Amy C. Edmondson

A detected-error paradox became a measurement problem

Harvard Business School's institutional biography identifies Amy C. Edmondson as its Novartis Professor of Leadership and Management, records her Harvard degrees in engineering and design, psychology, and organizational behavior, and says that she worked as Buckminster Fuller's chief engineer in the early 1980s before publishing A Fuller Explanation in 1987. Those facts establish the professional chronology, not a causal lineage from Fuller's geometry to her later organizational research.¹

The formative research problem came from drug-administration errors in hospital patient-care groups. Edmondson's 1996 article reports differences among groups in both detected-error frequency and the likelihood that errors would be found and learned from. In a later first-person account, she recalls predicting that stronger teamwork would accompany fewer errors, finding the opposite association, and commissioning a researcher who was blind to the survey results and her new hypothesis to observe the units.²³

The follow-up observations supported Edmondson's interpretation that openness changed error detection: units in which people spoke more openly about mistakes also had higher detected-error rates. They did not reveal the true underlying error rate, prove that the more open units made fewer errors, or turn a higher reported count into a general sign of safety. A later independent review likewise finds that reported patient-safety problems can indicate either disclosure or unsafe care, depending on the outcome and setting.³⁴

That distinction makes error reporting a problem of organizational intelligence, executive attention and sensing, decision-making under bounded information, and organizational ignorance. The connection is analytical rather than a documented influence claim: a count reflects what happened, what observers could detect, and what people were willing to disclose; the cited studies do not identify those components separately.²⁴

Edmondson made an older idea measurable at team level

“Psychological safety” predates Edmondson's research. An independent meta-analytic review traces organizational uses to Schein and Bennis in 1965 and Kahn's individual-level formulation in 1990. It identifies Edmondson's 1999 contribution as the group-level definition of team psychological safety: a shared belief that a team is safe for interpersonal risk taking.⁵

The 1999 study combined surveys and qualitative fieldwork across 51 teams in one manufacturing company. Its seven-item scale asked, among other things, whether mistakes were held against members, difficult issues could be raised, help could be requested, and distinctive skills were valued. Psychological safety was associated with self-reported and observer-rated learning behavior, and the article modeled learning behavior as a mediator between safety and performance.⁶⁷

The article itself limits the strength of that account. Its cross-sectional design could not demonstrate causality or the proposed feedback cycle, and tests of leadership and contextual antecedents drew variables from the same survey. Edmondson therefore described those antecedent findings as tentative and the design as a snapshot of a dynamic process.⁸

In The Fearless Organization, Edmondson later distinguishes psychological safety from niceness, comfort, dyadic trust, and relaxed performance standards. The book presents safety and standards as separate dimensions and treats candor plus demanding goals as a managerial ideal. That is Edmondson's participant-authored framework; the quadrant labels are not themselves experimental findings.⁹

This framing places the construct beside culture, informal organization, trust, and voice and cooperation, incentives, and equilibrium. It does not make a climate score a measure of kindness, legal protection, or effective remedy. The original instrument measures beliefs about expected interpersonal consequences inside a team.⁶⁹

Independent evidence supports associations, with important boundaries

Frazier and colleagues independently synthesized 136 samples representing more than 22,000 individuals and nearly 5,000 groups. Their corrected correlations linked psychological safety with learning behavior, information sharing, voice, and task performance at the individual level; the smaller group-level evidence also showed positive associations with learning, information sharing, and task performance. This is broad corroboration of a network of associations, not a replication of every causal step in Edmondson's model.¹⁰

The same meta-analysis supplies a material warning. Only 13 percent of extracted correlations used different sources for the predictor and outcome, effect sizes were 27 percent higher when measures came from the same source at the same time, and most included studies were cross-sectional. The authors found that time-separated, multi-source relationships remained significant, but said the literature could not support strong causal assertions and detected possible publication bias for prominent outcomes.¹¹

Task context also matters. Sanner and Bunderson's independent meta-analysis found stronger safety–learning and safety–performance associations in knowledge-intensive settings involving complexity, creativity, and sensemaking; they concluded that psychological safety may be insufficient to stimulate learning when the task environment does not require learning.¹²

More recent evidence challenges a simple “more is always better” rule. Across five multi-source studies, Eldor, Hodor, and Cappelli reported that moderate psychological-safety climates accompanied better routine in-role performance, while very high levels accompanied declining routine performance; collective accountability buffered the decline. Their result concerns routine tasks and their operationalization of climate and performance, not candor in every kind of work.¹³

A different boundary concerns purpose. In a study of 126 teams, Pearsall and Ellis found that psychological safety strengthened the association between a utilitarian team orientation and unethical team decisions and behavior. One study cannot establish a general “dark side,” but it shows why interpersonal openness cannot substitute for ethical constraints on what a group is trying to do.¹⁴

Newman, Donohue, and Eva's independent systematic review covered 83 published studies available by the end of 2015. It documented extensive development while calling for better measurement, alternative methods, cultural research, multi-level analysis, and investigation of possible negative effects. Its role is to map a field and its gaps; it does not estimate the effect of a particular leadership intervention.¹⁵

Leadership and hierarchy remain relevant without becoming a guaranteed recipe. Nembhard and Edmondson's survey of 23 neonatal intensive-care units associated professional status and leader inclusiveness with psychological safety and quality-improvement engagement. An independent review of 36 healthcare studies also identified status, hierarchy, inclusiveness, support, familiarity, and learning orientation among recurring enablers, but synthesized heterogeneous observational evidence rather than trials of one intervention.¹⁶¹⁷

Nor does speaking make voice consequential. Morrison, See, and Pan found that a felt lack of power contributed to silence and that perceived openness in the person being addressed mitigated that relationship. Burris found in one field study and two experiments that managers rated challenging speakers as worse performers and endorsed their ideas less than supportive speakers, with perceived threat and loyalty helping explain the difference. Those studies separate an invitation to speak from authority, uptake, and protection after speaking.¹⁸¹⁹

Teaming and intelligent failure are broader author frameworks

Teaming shifts the unit of attention from a stable team to the activity of coordinating in fluid, flexible collaborations. Edmondson's 2012 book connects that activity to reflection, experimentation, knowledge sharing, and psychological safety; its publisher describes cases across healthcare, industry, government, and design. The book is a participant-authored synthesis, not evidence that every temporary group can replace familiarity with process.²⁰

The conceptual move links psychological safety with coordination, communication, and common understanding: people who have not developed a long shared history still need ways to expose uncertainty and integrate specialized knowledge. Whether a given setting has the roles, time, resources, and authority to do so remains an empirical question, not an implication of the word “teaming.”²⁰¹²

Edmondson's 2011 failure typology distinguishes preventable failures in predictable operations, complexity-related failures produced by interacting conditions, and “intelligent” failures in new territory. Her management article argues that an informative frontier failure should occur quickly and at small scale. This is an author-defined classification for learning, not evidence that calling an event intelligent makes its costs acceptable.²¹

Viewed through benefit for all life, the missing questions are authorization and burden: who may expose a worker, patient, customer, community, nonhuman animal, or ecosystem to an experiment, and who can stop it? The failure typology bounds novelty and scale but does not measure consent, remedy, distributional effects, or ecological consequences. Those dimensions must be evaluated separately before the label guides action.²¹

The evidence does not support an impact shortcut

The strongest recent check on consequential outcomes is a 2025 systematic and narrative review of psychological safety and objective patient-safety measures. Only nine studies met its criteria; five found a significant relationship, the outcomes were heterogeneous, 88.4 percent of participants were nurses, and eight of nine studies were conducted in the United States. The reviewers concluded that no clear relationship or causal direction could be extracted.²²

The source base used here contains employee surveys and interviews but no worker-controlled or union source, and it does not establish whether the scale works equivalently across employment arrangements or supplies legal protection, stop-work authority, job security, or a route to remedy. The healthcare evidence also centers clinicians more than patients and leaves community and ecological effects outside its measured outcomes. These are affected-party evidence gaps, not evidence that such effects are absent.²³

No cited study treats reading, teaching, or adopting Edmondson's books as a causal treatment. No causal impact on workers, patients, organizations, communities, nonhuman life, or ecosystems is therefore claimed for the works themselves. The research concerns measured climates, behaviors, and outcomes in particular settings; moving from that evidence to a branded program requires a new evaluation.¹¹²²

A responsible use is diagnostic and conditional: interpret reporting counts as visibility as well as incidence, pair invitations to speak with authority and follow-through, retain ethical and performance constraints, and test effects in the actual task and population. That preserves Edmondson's contribution to learning, quality, and reliability without turning a team-perception measure into a universal theory of safety or power.²⁴