Why People Don’t Speak Up: The Calculus of Voice

Understanding what psychological safety is requires understanding why, in its absence, people stay silent even when they have something important to say. This is the psychological mechanism beneath speaking up.

Before “speaking up” (and speaking up may not be verbal – it may be a text, an email, a note, sign language, a nod, or something else), we run a rapid, usually unconscious risk calculation: what are the likely costs of speaking, and the likely benefits? It is fast, affective rather than deliberate, and it is running continuously, updated in real time as we read the room. It also doesn’t produce a simple yes or no. It has four possible outputs: say the thing fully; stay silent; soften and dilute what we meant to say; or change it entirely to please the listener. That third option — softening — is the most common, and it is exactly the move a junior officer makes when, instead of “stop, this is unsafe,” they ask a tentative question and hope it draws attention to the thing they’re concerned about.

Positive interpersonal predictability

To run that calculation at all, we need something to run it on: a prediction of how our vulnerability will land. This is why one of the most useful ways to understand psychological safety is as positive interpersonal predictability — not a vague sense of comfort, but enough confidence in how a group will respond to make the risk worth taking. We are, in other words, usually not managing danger so much as navigating uncertainty, which is why the word “safety” can mislead.

That prediction runs on data: what we’ve seen happen to others, how this person has responded before, the norms of the group, and (when local data is thin) more distant reference points like a previous boss or what we learned long before we ever entered this workplace. This is why telling a team “this is a safe space” rarely works: the declaration generates no data, and data is what the calculus needs. It also explains the counter-intuitive case of the hot-and-cold manager, who is in some ways harder to work with than a consistently harsh one. A predictably negative boss can at least be modelled and planned around; an unpredictable one resists modelling entirely, and under that kind of ambiguity our default is simply not to take the risk.

Two gates: ambiguity, then valence

This gives us two distinct points at which voice can fail — and crucially, they are sequential, not parallel:

• Gate 1 — ambiguity. Do we have enough data to form a confident prediction at all? This is about making the implicit explicit: clear norms and contracts, defined expectations, team charters, shared understanding of what is sayable here.
• Gate 2 — valence. Given that prediction, does it lean positive enough to be worth the risk? This is about reducing the real costs of speaking and increasing the benefits, demonstrated repeatedly over time.

If Gate 1 doesn’t open, we never reach Gate 2 — with no prediction to weigh, the calculus can’t run, and silence is the safe default. So a lot of interventions misfire by treating an ambiguity problem as if it were a valence one: assuming people expect a bad response, when in fact they simply can’t predict the response at all. The useful diagnostic question is therefore: is this an ambiguity problem or a valence problem? — because the fixes are different. Gate 1 is largely structural and, once built, fairly durable; Gate 2 is built slowly through consistent behaviour and can be eroded quickly.

Power Gradients 

This is where power gradients become critical. The steeper the hierarchy — the greater the status difference between the person with something to say and the person they need to say it to — the higher the perceived cost of speaking. Decades of work in aviation safety, where this dynamic was called “authority gradient” or “cockpit gradient”, show that power differentials reliably suppress voice, and that the consequences of that suppression can be catastrophic.

Predictability is also unevenly distributed. Someone with tenure and informal networks has a rich predictive dataset; someone newer, more junior, or from a different background has fewer reference points and usually pays a higher price for getting it wrong. Making norms explicit isn’t only good practice, then — it is redistributive, transferring predictive capacity to the people who have least of it. And note where the mechanism sits: the calculus runs inside the individual, but the conditions that load it are structural. Locating the mechanism in the person is not the same as locating the responsibility there. The work is to reduce the cost of speaking up, not to coach people to be braver in conditions that haven’t changed.

We go deeper into the calculation itself in our Calculus of Voice article, and into predictability as its essential precursor in Ambiguity, Predictability and Psychological Safety.