Google’s Project Aristotle and Psychological Safety

The Answer is Kind of Simple

Around 2012, Google decided to apply its signature approach, vast datasets, rigorous analysis, and a deep institutional faith in the quantifiable, to one of the oldest questions in management: why do some teams work brilliantly and others collapse into mediocrity despite having equally talented members?

The initiative was called Project Aristotle, a name chosen in homage to the philosopher’s observation that the whole is greater than the sum of its parts.

What followed was a piece of corporate research that, for all its limitations, has done more to popularise the concept of psychological safety than perhaps any academic paper ever could.

What Google Expected to Find

To understand why Project Aristotle’s findings landed with such force, you need to appreciate what Google’s researchers assumed going in. Google is, at its core, an engineering culture. It is a place that believes deeply in optimisation, in the idea that with enough data you can solve any problem, including the problem of human collaboration. The initial hypothesis was entirely consistent with this worldview: surely, the key to building great teams was assembling the right people. Find the optimal combination of skills, experience levels, personality types, educational backgrounds, and working styles, and the team would perform. It was, in essence, a composition hypothesis, the idea that team effectiveness is primarily a function of who is in the room.

This is not a foolish assumption. It is, in fact, the assumption that most organisations operate under most of the time. Recruitment processes, talent mapping exercises, team restructures, and the entire apparatus of workforce planning are built on the premise that getting the right people in the right seats is the fundamental challenge. And to be fair, it is not entirely wrong. You cannot build an effective software engineering team out of people who cannot write code. Competence matters. But Google already had competence in abundance. What they wanted to know was why, given a baseline of exceptional talent, some of their teams soared while others stalled.

The research team, led by Julia Rozovsky (a People Analytics manager, not a software engineer, which is itself an interesting detail), spent the first phase of the project looking at team composition variables. They examined whether teams performed better when members were friends outside of work. They looked at educational backgrounds. They analysed personality mixes. They studied gender balance, tenure, seniority, and the degree of overlap in members’ social networks. And they found, to their considerable frustration, essentially nothing. No consistent pattern of composition predicted team effectiveness. Some brilliant teams were made up of close friends; others were composed of near strangers. Some had dominant personalities; others were notably egalitarian. The data refused to cooperate with the hypothesis.

The failure of the composition hypothesis is arguably the most important finding of the entire project, even though it is the one most frequently skimmed past in popular accounts. Google, a company with access to more employee data than virtually any organisation on earth, could not identify a reliable recipe for team composition. If they could not do it, we should probably stop pretending we can.

What They Actually Found: The Five Dynamics

Having exhausted the composition approach, the research team shifted its focus from who was on the team to how the team worked. And here, at last, the data began to speak. Through a combination of surveys, interviews, and performance data, the researchers identified five dynamics that distinguished Google’s most effective teams from the rest. These were not personality traits or demographic characteristics; they were norms, behaviours, and shared experiences that characterised the team’s internal culture.

The five dynamics, listed in order of importance as Google reported them, were:

• Psychological safety – team members feel safe to take risks and be vulnerable in front of each other. They do not fear punishment or humiliation for admitting mistakes, asking questions, or proposing ideas that might not work.
• Dependability – team members reliably complete quality work on time. People do what they say they will do, and the team can count on each member to fulfil their responsibilities.
• Structure and clarity – the team has clear roles, plans, and goals. Members understand what is expected of them individually and what the team is collectively trying to achieve.
• Meaning – the work is personally important to team members. People find a sense of purpose in what they are doing, whether that purpose is financial security, creative expression, supporting their family, or contributing to something they believe in.
• Impact – team members believe their work matters and creates change. There is a sense that the team’s output makes a difference, that it is not merely busywork disappearing into an organisational void.

All five dynamics mattered, but psychological safety was not simply first among equals. It was, in Google’s analysis, foundational. It was the dynamic that underpinned and enabled the other four. A team without psychological safety

• Could not develop genuine dependability, because people would not admit when they were struggling,
• Could not achieve real clarity, because people would not ask the clarifying questions that felt too basic or too challenging,
• Could not sustain meaning or impact, because people would not voice the concerns or aspirations that gave their work personal significance.

Psychological safety, in short, was the soil in which everything else grew. Without it, even the most talented, well-resourced teams underperformed.

Psychological Safety: Edmondson’s Foundational Work

Google did not invent the concept of psychological safety. That credit probably belongs to Carl Rogers. But it wasAmy Edmondson, whose 1999 paper “Psychological Safety and Learning Behavior in Work Teams” that really brought the idea into the mainstream. Edmondson was studying medication error rates in hospital teams when she noticed something counterintuitive: the teams with the best leadership and the strongest interpersonal relationships reported more errors, not fewer. This did not mean they were making more mistakes. It meant they were more willing to admit to them. The “better” teams had created an environment where reporting errors was safe, which meant errors were caught, discussed, and learned from rather than hidden.

From this observation, Edmondson defined psychological safety as a shared belief held by members of a team that the team is safe for interpersonal risk-taking. The key word is “shared.” Psychological safety is not an individual trait or a generalised sense of confidence. It is a property of the team, a collective perception about the social consequences of speaking up, admitting ignorance, challenging the status quo, or reporting failure. In a psychologically safe team, people believe they will not be punished, humiliated, or marginalised for doing any of these things. In an unsafe team, they believe they will be, and so they don’t.

Edmondson’s research demonstrated that psychological safety predicted learning behaviour in teams, which in turn predicted team performance. Teams that felt safe were more likely to seek feedback, discuss errors, experiment with new approaches, and ask for help. These behaviours drove learning, and learning drove performance. The causal chain was not safety leading directly to better results; it was safety leading to better learning leading to better results. This distinction matters, because it means psychological safety is not a feel-good end in itself (though it may feel good). It is a functional precondition for the kind of candid, iterative, sometimes uncomfortable learning that complex work demands.

The Fearless Organization: Edmondson’s 2019 Synthesis

Two decades after her original paper, Edmondson published The Fearless Organization (2019), which synthesised twenty years of research and practice into a comprehensive account of what psychological safety is, why it matters, and how leaders can build it. The book arrived at a moment when the concept had already begun its migration from academic journals to corporate slide decks, and part of its value lies in Edmondson’s careful effort to preserve the nuance that popular accounts were busily stripping away.

Several points from The Fearless Organization are worth highlighting for practitioners. First, Edmondson is explicit that psychological safety is not about being nice. It is not an absence of conflict, disagreement, or high standards. In fact, the opposite is closer to the truth: psychologically safe teams tend to have more productive conflict, because people feel able to challenge ideas without it becoming personal. A team where everyone agrees with the leader and no one raises concerns is not psychologically safe; it is psychologically terrified. The silence looks like harmony, but it is the silence of self-censorship, and the cost is paid in poor decisions, missed risks, and innovations that never get proposed.

Second, Edmondson distinguishes psychological safety from trust. The two are related but not identical. Trust is typically a dyadic phenomenon, something that exists between two individuals. Psychological safety is a group-level phenomenon, a shared climate that characterises the team as a whole. You might trust your colleague deeply but still feel that the team’s culture does not permit candour. Or you might not know your teammates especially well but feel, because of how the team operates, that it is safe to speak up. The group climate can override individual relationships in either direction.

Third, Edmondson offers a practical framework for leaders, organised around three clusters of behaviour:

1. Setting the stage (framing the work as a learning problem rather than an execution problem, and acknowledging one’s own fallibility),
2. Inviting participation (asking genuine questions, creating structures for input, and demonstrating situational humility), and
3. 3 Responding productively (expressing appreciation for candour, destigmatising failure, and sanctioning clear violations while distinguishing them from honest mistakes).

These are not revolutionary behaviours, but their consistent absence in most organisations is what makes them so important.

Why Psychological Safety Matters More Than Talent

The most provocative implication of Project Aristotle, and the one that has generated the most resistance in talent-obsessed corporate cultures, is the suggestion that psychological safety matters more than the composition of the team. This does not mean talent is irrelevant. It means that talent, without the conditions that allow it to be expressed, is wasted. And the primary condition, the thing that determines whether talented people actually contribute their best thinking, is whether they feel safe enough to do so.

Consider what happens in a team without psychological safety. The junior member who spots a flaw in the strategy stays quiet because the last person who challenged the director’s thinking was publicly dressed down. The engineer who does not understand the requirements pretends she does because asking for clarification would make her look incompetent. The team member who has a creative but unconventional idea keeps it to himself because the team’s unwritten rule is that only “safe” suggestions are welcome. In each case, the team has the talent it needs. What it lacks is the environment that allows that talent to function.

This is not merely a theoretical concern. Woolley and colleagues’ 2010 research on collective intelligence found that a group’s collective intelligence, its ability to perform well across a range of tasks, was not strongly correlated with the average or maximum individual intelligence of its members. Instead, it was predicted by the equality of conversational turn-taking (whether everyone contributed roughly equally to discussion / share of voice), the social sensitivity of group members (their ability to read one another’s emotional states), and, notably, the proportion of women in the group (which the researchers attributed to women’s higher average social sensitivity). These findings complement Project Aristotle’s conclusions: what makes a group smart is not how smart the individuals are, but how well they interact.

Charles Duhigg’s 2016 account of Project Aristotle in the New York Times Magazine brought this point to a wider audience, telling the story through the experience of specific Google teams and making vivid the contrast between teams where people felt free to be themselves and teams where they did not. The article did an enormous amount to popularise the concept, though, as with any popularisation, something was inevitably lost in the translation from research finding to narrative journalism.

This is actually the article that first really ignited our interest in the subject. You can still read it here.

Practical Applications: Building Psychological Safety

The value of any framework lies ultimately in what you can do with it, and the psychological safety literature offers a number of practical implications that we have found genuinely useful in our work with teams and leaders.

Leaders go first. Psychological safety is not built by announcing that the team is a safe space. It is built by leaders who model the behaviours they want to see. This means admitting mistakes openly, saying “I don’t know” without treating it as a confession of failure, asking for feedback and visibly acting on it, and responding to bad news with curiosity rather than blame.

Delizonna’s 2017 work on creating psychological safety identifies this leadership modelling as the single most powerful lever available. In our experience, she is right. Teams watch what leaders do, not what they say, and a single punitive response to honest feedback can undo months of stated good intentions. Edgar Schein also notes that leader role-modelling is a primary factor in any form or organisational culture.

Normalise productive failure. This does not mean tolerating incompetence or lowering standards. It means creating clear distinctions between blameworthy failures (negligence, ethical violations) and praiseworthy failures (well-designed experiments that did not produce the hoped-for result, honest mistakes made in good faith). Edmondson’s framework for distinguishing failure types is particularly useful here. Most organisations treat all failure identically, which means they get less of it reported and more of it hidden. The goal is not to celebrate failure for its own sake (which is a somewhat tiresome Silicon Valley affectation) but to make it safe to surface, discuss, and learn from failure when it occurs.

The predictive processing framework offers a useful neurological grounding for why this matters: at the neural level, learning is driven by prediction error — the gap between what was expected and what actually happened. An environment where mistakes are punished suppresses prediction error, and with it the brain’s capacity to update its models. Psychological safety is, in this sense, the organisational condition that keeps the learning system switched on. See our article on Predictive Coding: The Brain as a Prediction Machine for a fuller account of this mechanism.

Design for voice, not just for presence. Having people in the room is not the same as having their input. Meetings dominated by the most senior or most confident person are not psychologically safe simply because no one is being shouted at. Safety requires active structures that draw out quieter voices: round-robins, pre-meeting written input, anonymous idea submission, or simply the discipline of asking “who hasn’t spoken yet?”

These are small interventions with disproportionate effects, and they signal, more powerfully than any values statement, that every perspective is expected and valued.

Address the interpersonal risks directly. People in teams face four specific interpersonal risks: the risk of being seen as ignorant (for asking questions), incompetent (for admitting mistakes), negative (for raising concerns), or disruptive (for challenging the status quo).

Effective leaders name these risks explicitly and create norms that mitigate them. When a leader says “I would rather you ask a question that feels obvious than proceed with an assumption that turns out to be wrong,” they are directly addressing the ignorance risk and giving people explicit permission to take it.

Remember that psychological safety is local. It exists at the team level, not the organisational level. A company can have pockets of extraordinary psychological safety alongside teams where people are terrified. Organisational culture sets the conditions, but the team leader is the proximate architect. This is simultaneously empowering and sobering: empowering because any leader can begin building safety in their own team regardless of the broader culture, and sobering because it means the problem cannot be solved by a company-wide initiative alone.

Limitations and Criticisms:

Project Aristotle has been enormously influential, and much of its influence has been beneficial. But there are limitations of the research and the ways in which its findings have been oversimplified, overclaimed, or misapplied. We’re not going to spend much time on them, but they are basically that:

• Google employees are, by any measure, an extraordinarily non-representative sample of the working population. They are disproportionately highly educated, well-compensated, working in knowledge-intensive roles, and self-selected into a culture that prizes intellectual confidence. Findings in relation to them might not generalise well.
• Project Aristotle was fundamentally an observational study. It identified associations between team dynamics and team effectiveness, but it did not (and could not, given its design) establish causal direction. Do psychologically safe teams perform better because safety enables better collaboration? Or do high-performing teams develop psychological safety as a consequence of their success, because winning together builds trust and openness? Correlation is not causation, and all of that.
• There is a risk of “psychological safety washing”, organisations that adopt the language of psychological safety without doing the actual work and making the meaningful cultural changes needed for teams to actually feel psychological safety. 
• Both Project Aristotle and Edmondson’s broader work are rooted primarily in Western, and specifically American, organisational contexts. Cultural context matters, and the degree to which psychological safety impacts performance may differ in different cultures.
• Dependability, structure and clarity, meaning, and impact matter. These other four dynamics identified by Project Aristotle have been somewhat neglected. Perhaps the five factors need to be considered as a system, and these four factors less down-weighted.

Project Aristotle in Context: Connecting the Threads

It is worth stepping back to consider where Project Aristotle sits within the broader landscape of team effectiveness research. Its findings are remarkably consistent with several other established frameworks. Lencioni’s Five Dysfunctions of a Team places trust (which functions similarly to psychological safety) as the foundational layer upon which everything else rests. Tuckman’s model of team development implicitly acknowledges that teams must navigate the interpersonal vulnerability of the storming phase before they can reach genuine norming and performing. Hackman’s conditions for team effectiveness include a supportive organisational context and strong group norms, both of which overlap substantially with the dynamics Google identified. Adair’s ACL model calls out the importance of team dynamics, and so on.

What Project Aristotle added was not a fundamentally new insight but rather a powerful confirmation of existing theory, backed by the data resources and cultural cachet of one of the world’s most prominent companies. Academic researchers had been saying for decades that team climate matters more than team composition. Edmondson had been publishing on psychological safety since the late 1990s. But it took Google, with its brand, its data, and its willingness to be publicly surprised by its own findings, to make the idea stick in the popular imagination. There is something slightly ironic about the fact that a concept rooted in organisational behaviour research only achieved mainstream recognition when a technology company arrived at the same conclusions through its own proprietary analysis.

Learning More

To go deeper, explore these resources on people-shift.com:

• A Simple Introduction to Trust
• The Five Dysfunctions of a Team
• Fearless Feedback In The World of Work
• A Group or a Team: What’s the Difference?
• David Rock’s SCARF Model
• Predictive Coding: The Brain as a Prediction Machine

The People Shift View

At PeopleShift, we reference Project Aristotle a lot and really like the cache that name-dropping Google brings to training, whatever sector we’re in. The project really lent the weight of data and the prestige of a global brand to something that practitioners and researchers had been saying for years: that team effectiveness is fundamentally about how people interact, not merely about who they are individually.

The primacy of psychological safety in Google’s findings confirms what we see consistently in our own work with teams. The most talented groups underperform when people do not feel safe to contribute honestly, and the most ordinary-seeming groups can produce extraordinary results when they do.

What we would caution against, however, is the reductive version of this story that has taken hold in too many organisations. Psychological safety is not a programme, a workshop, or a declaration. It is a lived experience that must be built through consistent leadership behaviour, thoughtful team design, and an honest reckoning with the systemic conditions that make teams unsafe. Adding “psychological safety” to your corporate values while maintaining a culture of blame, competitive ranking, and punitive responses to failure is not building safety. It is building a more sophisticated form of denial.

Start with the honest question: in your team, right now, is it genuinely safe to say “I made a mistake,” “I don’t understand,” or “I disagree”? If the answer is no, or even “it depends on the topic,” then you have work to do. And that work begins not with a training programme but with a mirror. Leaders really are the starting point for this, and they might need to go through some personal change to really create psychologically safe teams.