Psychological safety: make questions, mistakes and dissent discussable

Psychological safety is a team’s shared expectation about what happens when someone takes an interpersonal risk: asks a question that might sound naive, admits a mistake, brings bad news or disagrees with a more senior colleague. Amy Edmondson, who introduced the team-level concept, defined it in her 1999 study of work teams as “a shared belief held by members of a team that the team is safe for interpersonal risk taking.” Where that belief is strong, people expect interest rather than embarrassment, rejection or punishment.

It is not the same as harmony. A psychologically safe team can hold sharp disagreements, demanding goals and uncomfortable conversations. What it reduces is the quiet calculation that keeps useful information out of the room: “If I say this, I will look incompetent, disloyal or difficult.” The practical question for a team lead is therefore not “Are people happy?” but “How early do questions, mistakes and dissent reach someone who can act on them?”

Consider a fictional travel-booking company whose support agents have known for months that the refund screen confuses customers. The product team takes it seriously only when card disputes start to rise. This invented case runs through the guide.

What the definition actually says

Edmondson’s paper makes three points that everyday use of the term tends to lose. First, the belief is usually tacit: members infer it from what happened the last time someone admitted an error or questioned a decision. Second, it describes the team rather than a personality trait, so members need broadly similar perceptions for it to count as a shared climate. Third, it goes beyond interpersonal trust: the paper describes a climate of trust and mutual respect in which people are comfortable being themselves.

The same study linked that belief to learning behaviour: seeking feedback, sharing information, asking for help, talking about errors and experimenting. Across 51 teams in one manufacturing company, psychological safety was associated with learning behaviour, which in turn mediated its relationship with team performance. Team leader coaching and supportive organisational context were among the conditions associated with it.

Later reviews widened the frame. Edmondson and Lei’s 2014 review describes psychological safety as people’s perceptions of the consequences of taking interpersonal risks in a particular context, and traces research at individual, team and organisational levels, including work on voice and team learning. Edmondson and Bransby’s 2023 review calls it “a state of reduced interpersonal risk” and groups the recent literature under four themes: getting things done, learning behaviours, improving the work experience and leadership.

What it is not: harmony, comfort or lowered standards

The 1999 paper says the term is meant to suggest “neither a careless sense of permissiveness, nor an unrelentingly positive affect”. It also separates psychological safety from group cohesiveness, which can make people less willing to challenge one another. In a 2019 HBR IdeaCast interview, Edmondson said the label can imply cosiness and everyone being nice to each other, “and that’s not what it’s really about. What it’s about is candor”.

Niceness that suppresses a concern reduces the information available to the team. Comfort is not the test either: raising a problem can still feel awkward, and the question is whether the response makes that worth bearing. Nor is it consensus. Disagreement is expected, and decisions still need an owner. Google’s re:Work guide on team effectiveness lists consensus-driven decision-making among the variables that were not significantly connected with team effectiveness in the company’s own internal study.

Psychological safety does not mean freedom from consequences. Honest mistakes in uncertain work call for learning. In the same interview, Edmondson says someone who keeps making mistakes needs real help, such as coaching or training, or else a decision to “free up their future”. Harassment, discrimination and fraud belong in formal processes, not in a climate conversation.

Edmondson treats standards as a separate dimension. A 2023 qualitative study of medical students’ clerkships, which she co-authored, sets out her two-by-two, adapted from her book Teaming: psychological safety on one axis and accountability, “the degree to which people are expected to adhere to high standards and pursue challenging goals”, on the other. Low on both is the apathy zone. High safety with low accountability is the comfort zone; high accountability without safety is the anxiety zone. High on both is the learning zone. It is a thinking tool, not a measured classification, but it corrects a common mistake: treating safety and demanding goals as a trade-off.

Why the refund problem surfaced so late

In our invented company, the refund confirmation screen shows “€0.00 – pending” until the payment provider confirms the transaction, which can take several days. Customers read the zero as a refusal. Many contact support; some dispute the charge with their bank. Agents use a saved reply that explains the delay and closes the contact quickly.

Priya, the support team lead, heard about the screen in spring. An experienced agent mentioned it at a monthly product sync, and Helen, the senior product manager, replied that the design had passed usability testing and the roadmap was fixed for the quarter. Her tone was not hostile, but nobody raised it again. Tomás, a support agent three weeks into the job, assumed he had misread the screen and once told a customer that their refund had been declined. It became a product priority only when finance flagged the disputes.

Nothing here requires a villain. Edmondson’s 1999 paper notes that people who admit an error or ask for help risk appearing incompetent and may bear tangible costs with people who influence their prospects. In the IdeaCast she adds that wanting to look good, divert blame and agree with the boss are instinctive, and that hierarchies exaggerate these instincts. For a new agent, the risk of looking ignorant is immediate. The benefit of speaking, a clearer screen for future customers, is uncertain and delayed.

The work system played a part too. Agents were measured on handling time, so the saved reply made the problem cheap to absorb and costly to escalate. Encouraging agents to speak up will not change that on its own: double-loop learning asks whether the measure itself is teaching people to stay quiet. To understand such patterns in your own setting, investigate the unwritten rules around bad news by tracing what happened to recent concerns rather than surveying attitudes in general.

What leaders do that makes speaking up easier

In the IdeaCast, Edmondson describes three activities that leaders need to keep repeating: setting the stage, inviting engagement and responding productively. The re:Work guide summarises three things she recommends in a TEDx talk: frame the work as a learning problem rather than an execution problem, acknowledge your own fallibility, and model curiosity by asking lots of questions. These describe what the originator advises, not tested interventions.

Frame the work as learning. Explain why voice matters in this particular work. Edmondson links it to how uncertain, complex and interdependent the work is. Priya might open the week with: “Refunds pass through systems we don’t fully control. You will see problems before anyone else does.”

Acknowledge fallibility. The person with most authority goes first. Priya says she heard about the screen in spring and let it drop. This is not a confession ritual. It shows that being wrong is survivable here.

Ask genuine questions. A genuine question is one whose answer you do not already know and which could change what you do. “What are you seeing with refunds this week?” invites information; “Everyone’s fine with the new flow, right?” invites agreement. Systemic questions offer a wider repertoire for exploring a pattern without smuggling in the answer.

Respond productively. The response to bad news is where a climate is made or lost. Edmondson’s suggested reply to someone reporting a delayed project is “Thank you for that clear line of sight”, followed by asking what help they need. Disappointment is fine; anger teaches people not to bring the next piece of news.

Follow through. This one is our addition: report back on what happened to the input, including when the answer is no.

When Tomás tells the next stand-up that he gave a customer the wrong answer, Priya thanks him, asks what on the screen led him there and arranges a correction for the customer. His mistake becomes the clearest evidence yet that the screen misleads even trained staff.

Meeting practices that make voice routine

Leader behaviour is easier to sustain when it is built into how meetings run, so that speaking up does not depend on someone’s courage that day. The practices below are practical adaptations.

Hold a round before discussion. Ask everyone to note a response privately, then go round the group before open debate. Newer people get a turn before the most senior voice has set the direction. Lean Coffee applies a similar principle to the agenda itself, letting participants propose and vote on the topics.

Run a premortem on the plan. Asking people to imagine that the fix has already failed makes criticism the assignment rather than an act of disloyalty. Our premortem guide sets out a 60-minute version.

Separate generating ideas from judging them. Evaluation during generation rewards the confident and discourages the tentative. Written formats such as brainwriting protect the first stage, and a 90-minute idea workshop shows how to schedule the two phases apart.

Thank the messenger specifically. “Thanks for raising that before the release” names the behaviour you want repeated.

Act visibly on input. Keep a short record of what was raised and what happened to it: changed, tested, parked with a reason or declined with a reason.

In the fictional company, Helen’s first joint session with support went badly. She opened by presenting her fix, relabelling “pending” as “refund on its way”, and asked for reactions. The room nodded. At the second session, Priya asked everyone to write down one way the fix could fail before anyone spoke. Kwame, the payments engineer, wrote that the zero was a display default: the refund amount is known at cancellation, so a new label would still show customers €0.00. Helen wanted the label change in the current sprint. She asked Kwame what would settle the question, and he offered to check a sample of recent cancellations the next day.

Responses that close or open the conversation

Small replies carry much of the signal. The table sets moments from the example beside a reply that tends to shut voice down and one that keeps it open. The wording is illustrative, not a script.

Situation	Shuts voice down	Keeps it open
A new agent asks why the refund shows €0.00	“It’s in the onboarding pack.”	“Good question. Show me the booking where you saw it.”
An agent admits giving a customer the wrong answer	“How did you miss that?”	“Thanks for telling us quickly. What on the screen led you there?”
An engineer disputes the product manager’s fix	“We’ve already decided this.”	“What would show us which view is right? Let’s check before we ship.”
Bad news arrives late	“Why am I only hearing this now?”	“I’m glad you raised it. What made it hard to raise earlier?”
An idea sounds unworkable	“Finance will never accept that.”	“Let’s capture it and test it against the constraints when we evaluate.”
A concern cannot be acted on this quarter	No reply	“We can’t change it before March because of the provider contract. I’ll bring it back then.”

The middle column is not always rude, and “We’ve already decided” may be true. The difference is whether a contribution is treated as information or as an interruption. The last row matters most over time: explaining a no, and when the issue will be revisited, is itself follow-through.

Psychological safety during change

Change raises the interpersonal stakes. People must use unfamiliar tools, follow new rules and admit what they do not yet know, often in front of the managers who sponsored the change. The information a change team most needs, what does not work in practice, is exactly what people hesitate to report.

That is why assessing change readiness depends on the climate in which you ask. If the first honest answer is corrected or dismissed, later answers become agreeable rather than accurate. Objections are often early reports about workload, missing access or a flawed design, which is why our guide to resistance to change treats them as information rather than a motivation problem.

In the example, Kwame’s sample supports his point. Helen drops the label change and asks for the actual refund amount and an expected date to be shown instead, accepting that it will take longer than one sprint. Support gets its own change: a tag for product problems that does not count against handling time, and a fortnightly thirty-minute review with Helen and Kwame. Before redesigning the screen, Helen proposes to map the refund journey with support so that the front-line account and the system’s backstage steps sit side by side. Priya plans to check after a month whether newer agents use the tag as well as experienced ones. These are the invented team’s decisions, not reported results.

Measuring psychological safety without turning it into a score

The most widely used instrument is the seven-item team survey from Edmondson’s 1999 paper. It includes statements such as “If you make a mistake on this team, it is often held against you” (reverse-scored) and “Members of this team are able to bring up problems and tough issues.” The re:Work guide reproduces the items and describes Google teams receiving aggregated, anonymised scores from the company’s own survey to inform a discussion. Used as a prompt for conversation, a short survey can help. Several cautions apply.

It is self-report, and people who feel unsafe may not answer candidly, especially in small teams where responses feel identifiable. It is also a snapshot: Edmondson noted that her cross-sectional design could not show how psychological safety develops over time.

Averages hide subgroups. A team mean can conceal that new agents or evening staff feel very differently from long-serving colleagues, and if perceptions diverge widely there may be no single shared climate to describe. Climates also vary within one organisation. In the 1999 study, psychological safety differed significantly from team to team inside a single company, so an organisation-wide figure says little about a particular team.

Targets distort. Once a score becomes a manager’s objective, people learn which answers are expected. When an area manager in our example proposes adding the survey score to team-lead goals, Priya argues for watching observable signals instead: how long a known problem takes to reach someone who can act, who speaks in meetings and who does not, and what happens to raised items.

Expect reported problems to rise. The IdeaCast recounts Edmondson’s early hospital research, in which better teams appeared to report more errors, possibly because they were more willing to talk about them. More raised issues can mean better visibility rather than worse work.

Evidence and limitations

The research base is large. Frazier and colleagues’ meta-analysis drew on 136 independent samples, covering more than 22,000 individuals and nearly 5,000 groups, to examine the antecedents and outcomes of psychological safety. A 2020 umbrella summary from the Quality Improvement Center for Workforce Development, then funded by the US Children’s Bureau, reads the meta-analysis as showing, at the individual level, strong associations with engagement, job satisfaction and commitment and moderate associations with task performance and helpful discretionary behaviour. It also reports positive links with leader, peer and organisational support and with autonomy, interdependence and role clarity.

Most of this evidence is correlational. The same summary states that a lack of experimental research leaves little evidence about what causes psychological safety to develop. Newman, Donohue and Eva’s systematic review likewise called for better theory on how psychological safety develops and influences outcomes at different levels. Edmondson’s original study covered 51 teams in one company; she described its findings on antecedents as tentative and cautioned against generalising to other organisations.

Google’s re:Work guide is an organisation’s account of its own internal study of 180 teams, not a peer-reviewed paper. It lists psychological safety first of five dynamics in order of importance at Google and notes that factors which did not matter there may matter elsewhere. Treat it as an influential illustration, not proof that a survey programme improves performance.

Culture and hierarchy may shape how voice works. Frazier and colleagues examined whether national culture and research design alter the relationships. In the IdeaCast, Edmondson argues that psychological safety matters in hierarchical, high power-distance settings but is harder to achieve there, and points to structural aids such as the andon cord, which lets a production worker flag a problem without making a speech.

Psychological safety also has boundaries. It does not replace whistleblowing routes, grievance procedures or proper investigation of harassment and discrimination, and a meeting format cannot create safety where people have recently been punished for speaking. What the research supports is more modest and still useful: teams in which people expect interest rather than punishment tend to show more learning behaviour, and leader support is among the conditions associated with that expectation.