The Four Stages of Psychological Safety: a Six-Stage Critique
A Critique of The Four Stages of Psychological Safety – a second edition
We’ve written about the Four Stages of Psychological Safety before (way back in 2020 in fact), and that earlier piece was somewhat generous. It concluded that the model is potentially useful despite being wrong (though I didn’t, at the time, clarify in which way it was wrong), and potentially harmful if applied uncritically. Now I’m revisiting the Four Stages and reflecting on a few years of practice since then.
In the spirit of applying an inappropriate numerical structure to a complex phenomenon that doesn’t actually arrange itself in numbered steps, here is my Six Stage critique of the Four Stages model. They’re not really stages, and don’t need to be read in order. You can skip some, double back, or read it in reverse, which is kind of the point.
Before we begin, I want to be fair: the Four Stages model does some potentially useful things. It gives people a vocabulary for noticing that psychological safety is not a binary on/off, and that (for example) feeling safe to suggest an idea is not the same as feeling safe to dissent. That’s useful, and conversations that start with “which kind of thing isn’t safe to say here?” are probably better than conversations that don’t happen at all. I’m not arguing that the model is worthless. I am suggesting that it is far weaker, and in places actually counterproductive and harmful, than its popularity suggests.
Stage One: The Empirical Problem
The first thing to ask of any model that arranges human behaviour into stages is: how do we know the stages actually exist?
For the Four Stages, the honest answer is that we don’t. There is no body of peer-reviewed research demonstrating that teams move through stages of inclusion, learner, contributor, and challenger safety in that sequence (or at all), or that the two proposed drivers — respect and permission — are the variables that move them. The model is presented with the confidence of a finding, without the evidence that a finding requires.
This is the moment at which one may reach for George Box: all models are wrong, but some are useful, so why are we being so demanding? But “wrong” is not just one type or degree of wrongness, and as Dave Snowden has argued in this excellent piece (which inspired this article), the habit of treating wrongness as one thing is precisely how the aphorism gets weaponised to shut down inquiry and facilitate the uncritical commercialisation of a model.
So what kinds of wrong can a model be? Firstly there’s the justifiably wrong model that is merely a simplified version of reality — a map that’s wrong only in the unavoidable sense that the map isn’t the territory. That kind of wrong is usually useful. Secondly there is the model that is incoherent, that contradicts itself or makes unfalsifiable claims that cannot be tested. Like a map that claims a teapot that is too small to detect with a telescope exists in solar orbit between Earth and Mars – we can’t prove that it’s wrong, but we know it is. Thirdly there is the model that is coherent in itself but applied far outside the domain where it holds – a map that isn’t actually of the place it claims to be. Box was talking about the first kind of wrong (the simple map of the real territory). The Four Stages, as we will see, is also wrong in the other two ways: incoherent in the sense that its central claim can’t be tested (Stage Two), and applied outside the domain where any stage model could hold (Stage Three).

A defender might say the Four Stages is not a model but a framework, and that frameworks, as Snowden notes, cannot be wrong in Box’s sense. That distinction is real, but it doesn’t come to the rescue here. The Four Stages asserts a numbered sequence, names two driving variables, and claims that we can progress through them. That’s the grammar of a model: claims that can in principle be tested. A thing that helps itself to the authority of a model does not get to retreat to the modesty of a framework when it’s convenient.
There is possibly a structural reason the claims have never been seriously tested. The model’s vocabulary is trademarked, its assessment instrument is proprietary, and the certification that authorises one to teach it is sold. A framework that generates consultancy revenue, certification fees, and assessment licences has a powerful interest in remaining stable, recognisable, and certifiable, and if not proven to be true, at least not proven to be false. Commercialisation does not make a model wrong but it does create disincentives for genuine testing through which a model may be challenged, altered or abandoned. The commercial architecture helps to explain why many commercial (and often weak) models tend to go unexamined.
Academically, whilst the model has been cited as a descriptive framework, its vocabulary, “inclusion safety,” “learner safety,” “contributor safety” and “challenger safety,” has barely been taken up by researchers even as terminology to argue with. A search for the Four Stages model in Google Scholar returns a few hundred results, many of them the creator’s own. Edmondson and Kahn have a two- or three-decade head start, so the raw numbers aren’t a fair fight, but the pattern is. Edmondson’s 1999 paper was being tested, extended and argued with within a few years of publication. Six years on, the Four Stages vocabulary is still rare in the academic literature. Edmondson’s terms saturate the field precisely because people test, extend and evolve them, and (importantly) dispute them. This marks the difference between an idea circulating openly in research and practice, where it can be built on and challenged by anyone, and a product in a market, where it’s less likely to.
Fundamentally, a model that adopts the visual and rhetorical grammar of leadership and organisations — numbered stages, 2×2 diagrams, and a progression curve (if only it also had an iceberg image it would qualify for the full business-grade quadrumvirate) — while resting on the evidential basis of a well-organised opinion borrows the authority of research without having done the work. And this is largely invisible to the person who sees the neat diagram and reasonably assumes someone, somewhere, actually checked it.
Of course, we don’t need to demand years of randomised controlled trials before anyone is allowed to try something. The only RCT ever conducted on the effectiveness of parachutes involved jumping from a plane parked on the ground, which, if I relied solely on research trials instead of experience, would leave me rather nervous about flying. Practitioner experience is real evidence, but the absence of any significant attempt to rigorously test the claims of the Four Stages over several years stops being a minor footnote and starts being the headline.
Stage Two: The Structural Problem
Even setting the evidence aside, the model makes a fundamental structural error. It takes a thing that doesn’t exist in discrete stages, and arranges it in discrete stages.
Inclusion, learning, contributing, and challenging are not rungs on a ladder. They are different aspects, outcomes and interpretations of the same underlying thing, and they vary independently, contextually and culturally. We’ve worked with teams where people challenged ideas from authority robustly while remaining anxious of raising concerns. We’ve encountered software engineers who contribute ideas freely whilst being reluctant to ask for help. (These are anecdotes, offered as exactly the class of evidence the model itself rests on.) And there are many other examples too. These are not people stuck at one stage on their way to stage four. They are people whose experience of psychological safety is shaped by complex power dynamics, and by their own history, culture and context.
This is where the model slides into Snowden’s second kind of wrong. Ask what observation would count against the sequence. A team that challenges freely while being anxious to ask for help is not treated as a counterexample; it is described as having “skipped” a stage, or having “incomplete” learner safety. A programme that fails is explained by the team being “stuck” at a stage rather than by the stages being wrong. The two drivers, respect and permission, are defined loosely enough that any outcome can be read back into them. A claim that no evidence could contradict is not a finding.
The ladder metaphor then does not merely fail to describe this; it actively misdirects the response to it. If we believe safety develops in sequence, we intervene sequentially: shore up “inclusion”, then “learning”, then “contribution”, and challenge will follow. But if the real constraint is, say, the threat of a public inquiry and negative newspaper headlines, then regardless of how warm and inclusive the team feels, the sequential intervention is not just inefficient. It is aimed at the wrong layer entirely, and the leadership comfort it generates often masks the problem it was meant to solve. The frame we adopt determines what we see, and a frame that says “ladder” makes the sociotechnical, ecological, multidimensional reality of psychological safety harder to see, not easier.
Psychological safety is complex, multidimensional and both collective and individual. The Four Stages structure is not a simplification of reality, it is a distortion of it.

Stage Three: The Complexity Mismatch
There is a deeper version of the structural problem, and it concerns the kind of phenomenon psychological safety actually is.
Some systems are merely complicated. They have many parts, but the parts relate to each other in stable, predictable ways, so that with enough expertise we can diagnose the problem, apply the correct intervention, and predict the result. A model that says “do these things in this order and you will get this outcome” is well suited to a complicated system such as a gearbox, an electrical circuit or a Rubik’s Cube.
Psychological safety does not live in a complicated system. It lives in a complex one: non-linear, emergent, sensitive to context and history, prone to producing outcomes that were not intended and in which the same intervention, a second time, produces different results. The same intervention (let’s say, visualising work) may facilitate candour in one team and suspicion in another, not because it was applied “wrongly” but because the teams and their contexts were different. In a complex system, the relationship between what we do and what happens is probabilisitc, and any model that promises a predictable path from intervention to outcome is selling a promise that the underlying reality cannot deliver (though I must point out that this doesn’t excuse inaction).
This is the trap, and it is the third kind of wrong from Snowden’s taxonomy: not a model that is incoherent, but a model applied outside the domain where it works. A stage model applied to a complex phenomenon creates a false sense of control. It tells leaders that psychological safety is a project with milestones, that progress can be tracked against deliverables, that stage three is behind them and stage four lies ahead. The map looks so much more navigable than the territory that we begin managing the map – comfortingly believing that it is the territory. And when the team’s actual, messy, non-linear experience stubbornly refuses to match the diagram, the conclusion is often that we did it wrong (or they – the team themselves – did), rather than the more simple and correct answer – the model was never going to fit.
Stage Four: The Leader as Gatekeeper
Outside of Snowden’s taxonomy, there’s another problem with the model; in its politics.
The Four Stages places the leader at the centre of the field. It is the leader who cultivates safety, the leader who extends respect, and — in the model’s own terminology — the leader who grants permission. Psychological safety, in this telling, flows downward from the person with the power to bestow it. The whole apparatus is organised around what those in charge do for those who are not.
I should be precise here, because it would be easy to overshoot. Leaders do have a role, and a large one at that. The behaviour and decisions of people with power often shape the conditions for psychological safety more than anyone else, and pretending otherwise would be dishonest. The problem is not that the model involves leaders. The problem is that it positions the leader as the architect and grantor of a thing that, by its own account, reaches its summit in the freedom to challenge that very leader. The model itself describes contributor safety as offered “in exchange for effort and results”, and challenger safety as “air cover” offered “in exchange for candor”.
Nor is the model unusual in centring leaders; Edmondson’s own toolkit is largely a list of leader behaviours. The difference is in the verb. Edmondson’s leader frames, invites and responds: she works primarily on conditions. Clark’s leader grants. Safety that is cultivated can outlast the person who cultivated it; safety that is granted is held on licence.
In a model where safety is something authority dispenses, permission to challenge is itself granted by the person being challenged. The summit of the model depends on the leader voluntarily authorising dissent against themselves. The model thus reproduces, at its highest level, the very power asymmetry it claims to dissolve. A framework for distributing the capacity to speak truth to power turns out to route that capacity through power’s permission, and a permission that can be withheld is not psychological safety. It is a licence, and licences are revocable.
None of this is particularly hard to see. Mary Parker Follett saw it over a century ago: the difference between power-over, dispensed downward by those who hold it, and power-with, which is co-active and develops jointly between people. Her “law of the situation” is power-with (and predates later concepts like Gemba and deference to expertise): authority that derives from what the circumstances genuinely demand, rather than from rank, so that the person who defers is responding to the situation and not to status. We cannot create power-with by perfecting the dispensation of power-over, and a model whose summit is “the leader permits dissent” has not flattened the power gradient; it has merely made the leader at the top of it appear gracious.. Follett’s point was sharper: the goal is not for authority to permit challenge but to dissolve the premise that challenge even needs permitting.
Stage Five: The Equity Problem
As we’ve just seen, in the Four stages, the two drivers of “respect” and “permission” are things conferred by those with more power on those with less. You are respected; you are permitted: they describe a transaction in which psychological safety is something the privileged extend to the disadvantaged, conditional on the goodwill, attention, and continued approval of the person doing the extending.
A team is only as safe as the least safe person. The Four Stages model has little concept of deeper organisational substrate or intersectionality: structural power, individual upbringing, cultural (particularly non-western) defaults, neurodiversities, and the socioeconomic conditions that existed before the team even formed. It operates entirely in the interpersonal here and now, as though every team exists only in that context at that time, rather than bringing all our baggage with us. But for the woman who is talked over in meetings by a man, for the neurodivergent employee whose access needs are treated as an imposition, for the minoritised worker whose accented speech is read as lower competence, for the junior member of staff from a background that taught them deference to authority as survival — for these people the constraint was never a shortage of respect or permission.
The constraint is that the cost of speaking is borne unequally. To be fair, the model’s more recent material does name this: hierarchy, tenure, upbringing, socialisation are all acknowledged as setting each person’s “price”. It names the substrate and then does nothing with it. Every remedy on offer is interpersonal: a phrase for the leader to use, a question to ask by name, a one-to-one to book. The conditions that set the price are noticed and then left exactly where they were. It assumes the only barrier is whether the leader has opened the door, when for a great many people the door was never the problem. Many people didn’t even know that there was a door at all.
Psychological safety is impacted most where power gradients are steepest, and a framework built on respect-and-permission reinforces power gradients rather than reduces them. Granted, the model may be well fitted to a relatively homogeneous, flat, equally privileged team in which the main obstacle to speaking up is that no one thought to actually invite contributions. Those cases, however, are somewhat rare. Conditional psychological safety, especially for the people who most need it, is not psychological safety at all
Stage Six: The Iatrogenic Risk
Which brings us to the most serious point, and the one the previous five have been building toward. It is not that the model is empirically thin, structurally confused, mismatched to its domain, leader-centric, or blind to structural power, though it is all of those. It is that being wrong in these particular ways does not fail neutrally. Under certain conditions, its simplicity and apparent success can make the underlying problem harder to see. It thus risks harming the people it claims to support, while signalling to everyone that the opposite has happened.
If an organisation adopts the Four Stages, runs some assessments, delivers some workshops, and writes a report, leadership, reasonably enough, may conclude that the work has been done. The box on psychological safety is ticked and attention moves on to the next priority.
But the people that the Four Stages model serves least well (those who were already least safe – including neurodivergent, LGBTQ+, and people who don’t fit the Western cultural ideals that the model assumes), are not psychologically safer. In many cases they are less safe, because the organisation has now officially concluded The Problem Is Solved. In this way, it doesn’t just harm the people in the organisation subject to the programme, but it harms the field of psychological safety itself, through damaging the credibility of the concept and distorting it away from the original and true meaning.
This is the iatrogenic stage: harm caused by the treatment itself. It’s worse than doing nothing, because doing nothing at least leaves us aware the problem is still there. The Four Stages doesn’t just fail to help the least safe, but risks removing them from view, whilst congratulating leadership on a job well done.
We haven’t conducted controlled trials on this; that literature barely exists for any of the model’s claims, in any direction. What I’m describing is what the mechanism predicts, and what we see repeatedly in our own work with organisations. A meaningful share of the organisations that come to us arrive because the Four Stages didn’t work for them, or made things worse: scores improved while reality did not, and the gap was carried, as usual, by the people with least capacity to carry it. I’m aware of the sample bias: nobody calls us to say their Four Stages programme went great. So this can’t tell us how often it happens. What it does show is that when it happens, it happens in the way the mechanism predicts, and it lands on the people the model was least built for.
After the Stages
None of this means we should never use the Four Stages, though we must use it with great caution. A model held lightly, in full knowledge of its limits, can still open a useful conversation, and I would rather a team talked about inclusion safety than did not talk about safety at all. It can also be useful to explore many different models, critically and exploratively, to see what lands with us and what doesn’t.
But the Four Stages is a misleading metaphor, the complex thing it describes will not march in step with it, and safety dispensed by authority is not the same as nurturing the conditions in which psychological safety can emerge for everyone.
Psychological safety is better understood as a shifting profile than a ladder: nebulous, multidimensional, relational, context-specific and unequally distributed. It varies by risk, person, audience, history and power gradient. We should examine patterns, trends and distributions rather than assign teams to stages; we should attend to the deeper substrate as well as just leader behaviour; and we should treat interventions as probes from which to learn, rather than steps in a predetermined sequence of milestones.
All models are wrong, but they are not all wrong in the same way: there is the forgivable wrong of the honest simplification, there’s the wrong of the incoherent model, and there’s the wrong of the model dragged far outside its relevant domain. And then there are ethical and political wrongs, the ones that matter most: resulting in a model that hurts the people it claims to protect, while telling everyone it has done the opposite. The Four Stages manages all of these. “All models are wrong” can forgive the first of them, but cannot forgive the last.
The post The Four Stages of Psychological Safety: a Six-Stage Critique appeared first on Psych Safety.







