The Better Teams Reported More Errors

A researcher studying hospital teams expected the good units to make fewer mistakes. The data said the opposite, clearly and consistently, and the obvious conclusion was that something had gone wrong with the study.

Our Grade For This Claim

Applying the scheme set out in the first article of this series.

The association is Grade A. Psychological safety's relationship with learning behaviour and performance is supported by a meta-analytic review in Personnel Psychology[3], a systematic review in Human Resource Management Review[4], and roughly twenty-five years of independent field studies across multiple industries and countries.

The causal claim is Grade C. The dominant designs in this literature are cross-sectional and correlational at team level. That a team high in psychological safety performs better is well evidenced. That raising psychological safety in a given team causes its performance to rise is a further claim, and we have not found it demonstrated to the standard the first claim meets.

Project Aristotle is Grade C in our scheme regardless of how often it is quoted, for reasons given in its own section.

Two consequences of grading it this way, and they are ours.

A manager is on firm ground treating low psychological safety as a warning indicator, because the association is robust.

A manager is on much softer ground promising that an intervention will produce a performance gain, and should say so rather than borrowing the authority of the underlying research for a claim it does not make.

A Note On Method

Everything here is verified against the primary literature or its published abstracts as at August 2026.

Our sources conflict on what the 1999 study actually examined. We set that out in its own section rather than choosing the version that reads better.

We did not obtain the full text of Edmondson (1999), the Frazier and colleagues meta-analysis, or the Newman and colleagues review. We rely on abstracts, on the journals' own summaries, and on subsequent papers that describe them.

Our sources also give materially different figures for Project Aristotle, and different date ranges. We report the divergence rather than picking one.

One statement about what the meta-analysis found regarding transformational leadership comes from a single secondary source and we flag it where it appears[10].

This article describes organisational research. It is not psychological, clinical or employment law advice, and nothing here is a substitute for professional guidance on any individual situation.

The Anomaly

The observation that generated the field.

A history of the concept records that Edmondson was working from an apparently odd finding: teams that reported better relationships also reported more errors[2].

Consider what that looks like from inside the research. The hypothesis is that well-functioning teams make fewer mistakes. The measure of team functioning is a relationship quality survey. The measure of mistakes is reported error rates. The correlation comes back positive: the better the relationships, the more errors.

Three readings are available at that moment, and this framing is ours.

The hypothesis is wrong and good relationships genuinely cause errors, perhaps through complacency.

The study is broken, and something in the measurement has gone astray.

Or the dependent variable is not what it appears to be.

The third turned out to be the answer, and it is the interesting one because it is available in a great many organisational datasets that nobody re-examines.

Reading It The Other Way Round

The resolution, and its consequences for anyone who measures anything about their own business. This section is our own analysis.

Reported errors are not errors. They are errors that somebody chose to report.

That single distinction inverts the finding. If teams with better relationships report more errors, the most plausible explanation is not that they make more. It is that in those teams, reporting one is survivable.

Three consequences follow, and they generalise well beyond hospitals.

A low error count can be the worse signal. A team reporting nothing is either flawless or silent, and silence is far more common than flawlessness.

Any metric that depends on voluntary disclosure measures two things at once: the underlying rate, and the willingness to disclose. Near-miss reports, safety incidents, quality escalations, customer complaints logged internally, and project status reporting all have this structure.

And improvement in such a metric is ambiguous. A falling internal complaint count can mean the problem is resolving or that people have stopped bothering.

Our own observation is that this is the most transferable idea in the whole psychological safety literature, and it is usually skipped in favour of the management advice. It costs nothing to apply, requires no intervention, and it tells a business owner something real about numbers they already collect.

A Conflict In The Record

Something our sources disagree about, reported rather than smoothed.

One account describes the 1999 paper as Edmondson's study of hospital nursing teams[2].

Another states that the paper studied 51 work teams in a manufacturing company, not the hospital setting that inspired the question[6].

The journal's own abstract says only that the paper presents a model of team learning and tests it in a multimethod field study[5].

These are not reconcilable from what we obtained. The second version, in which an earlier hospital study produced the anomaly and the 1999 paper tested the resulting model in a different setting, is coherent and would explain the confusion. We did not verify it and do not assert it.

Two observations, ours.

The distinction matters more than it looks. If the anomaly and the model test come from different settings and different samples, that is a stronger evidentiary position than if one study did both, because the model was tested somewhere other than where it was generated.

And the fact that a foundational study in a heavily cited literature is described inconsistently by careful secondary sources is itself the point of this series. Anyone building an argument on the hospital origin story should read the 1999 paper, which we have not.

The Definition, Precisely

The construct as its author defined it, because the popular version has drifted.

Edmondson introduced the construct of team psychological safety as a shared belief held by members of a team that the team is safe for interpersonal risk-taking[5][2], later phrased as a shared belief that people feel safe about the interpersonal risks arising in a team context[8].

Four elements of that definition do real work, and this parsing is ours.

Shared. It is not an average of individual comfort levels; it is a belief held in common. A team where two people feel safe and four do not is not moderately psychologically safe.

Belief. It is a perception about consequences, not an emotional state and not a personality trait.

Safe for interpersonal risk-taking. The risk in view is social: looking ignorant, incompetent, negative or disruptive. It is not physical safety and not job security.

And the team. The unit is the team, which is the subject of the next section.

What is conspicuously absent from the definition is any reference to comfort, agreeableness, harmony, or the absence of conflict. Every one of those appears in the popular version, and none is in the construct.

It Is A Property Of Teams

A structural point with immediate practical consequences.

Edmondson explicitly defined it as a team-level construct, and by locating it at that level and providing a measurement instrument, she made the concept empirically studiable in a way it had not been before[2].

Three consequences, ours.

It can vary enormously within one organisation. A company does not have a psychological safety level; its teams do, and they can differ sharply under the same policies, the same values statement and the same chief executive.

Which means organisation-wide interventions are aimed at the wrong altitude. A company-wide training programme addresses a variable that lives in the interaction between a specific group of people and, usually, their immediate manager.

And it means an aggregate engagement score conceals the thing you care about. Averaging a high-safety team with a low-safety one produces a middling number describing neither.

For a small business this cuts both ways. A ten-person firm may genuinely be one team, in which case the construct applies cleanly. A thirty-person firm with three departments almost certainly is not, and the owner's impression of the culture is likely formed by the team they sit closest to.

The Path Is Mediated

The finding this article is built around.

The 1999 paper's core result was that psychological safety predicted learning behaviour, and learning behaviour predicted performance, and that the relationship between safety and performance was not direct: it ran through a mediating variable[6].

The history of the concept states the same structure: Edmondson demonstrated empirically that psychological safety predicts learning behaviours and, through learning, team performance[2].

The subsequent literature treats this as established. A later review notes that team learning behaviour is known to mediate the relationship between psychological safety and performance, citing Edmondson and a series of later studies[8].

The learning behaviours themselves are specific and observable: seeking feedback, experimenting, discussing errors, and asking for help[6].

So the causal chain as the literature describes it has three links rather than two.

Psychological safety makes interpersonal risk survivable.

Which enables learning behaviours, being the specific acts of asking, admitting, testing and seeking.

Which produce performance.

Why That Single Word Changes Everything

Our own analysis, and the practical heart of this article.

A mediated relationship means the first variable does nothing on its own. It works only through the second.

Three consequences follow, and each contradicts something commonly done in the name of psychological safety.

First, safety without learning behaviour produces nothing. A team can be made to feel entirely safe and, if nobody consequently asks a difficult question or surfaces an error, the mechanism has not engaged. On this model, the comfortable team that raises nothing has psychological safety and no benefit from it.

Second, the observable outcome is behavioural, not attitudinal. The check on whether an intervention worked is not whether people report feeling safer. It is whether the specific behaviours increased: more questions asked, more errors surfaced, more requests for help, more small experiments run.

Third, and most usefully, the middle link can be measured directly. A business that finds surveying attitudes uncomfortable or unreliable can skip it and count the behaviours instead, which are far harder to fake and require no instrument.

We would put the practical instruction plainly. If you want to know whether your team is psychologically safe, do not primarily ask them. Look at whether anyone brought you bad news this month that you did not already know.

The Wrong Target

Where implementation goes wrong, in our assessment.

Most organisational efforts under this banner target the first link and stop. Values statements, civility training, listening skills workshops, open-door declarations.

None of those is objectionable, and none of them is the mechanism. On the model above they are attempts to move an antecedent, with success measured by an attitudinal survey, while the variable that actually connects to performance goes unmeasured.

Three failure modes we would expect from that design.

The survey improves and nothing else does, because people have learned what the programme wants to hear.

The intervention increases comfort and decreases candour, if the practical message received is that disagreement is now a civility problem. Nothing in the construct suggests this is inevitable, but nothing prevents it either, and it is the specific risk of implementing safety as niceness.

And the organisation concludes the concept does not work when the behavioural change never arrives, having never targeted the behaviour.

We flag that these are our own predictions from the model's structure, not findings we obtained from the literature.

Safety Is Not Comfort

The distinction stated directly, because it is the most consequential misreading.

A summary of the research puts it in one line: the mechanism is learning, not comfort, and it is the learning behaviours that safety enables which drive performance[6].

Three implications, ours.

The behaviours the construct is about, being admitting mistakes, asking questions and raising problems[6], are individually uncomfortable. A team where these happen routinely is not a team where nobody is ever uncomfortable. It is one where discomfort is survivable.

Which means a demanding team can be psychologically safe and a pleasant team can fail to be. High standards and safety are orthogonal, not opposed, and a workplace can be low on both.

And the diagnostic question is not do people get along. It is what happens to the person who says the difficult thing.

Our own observation is that owner-managed businesses are especially prone to the pleasant-but-unsafe combination. Small firms are frequently warm places with genuine mutual regard, in which nobody will tell the owner their pricing is wrong, because the relationship is exactly what makes it awkward.

What The Meta-Analysis Adds

The evidentiary backbone.

Frazier, Fainshmidt, Klinger, Pezeshkan and Vracheva published a meta-analytic review and extension in Personnel Psychology in 2017[3]. A description of its scope records that it aimed to determine the effectiveness of antecedents, examine influence on task performance, examine the nomological network of the research, and test the homology assumption[9].

Alongside it, Newman, Donohue and Eva published a systematic review in Human Resource Management Review in the same year[4].

Two observations, ours.

The presence of both a meta-analysis and a systematic review in mainstream journals in the same year indicates a literature substantial enough to be worth synthesising, which is itself evidence of a kind. Fields resting on one striking study do not generate these.

And the reference to testing the homology assumption is a sign of methodological seriousness. That is the question of whether a construct behaves the same way at individual and group levels, and it is precisely the sort of thing a field checks when it is being careful rather than enthusiastic.

One further claim appears in a secondary source: that the 2017 meta-analysis found transformational leadership's effect on team performance largely runs through psychological safety as a mediator[10]. We obtained this from a single secondary source and did not verify it against the meta-analysis, and we report it as that source's characterisation rather than as an established finding.

What Leaders Actually Do

The most operationally useful strand, from a study of health care teams.

Nembhard and Edmondson introduced the construct of leader inclusiveness, being the actions by which leaders encourage psychological safety, in a study of the effects of leader inclusiveness and professional status on psychological safety and improvement efforts in health care teams[11].

Two observations, ours.

The framing is behavioural rather than dispositional. It concerns actions a leader takes, not qualities a leader has, which makes it a management variable rather than a personality one.

And the study pairs it with professional status, which is the subject of the next section and the part most relevant to any business with mixed seniority in a room.

We did not obtain the full text of this study and do not report its specific findings on which leader actions had which effects. What we take from it is the framing: psychological safety has identifiable leader antecedents, and they are things a person does.

The Status Problem

An extension of the construct that matters more in small businesses than in large ones. This section is our own analysis, built on the Nembhard and Edmondson framing[11].

If psychological safety concerns the survivability of interpersonal risk, then the cost of that risk is not equal across a team. The junior person, the newest hire, the contractor and the person whose professional status is lower in the room are each risking more by speaking.

Three consequences.

A team can be genuinely safe for its senior members and unsafe for its junior ones, and since the senior members are the ones asked how the culture is, the assessment will be wrong in a predictable direction.

The people best positioned to notice certain problems, being those closest to the work, are frequently those with least status to raise them.

And in an owner-managed business the status gradient is unusually steep, because the person at the top is not merely senior but owns the enterprise, sets the pay, and is frequently in the room.

We are not aware of, and did not find, research quantifying this gradient in small firms specifically. We offer it as an implication of the construct rather than as a finding.

The Evidence Everybody Cites

Google's Project Aristotle, described accurately.

Project Aristotle was Google's internal research effort examining what made its teams effective. Accounts state that it identified psychological safety as the most significant predictor of team effectiveness[2], and that successful Google teams shared five elements: psychological safety, dependability, structure and clarity, meaning, and impact of work[8].

It reached a general audience through Charles Duhigg's article in the New York Times Magazine on 25 February 2016[7].

Our sources give materially different specifics, and we report all of them rather than choosing.

On scale: one source says hundreds of teams[2]; another says the project explored over 250 team-level variables[8]; another says Google studied 180 teams over two years[7].

On dates: one gives 2012 to 2016[2], another 2012 to 2015[7].

These are not trivial discrepancies. A reader citing a specific number should obtain it from Google's own published material, which we did not.

And Why We Grade It Separately

The epistemic point, and it is the most important thing in this article for anyone who quotes this study.

Project Aristotle is internal corporate research, conducted by a company on its own employees, and popularised through a magazine article[7].

It is not, on anything we located, a peer-reviewed publication.

Three consequences, ours, and none of them is that it is wrong.

It has not been through peer review, so its methodology has not been examined by independent specialists in the way the Edmondson literature has.

Its sample is one company, in one industry, in one country, at one point in time. Google's teams are not a random sample of workplaces.

And a company publicising a finding about its own culture has an interest in the result that an academic team does not, however carefully the work was done.

Under the grading scheme in the first article of this series, that places it at Grade C: interesting, widely known, not a basis for advice on its own.

The reason this matters is a matter of practical logic rather than pedantry. Project Aristotle is frequently deployed as the independent confirmation of Edmondson's academic work, and the phrase does a lot of persuasive work. The honest description is that a large company's internal analysis pointed in the same direction as a peer-reviewed literature. That is genuinely encouraging. It is not the same thing as replication, and the academic literature is the stronger evidence of the two despite being far less quoted.

The Causal Gap Nobody Mentions

The limitation we think this field states least clearly. This section is our own analysis.

Almost everything above establishes association. Teams high in psychological safety show more learning behaviour and better performance.

The management inference is causal and directional: raise safety, and performance follows.

At least three alternative structures are consistent with the same correlations, and we have not found them ruled out.

Reverse causation. Teams that are performing well may find candour cheaper. Success buys the slack in which admitting error is survivable.

A common cause. A competent manager may independently produce both higher safety and higher performance, in which case safety is a symptom rather than a lever.

Selection. People who do not fit a team leave it, and the survivors report a shared belief that may reflect who remained rather than what the team does.

Two things follow, ours.

This does not make the construct useless. As a diagnostic, association is enough: a team where nobody raises problems is worth investigating whichever way the causation runs.

But it does mean that a promised return on a psychological safety programme is not supported by the evidence in the way it is usually presented, and anyone selling one should be asked what their causal claim rests on.

Measuring It Without Fooling Yourself

Practical guidance, derived by us from the structure above rather than from a source.

The literature provides an instrument. Edmondson's 1999 psychological safety subscale is described as a seven-item survey[9], and it is the measure the field has largely used.

For a small business we would suggest a different starting point, for three reasons.

A survey in a ten-person firm is not anonymous whatever it says on the form, and people know it.

The mediating variable is behavioural and countable, which means it can be observed without asking anyone anything.

And behaviour is harder to perform for an audience than an attitude rating is.

Four observable proxies, ours, drawn from the learning behaviours the literature names[6].

Error surfacing. When did someone last bring you a mistake you had not already discovered? If the answer is never, that is data.

Help-seeking. Do people ask before they are stuck, or after?

Question asking. In a meeting where you explained something, did anyone ask you to explain it again?

Dissent. When did someone last tell you a decision of yours was wrong, before it was made rather than after?

None of these requires an instrument, and all four are things an owner can answer honestly in a minute. If all four answers are uncomfortable, the attitudinal survey is unlikely to tell you anything you do not already know.

The Small Business Version

Where the construct behaves differently at small scale, in our assessment.

Three features of an owner-managed firm change the picture.

The team and the company are the same thing. The team-level framing that complicates large-organisation measurement is a simplification here, because there is frequently only one team.

The status gradient is steeper. The senior person owns the business, which compounds every ordinary reason not to contradict a manager.

And exit is more visible. In a large organisation a person who does not feel safe can move teams; in a small one, leaving means leaving, which raises the cost of every interpersonal risk.

Two consequences.

The owner's own behaviour is a far larger share of the variable than any manager's would be in a large firm, because there is nobody else to dilute it.

And the single most informative signal is what happens the first time somebody disagrees with the owner in front of others. Everyone present updates on that, and the update persists.

We offer these as reasoned extensions of the construct to a setting the research we located does not specifically address.

Contractors, Advisers And The Outside Voice

A case this publication has a particular view of, stated as our own analysis.

A contractor, a fractional executive or an external adviser occupies an unusual position on the status gradient. They frequently have lower formal standing and lower personal cost of speaking at the same time, because their livelihood does not depend on this one relationship in the way an employee's does.

Three consequences.

An outside voice can sometimes say the thing the team cannot, and that is a substantial part of what such a person is actually being paid for, whatever the engagement letter says.

But an adviser who wants the engagement renewed faces exactly the incentive that suppresses candour internally, and should notice it in themselves.

And a business that has hired an outside adviser and finds itself receiving only agreement has learned something about its own environment, not about whether it is doing well.

We include this because it applies to us. A firm in our position is subject to the same pressure, and a reader is entitled to weigh that when reading anything an adviser tells them.

What To Do

Read low error counts as ambiguous. Any metric depending on voluntary disclosure measures the rate and the willingness to disclose at once, and silence is more common than flawlessness.

Target the behaviour, not the feeling. The path from safety to performance runs through learning behaviours, and safety without those behaviours is not doing the work.

Count four things. Error surfacing, help-seeking, question asking, and dissent before a decision rather than after.

Do not equate safety with comfort. The behaviours the construct concerns are individually uncomfortable; the point is that they are survivable.

Assume your own read is biased upward. Safety varies by status, and the senior people who get asked about the culture are the ones facing the lowest risk.

Watch what happens the first time someone disagrees with you publicly. In a small firm that single event sets the level for a long time.

Treat it as a team-level variable. It can differ sharply between groups under identical policies, and an aggregate score describes none of them.

Cite Project Aristotle carefully, if at all. It is internal corporate research popularised by a magazine article, our sources disagree on its basic figures, and the peer-reviewed literature behind it is stronger evidence.

Be sceptical of promised returns. The association is well evidenced; the causal claim that raising safety raises performance is a further step we did not find demonstrated to the same standard.

Ask your outside advisers what they are not telling you. Their incentive to keep the engagement is the same incentive that silences employees.

The Limits Of This Analysis

Several caveats matter. This article describes organisational research and is not psychological, clinical or employment law advice; nothing here substitutes for professional guidance on an individual situation. Everything is verified to August 2026. We did not obtain the full text of Edmondson (1999), the Frazier and colleagues meta-analysis, the Newman and colleagues systematic review, or the Nembhard and Edmondson study, and rely on abstracts, journal summaries and subsequent descriptions. Our sources conflict on what the 1999 study examined, one describing hospital nursing teams and another 51 teams in a manufacturing company; we report the conflict and do not resolve it. Our sources give materially different figures for Project Aristotle, variously hundreds of teams, over 250 team-level variables, and 180 teams over two years, with date ranges of 2012 to 2016 and 2012 to 2015; we did not obtain Google's own published material. The claim that the 2017 meta-analysis found transformational leadership's effect running largely through psychological safety comes from a single secondary source we did not verify. We did not obtain the specific effect sizes reported by the meta-analysis and state none. The causal direction is not established by the designs we located, and we set out three alternative structures consistent with the same correlations. Our three predicted failure modes of implementation, the status-gradient analysis, the four behavioural proxies, the small business extension and the contractor discussion are our own reasoning from the construct, not findings we obtained. We did not research the employment law implications of any measurement practice described here, which vary by jurisdiction.

Related Reading

Behavioural Finance Thirty-Six Percent The grading scheme applied here, and why a correctly cited study can still be unreliable. No. 3 The Jam Study Is Not A Rule One famous experiment on 242 shoppers produced the advice to cut your product range. A meta-analysis of fifty studies put the average effect at roughly... Behavioural Finance Financial Dysmorphia The owner-level version of not seeing what the numbers are saying.

Frequently Asked Questions

What is psychological safety, precisely?

Edmondson defined it as a shared belief held by members of a team that the team is safe for interpersonal risk-taking. The risk is social: looking ignorant, incompetent, negative or disruptive. Comfort, harmony and absence of conflict appear nowhere in the definition, though they dominate the popular version.

Why did the better teams report more errors?

Because reported errors are not errors; they are errors somebody chose to report. In teams where reporting was survivable, more got reported. That inversion generalises to any metric depending on voluntary disclosure, which includes near-misses, internal complaints and project status reporting.

Does making people comfortable improve performance?

Not on this model. The relationship is mediated: safety enables learning behaviours such as seeking feedback, experimenting, discussing errors and asking for help, and those behaviours drive performance. A comfortable team in which nobody raises anything has not engaged the mechanism.

How do we measure it in a ten-person business?

We would not start with a survey, which is not anonymous at that size whatever the form says. Count behaviours instead: when did someone last bring you a mistake you had not found, ask for help before being stuck, ask you to re-explain something, or tell you a decision was wrong before you made it.

Is the Google study good evidence?

It is interesting and it is not peer-reviewed. Project Aristotle is internal corporate research at one company, popularised by a magazine article, and our sources disagree on its basic figures and dates. The academic literature behind the concept is considerably stronger evidence and considerably less quoted.

Will raising psychological safety raise our performance?

That is a causal claim, and the designs we located establish association rather than causation. Reverse causation, a common cause such as a competent manager, and selection effects are all consistent with the same correlations. As a diagnostic the construct is sound; as a promised return it is oversold.

References

1. Edmondson, A. C. (1999). Psychological Safety and Learning Behavior in Work Teams. Administrative Science Quarterly, 44(2), 350–383, introducing the construct of team psychological safety as a shared belief held by members of a team that the team is safe for interpersonal risk-taking, and establishing connections with team efficacy, team performance and individual learning behaviours. Note: we did not obtain the full text and rely on the published abstract and on subsequent descriptions. journals.sagepub.com
2. Psych Safety. The History of Psychological Safety, on Edmondson working from an apparently odd finding that teams reporting better relationships also reported more errors; on her demonstrating empirically that psychological safety predicts learning behaviours and, through learning, team performance; on her explicitly defining it as a team-level construct, being a shared belief that the team is safe for interpersonal risk-taking; on locating it at team level and providing a measurement instrument having made the concept empirically studiable; and on Google's Project Aristotle (2012–2016) identifying psychological safety as the most significant predictor of team effectiveness across hundreds of teams. Note: a specialist practitioner resource, not peer-reviewed; its description of the 1999 study's setting conflicts with reference 6. psychsafety.com
3. Frazier, M. L., Fainshmidt, S., Klinger, R. L., Pezeshkan, A., & Vracheva, V. (2017). Psychological safety: A meta-analytic review and extension. Personnel Psychology, 70(1), 113–165. Note: cited as the field's principal meta-analysis; we did not obtain it and report no effect sizes from it. psychiatry.ucsd.edu
4. Newman, A., Donohue, R., & Eva, N. (2017). Psychological safety: A systematic review of the literature. Human Resource Management Review, 27(3), 521–535. Note: cited as the field's principal systematic review; we did not obtain it. si-labs.com
5. Administrative Science Quarterly, publisher abstract of Edmondson (1999), stating that the paper presents a model of team learning and tests it in a multimethod field study, and introduces the construct of team psychological safety as a shared belief held by members of a team. Note: the journal's own abstract, which does not specify the study setting. journals.sagepub.com
6. Alfred. Psychological Safety: Amy Edmondson's Research on What Actually Makes Teams Perform, on the 1999 paper having studied 51 work teams in a manufacturing company rather than the hospital setting that inspired the question; on the core finding that psychological safety predicted learning behaviour and learning behaviour predicted performance, with the relationship between safety and performance running through a mediating variable rather than directly; on the learning behaviours being seeking feedback, experimenting, discussing errors and asking for help; and on the mechanism being learning rather than comfort. Note: a commercial publication, not peer-reviewed; its description of the study setting conflicts with reference 2 and we do not resolve the conflict. get-alfred.ai
7. SI Labs. Psychological Safety: Definition, Research, and Practical Guide for Teams, citing Duhigg, C., What Google Learned from Its Quest to Build the Perfect Team, New York Times Magazine, 25 February 2016, based on Google's internal research Project Aristotle (2012–2015); and stating that Google studied 180 teams over two years and found psychological safety by far the strongest predictor of team performance. Note: a commercial publication citing a newspaper article reporting internal corporate research. Its figures and date range conflict with references 2 and 8, and we did not obtain Google's own published material. si-labs.com
8. Frontiers in Psychology. (2020). How Psychological Safety Affects Team Performance: Mediating Role of Efficacy and Learning Behavior, on team learning behaviour being known to mediate the relationship between psychological safety and performance, citing Edmondson (1999) and a series of subsequent studies; on team psychological safety being a shared belief that people feel safe about the interpersonal risks arising in a team context; and on Project Aristotle having explored over 250 team-level variables and found five elements common to successful Google teams, being psychological safety, dependability, structure and clarity, meaning, and impact of work. Note: a peer-reviewed article, cited for its summary of the mediation literature and for its account of Project Aristotle's scope. frontiersin.org
9. University of California San Diego, Department of Psychiatry, research resource on psychological safety, on Edmondson (1999) introducing the construct and presenting the commonly used psychological safety measure as a seven-item survey subscale; and on the Frazier and colleagues meta-analysis aiming to determine the effectiveness of antecedents, examine influence on task performance, examine the nomological network of the research, and test the homology assumption. Note: an academic institution's research resource page. psychiatry.ucsd.edu
10. Chou, Y. Psychological Safety: Behavioral Designer's Guide, stating that the 2017 Frazier and colleagues meta-analysis explicitly found that transformational leadership's effect on team performance largely runs through psychological safety as a mediator. Note: a single secondary source, not peer-reviewed; we did not verify this characterisation against the meta-analysis itself and do not present it as an established finding. yukaichou.com
11. Nembhard, I. M., & Edmondson, A. C. (2006). Making it safe: The effects of leader inclusiveness and professional status on psychological safety and improvement efforts in health care teams. Journal of Organizational Behavior, 27(7), 941–966, introducing the construct of leader inclusiveness, being the actions by which leaders encourage psychological safety. Note: we did not obtain the full text and report none of its specific findings. psychiatry.ucsd.edu

This article describes organisational research and is not psychological, clinical or employment law advice. Full texts of the primary studies were not obtained. Sources conflict on the setting of the 1999 study and on the basic figures and dates of Project Aristotle; both conflicts are reported and neither is resolved. The causal direction is not established by the designs located. Several sources are commercial or practitioner publications rather than peer-reviewed, and are identified as such.