Google’s Project Aristotle: star players or better team norms?

The room

The psychology: psychological safety

Amy Edmondson of Harvard Business School introduced team psychological safety in a 1999 paper. Her point was that learning at work is socially risky. Asking for help, admitting an error or challenging a plan can make you look ignorant, incompetent or disruptive, so people often stay quiet unless they believe the group will not reject, embarrass or punish them. The belief is shared rather than individual: it describes the climate of a team, resting on mutual respect and trust.

Psychological safety is not the same as comfort or niceness. Edmondson and Michaela Kerrissey wrote in 2025 that it does not mean avoiding conflict, backing every idea, removing accountability or protecting people from layoffs. Candid disagreement is exactly what it is meant to enable, and hard debates may not feel pleasant while they happen.

In Google’s data the concept fits the pattern the researchers saw. Composition variables did not separate strong teams from weak ones; what differed was how members treated each other – whether they spoke up, relied on each other and understood the goal. A team of high performers that punishes dissent can still lose the information it needs, while an ordinary team that surfaces mistakes quickly gets the chance to fix them.

Why smart people fall for it

Why would how a team talks matter more than who sits on it? Three mechanisms connect psychological safety to results.

1. 1

   Errors surface while still cheap

   When people can admit mistakes without fear, problems are reported early, when they cost less to fix.

2. 2

   Learning behaviour does the work

   In Edmondson’s study, safety went with seeking feedback and discussing errors, and that learning behaviour was what linked to performance.

3. 3

   More voices reach the decision

   Even turn-taking lets a group use the knowledge it already has, rather than only the views of its loudest members.

What did Project Aristotle actually prove?

Project Aristotle was internal management research, not a peer-reviewed study. Google released its findings through a re:Work blog post and guide, which describe the sample, the measures and the five dynamics but do not publish the statistical models, effect sizes or underlying data. Readers have to take the ranking of the five dynamics largely on trust.

The design was also observational. The researchers related survey measures and team attributes to effectiveness ratings; they did not randomly assign teams to higher or lower psychological safety. A team that is doing well may simply feel safer, so the arrow could run both ways. Google’s write-up uses language that implies cause in places, but the data it describes are associations.

Measurement adds a second caution. Three of the four effectiveness measures were ratings by people – executives, team leaders and the team members themselves – and only sales against quota was an objective number. If the same members who report feeling safe also rate their team as effective, part of the link can come from shared mood rather than performance. The guide does not say how the four measures were weighted.

Finally, the sample is Google. Critics argued that a company which screens hard for personality during hiring will see less personality variation on its teams, so composition may look unimportant there and matter more elsewhere. Google’s guide makes a similar caveat about team size, colocation and workload, noting these may still matter in other organisations. Psych Safety, a consultancy, adds that the team size range studied may not have been wide enough to show where large teams break down.

What did the research later show?

The broader literature supports psychological safety as a real and useful construct, with limits. Frazier and colleagues’ 2017 meta-analysis in Personnel Psychology pooled 136 independent samples covering more than 22,000 individuals and nearly 5,000 groups. It examined what predicts psychological safety and what it predicts in turn, including whether it explains task performance beyond leader relations and work engagement.

A plain-language review of that meta-analysis by ScienceForWork reported strong links between psychological safety and information sharing, learning behaviour and satisfaction, as well as role clarity and peer support. It also flagged that most of the included studies were cross-sectional, so they cannot on their own establish that safety causes better results, and that possible downsides had received little study.

The collective intelligence research points the same way from a different angle. Woolley and colleagues found that a group’s average or maximum individual intelligence did not strongly predict its performance across varied tasks, while social sensitivity and even turn-taking did. That is consistent with Google’s finding that the ‘who’ mattered less than the ‘how’, though it was a laboratory study of short-lived groups rather than of real teams.

The most useful later correction came from Edmondson herself. As psychological safety spread, she and Kerrissey wrote, many organisations misread it as being nice, avoiding conflict or shielding people from consequences. It cannot be created by policy alone; it is built in a group, interaction by interaction, and any team at any level can build it without waiting for the top.

How would a modern team decide?

A leader facing Google’s choice today would treat the finding as a strong hypothesis rather than a law. The pattern is consistent across Edmondson’s field study, the meta-analytic literature and Google’s own data, which makes it a sensible place to start. It is still worth testing locally, because your workforce may vary far more on skills and personality than Google’s did.

The practical sequence is to measure, discuss and change one norm at a time. Google’s gTeams process followed that pattern: a short survey, a facilitated conversation about the results, and a team-chosen norm to try. The reported gains were modest – 6% and 10% on ratings – a reminder that norms shift slowly and that survey scores are not the same as business results.

Composition should not be abandoned entirely. Skills still have to match the work, and a team missing a critical capability will not be rescued by good norms. The lesson of Project Aristotle is narrower: once a company already hires carefully, the next gains are more likely to come from how people work together than from another round of reshuffling.

Pair safety with standards. Edmondson and Kerrissey stress that psychological safety does not remove accountability; leaders can still address mistakes and weak performance. Track objective outcomes next to any climate survey, so you can see whether better norms are actually improving delivery rather than just producing friendlier meetings.

The manager's playbook

1. Measure team climate before reshuffling peopleRun a short anonymous survey on the five dynamics, with items such as whether members feel safe expressing divergent opinions.
2. Discuss the results with the teamShare scores in a facilitated session and ask which behaviours explain them, rather than sending the report to the manager alone.
3. Pick one norm and try itAgree on a single concrete change, such as rotating who speaks first in meetings or reviewing mistakes without blame.
4. Model fallibility as the leaderName your own uncertainties and errors, frame the work as a learning problem and ask genuine questions, as Edmondson suggests.
5. Track hard outcomes alongside the surveyCompare climate scores with delivery, quality or revenue so you can tell whether better norms are changing results, not just ratings.

Questions people ask