Science

What Is Psychological Safety? A Definition for 2026 Teams.

Edmondson defined it in 1999. Google confirmed it in 2015. The signal is behavior, not opinion — and the gap between the two is exactly where attrition hides.

Bytru.place
Published2 Jun 2026
Read11 min
AudienceManager · HR Leader

In 1999, Amy Edmondson set out to prove that the best hospital teams made fewer medication errors. The data said the opposite — the best teams reported more, not fewer. The behavior was the signal, not the count, and that single finding became the foundation of psychological safety as a discipline. This article defines the term, traces the research from Edmondson through Google's Project Aristotle in 2015, and explains why most modern attempts to measure it — annual surveys, eNPS, engagement scores — fail at the thing they claim to capture.

What is psychological safety?

Psychological safety is the shared belief, held by members of a team, that the team is safe for interpersonal risk-taking. The term was defined by Amy Edmondson in her 1999 paper Psychological Safety and Learning Behavior in Work Teams, published in Administrative Science Quarterly. It is not a feeling of comfort or harmony. It is the structural condition that lets a colleague say "I think we are wrong about this," "I made a mistake," or "I do not understand," without expecting to be punished, ridiculed, or quietly sidelined.

The definition is interpersonal, not psychological in the clinical sense. It is about what a person can safely do in front of others, not what they feel privately.

Edmondson's full definition: "A shared belief held by members of a team that the team is safe for interpersonal risk-taking… a sense of confidence that the team will not embarrass, reject or punish someone for speaking up."

That precision matters. Psychological safety is not the same as trust, comfort, kindness, or harmony. It is the expectation that taking a small social risk — asking a basic question, admitting an error, raising a concern about a senior person's plan — will not cost a person their standing on the team.

Most popular treatments of the term collapse it into something broader: "psychological wellbeing," "feeling comfortable at work," general employee mood. Edmondson's original construct is narrower and more useful. It is the social cost of interpersonal risk, observed at the level of the team.

Why does psychological safety matter?

Because the team's ability to learn depends on it, and in 2015 Google demonstrated that the team's ability to perform does too.

Google's Project Aristotle was a multi-year internal study of 180 teams. The researchers expected to find patterns in seniority, tenure, gender mix, or sheer intelligence. They did not.

The single biggest differentiator was psychological safety.

"Psychological safety was far and away the most important of the five key dynamics we found — it's the underpinning of the other four." Google re:Work · Project Aristotle · 2015

Restated, the finding is sharper still: psychological safety is the single biggest differentiator between teams that perform and teams that don't. That restatement is the line we work from at tru.place. It is not Google's wording; it is the operational form of the same finding.

The Project Aristotle report was published on Google re:Work in 2015 and remains the most widely cited corporate replication of Edmondson's academic work. It also reframed the conversation. Psychological safety stopped being a soft HR topic and became a leading indicator for delivery, innovation, and retention.

When a team is psychologically safe, problems surface faster. Mistakes are caught earlier. Disagreement is processed in the meeting, not in the side conversations after it. When it is not, the costs are observable: delivery slips, quiet quitting, attrition that arrives "out of nowhere," and the slow erosion of judgment.

What did Amy Edmondson's 1999 research actually show?

Edmondson studied hospital teams to test a clean hypothesis: the better the team, the fewer medication errors it would report. The opposite was true.

The best teams reported more medication errors, not because they made more, but because the working environment made it safe to say "this happened, here is what we missed." Lower-performing teams reported fewer errors because the cost of reporting was higher. The errors still occurred. They simply went unreported, and therefore unlearned from.

"The best teams don't make fewer mistakes — they report more of them." Amy Edmondson · Administrative Science Quarterly · 1999

The finding produced a measurable construct. Edmondson built the seven-item Psychological Safety Scale, which remains the most widely used self-report measure in organisational research. The scale asks team members to rate statements such as "If you make a mistake on this team, it is often held against you" and "Members of this team are able to bring up problems and tough issues."

Two methodological notes matter for anyone using the scale today. First, it is a self-report, subject to social desirability bias and to the same psychological pressures it tries to measure. Second, it was designed for team-level analysis, not for individual diagnostics. The unit of measurement is the team, not the person.

Edmondson has since extended the work in The Fearless Organization (2018) and continued empirical research through the 2020s. The core claim, that interpersonal risk-taking is the substrate on which learning, innovation, and performance grow, has held up across more than two decades of replication.

How is psychological safety measured today?

The honest answer: poorly, in most companies. The standard methods — annual engagement surveys, pulse surveys, eNPS — share three structural weaknesses.

They measure opinion, not behavior. A respondent's answer to "I feel safe to speak up on my team" is filtered through mood, social desirability bias, fatigue, and the framing of every prior question. The behavior they actually exhibit on Tuesday at 11am — speaking up, asking for help, admitting confusion — is a more reliable signal.

They measure once or twice a year. By the time a quarterly engagement survey reaches HR, the data is months old. The team has moved on. The patterns that mattered have already produced their consequences.

They are not closed-loop. Employees fill in the survey. Nothing visibly changes. The next survey arrives. Response rates fall — below 40% within three months in most companies†. Once people stop trusting that their input will produce visible action, the data quality collapses.

The alternative is behavioral measurement: short, frequent, channel-native check-ins that capture what people actually do, not just what they say. The HR tech market, by and large, still measures self-report and calls the dashboard "data." The two are not the same thing — and the gap between them is exactly where attrition, burnout, and missed delivery hide.

Tatemae — what people show at work is not what they think

In Japanese, tatemae is the face one shows in public. Honne is what one actually thinks. The two are expected to differ, and in many social contexts the gap is functional and accepted.

Workplaces in every culture have their own version of tatemae. The polite answer in a one-on-one. The slightly upbeat response to "How is the project going?" on Slack. The careful phrasing in a channel where the manager is reading.

"What people show at work is not what people think at work. Most HR dashboards measure the show." tru.place

Most surveys measure tatemae. They ask employees what they think, and employees give a version of the answer that is safe to be seen. The dashboards then aggregate these answers into a score, and decisions are made on a number that is, by construction, the public face of the team rather than its private reality.

The closest observable proxy for honne is not asking again. It is reading behavior. Did the person speak up in the meeting? Did they ask a question that risked looking uninformed? Did they admit a mistake in front of the team? Did they push back on a senior colleague's plan?

These behaviors are visible. They are countable. They are far harder to fake than a 1–5 Likert response. And in aggregate, over weeks, they describe psychological safety more accurately than any single survey can.

This is the philosophical anchor of how tru.place measures psychological safety. Not by reading minds — by reading the behaviors that minds produce.

Six behaviors that predict psychological safety

In the SIGNAL™ Intelligence System — 5 layers · 6 nodes · 21 sub-clusters · 63 behavioral proxies — psychological safety is decomposed into six observable nodes. Each node represents a category of behavior that, when present, indicates a team is safe and, when absent, indicates risk.

The six nodes, by letter, name, and weight:

S
Safety to Speak
20% · highest
I
Initiative Behavior
15%
G
Group Belonging
15%
N
Need for Autonomy
15%
A
Authenticity Gap
20% · highest
L
Learning Comfort
15%

Source: tru.place SIGNAL™ framework · S and A weighted highest as most predictive of exit risk†

Authenticity Gap (A) carries the highest weight, at 20%, because in our internal modeling it is the most predictive single signal for exit risk over 90 days†. Safety to Speak (S) carries the same weight because it is the earliest signal to drop in the cascade.

This decomposition exists because "psychological safety" as a single score is too coarse to act on. A team scoring 65 overall can be either strong on S and weak on A (likely a team where people speak up but mask their identity) or strong on A and weak on S (likely a team that is authentic but afraid). The recommended action for each is different.

When psychological safety breaks — the cascade

Psychological safety does not collapse all at once. It cascades.

In the SIGNAL™ pattern library, the most common sequence runs as follows. Safety to Speak (S) drops first — people stop raising concerns in meetings. Within one to two weeks, Initiative Behavior (I) stalls — people stop proposing ideas without being asked. By weeks two to three, the Authenticity Gap (A) widens — people perform the role rather than show up as themselves. By weeks three to four, Learning Comfort (L) shuts down — questions stop, "I don't know" disappears, mistakes go underground.

By weeks six to eight, the visible consequences arrive: flight risk, burnout, unexplained delivery slips.

When the Authenticity Gap (A) exceeds 30 points, our internal modeling indicates a 67% probability of exit within 90 days†. The number is projected — see footnote. The cascade pattern itself, however, is well-documented in the academic literature. Morrison and Milliken's 2000 paper Organizational Silence describes how silence can become a collective property of a team, even when each individual member would, alone, choose to speak.

The structural conditions, not the people, produce the silence. This is why surveying individuals and asking "Do you feel safe to speak up?" misses the dynamic. The team-level pattern is what shifts, and behavior is the only thing that reveals it.

How is this different from employee engagement?

Engagement is the outcome. Psychological safety is the condition.

A team can score high on engagement-style metrics while psychological safety is quietly eroding underneath. The reverse is also true: psychological safety can be present in a team whose engagement score is depressed by an unrelated issue — bad tools, unclear strategy, slow procurement.

Engagement scores, by themselves, do not tell a manager what to change. Psychological safety, properly decomposed, does — because each of the six nodes maps to a different category of intervention. A drop in S suggests one set of moves: humble inquiry, framing the work as learning, modelling fallibility. A drop in A suggests another: reviewing identity safety, auditing the gap between stated values and observed promotions.

The diagnostic resolution is the point.

Pre-launch · 50 teams per month

tru.place is launching in 2026.

Join the waitlist at tru.place — we're letting in 50 teams a month. Risk intelligence for people decisions. Built in Europe. Designed for the world.

Join the waitlist →

Frequently asked questions

Who defined psychological safety?

Amy Edmondson, professor at Harvard Business School, defined psychological safety in 1999 in her paper Psychological Safety and Learning Behavior in Work Teams, published in Administrative Science Quarterly. The definition: a shared belief held by members of a team that the team is safe for interpersonal risk-taking.

Is psychological safety the same as being nice?

No. Psychological safety is the structural condition that lets people take interpersonal risk — speak up, admit mistakes, raise concerns, disagree with senior people. It is compatible with high standards, sharp disagreement, and direct feedback. Niceness without honesty is not psychological safety; it is closer to its opposite.

What did Google's Project Aristotle find?

Project Aristotle was Google's multi-year study of 180 internal teams. It concluded that psychological safety was the single biggest differentiator between high- and low-performing teams — more important than seniority, tenure, mix, or individual intelligence. The findings were published on Google re:Work in 2015.

How is psychological safety measured?

Academically, through Amy Edmondson's seven-item Psychological Safety Scale, a self-report instrument that asks team members to rate how their team handles mistakes, disagreement, and questions. In practice, most companies use annual or pulse surveys, which suffer from social desirability bias, low frequency, and lack of closed-loop feedback. Behavioral measurement, observing what people do rather than only what they say, produces a more reliable signal.

What causes psychological safety to drop?

Common causes documented in the research include leader response to bad news, restructures with unclear communication, performance review cycles that emphasise blame over learning, sudden senior departures, and workload shifts that remove time for reflection. Morrison and Milliken's 2000 work also shows that silence can become a collective property — a team can be silent even when each individual would, alone, choose to speak.

Can psychological safety be rebuilt once lost?

Yes, but rebuilding is slower than breaking. Edmondson identifies three practices: framing the work as learning, acknowledging fallibility as a leader, and modelling humble inquiry. The behavioral evidence of rebuilding tends to appear before the survey numbers do — more speak-up, more questions, more admitted mistakes.

Sources & notes

Projected from internal modeling of behavioral patterns. tru.place will publish validated retention probability data after six months of operational data, in line with our commitment not to misrepresent unvalidated figures.