult development in which a person stops identifying with any single belief system and can hold several frameworks in mind at once as objects of reflection. The concept denotes neither “being even more ruthlessly rational” nor abandoning rationality for intuition; it names a shift in the vantage point from which formal rationality is used and supervised.

Genesis and Global Lineage

The term in its current sense was popularized by David Chapman, a computer scientist trained in the MIT artificial intelligence lab of the 1980s who, after the breakdown of classical symbolic planning in AI (alongside Philip Agre), turned to the philosophy of practice. Since around 2013 in essays and since 2017 in his unfinished online book In the Cells of the Eggplant, Chapman has distinguished three modes: “rationality” (systematic, explicit, formal methods), “ordinary reasonableness” (action that makes practical sense and generally works even though it cannot be formalized), and metarationality as the art of weaving both together in a concrete setting. Three intellectual lineages supply the concept’s scaffolding: Harold Garfinkel’s ethnomethodology and its extension into human-computer interaction by Lucy Suchman (who showed that plans serve as resources for situated action rather than programs governing it), Martin Heidegger’s phenomenology of everyday coping as interpreted by Hubert Dreyfus, and Robert Kegan’s constructive-developmental psychology, whose “fifth order of consciousness” from In Over Our Heads (1994) Chapman explicitly reworks into what he calls the “fluid mode.” An additional catalyst was Chapman’s ongoing debate with the online “rationalist community” centered around the forum LessWrong, which treats Bayesian probability calculus as the universal sovereign of cognition. In popular internet parlance the position is often labeled “post-rationalism,” a tag that can mislead: Chapman rejects postmodern relativism just as firmly as he rejects formalist absolutism.

Central European Lineage

We found no documented link between this concept and the intellectual history or historical experience of Central Europe.

Why Now

The concept gained a wide readership between 2018 and 2026 for three converging reasons. First, the replication crisis in the empirical sciences made its core thesis tangible: following John Ioannidis’s 2005 paper, Brian Nosek’s Open Science Collaboration replication project in Science (2015), and re-analyses by Dénes Szűcs and Ioannidis (2017), it became unmistakable that executing statistical procedures without technical error offers no protection against false findings if nobody asks whether the statistical framework fits the phenomenon under study. Second, the rise of large language models since 2022 has confronted engineers with systems that excel inside well-defined formalisms (writing code, passing standardized tests, manipulating symbolic proofs) yet routinely fail where an agent must judge whether the task itself has been properly framed; debates around “AI evaluation” and autonomous “agents” are in practice debates over whether metarational judgment can be automated. Third, as parts of the online rationalist subculture matured and entered complex organizational roles, many practitioners sought a vocabulary for the experience that the formal decision tools of LessWrong do not suffice in institutional life; in that milieu, metarationality came to function almost as a stage of intellectual biography.

Explanatory Power

Metarationality makes it possible to see that most failures of “rational” decision-making stem not from calculation errors inside a procedure, but from the unnoticed moment when the procedure was selected, bounded, or kept running past its domain of validity. First, consider an analytics team that calculates the expected value of a project with flawless precision, while the list of scenarios over which they compute comes from a meeting held a year earlier under different conditions; the failure lies upstream of the calculation, not inside the spreadsheet. Second, consider a regulatory regime built around a quantitative indicator (emissions, student test scores, hospital readmissions, or citation counts): the metric works until the measured world begins reorganizing itself around the score, at which point fixing the system requires not a slightly more complex formula, but stepping outside the indicator to recover the substantive purpose it was meant to approximate. Third, consider an experienced clinician who knows the diagnostic guidelines thoroughly yet recognizes the specific patient for whom those guidelines were written in a different clinical world: the physician performs essential epistemic labor that the protocol neither describes nor entails. Across all three cases, the crucial competence remains invisible from inside the system because a formal system can only define errors of its own internal type.

Axes of Debate

Underappreciated Angle

Discussions of metarationality almost always treat it as a property of an individual mind, and almost never as a property of institutions. Yet in real-world technical and policy work, the competence Chapman describes is distributed: no analyst selects, recalibrates, or retires a model in isolation, but rather within a web of peer reviewers, managers, compliance rules, and deadlines. Modern bureaucracies are typically designed on the assumption that metarational judgment should be unnecessary — the procedure is supposed to work regardless of who executes it — which means that when a model drifts out of alignment with reality, the people who possess metarational skill perform invisible, unrewarded repair work in the margins. Asking how to design institutions that recognize and transmit that repair work is more consequential than asking how many individuals have attained fifth-order consciousness. For artificial intelligence deployment, this institutional angle is immediately practical: much of what is called “automated model evaluation” is an attempt to encode metarational oversight inside another rigid procedure — the exact move that the concept teaches us cannot succeed on its own.