Predictability and the “calculus of voice”

We’ve previously covered the “calculus of voice”: the mechanism by which we decide whether to speak up, stay silent, soften what we were going to say, or say something else entirely. This mental risk calculus is an ongoing, dynamic process, not a one-time assessment of conditions – we’re continuously updating our predictions within interactions, not just between them. It’s usually fast, unconscious, and affective rather than deliberate.

Some models of this cost/benefit calculation – what Detert and Burris call the “affect-laden expectancy-like calculus” – typically assume an even weighting of the costs of speaking up versus the benefits. However, Prospect Theory, as we’ve discussed previously, shows us that the calculus will tend to lean towards the costs due to our inherent loss aversion. This is not a quirk of particularly anxious individuals – it applies to almost all of us as human beings, to varying degrees.

But there’s another crucial layer missing to this voice calculus concept.

Positive interpersonal predictability

One useful way to think of psychological safety can be as “positive interpersonal predictability”. We take interpersonal risks not because we feel vaguely comfortable, but because we can run a rapid mental risk calculus, which requires being able to predict how our vulnerability will land. Positive interpersonal predictability names the actual mechanism by which we may have high or low confidence that the response will be good enough to justify the risk.

This helps to explain why people don’t speak up even in ostensibly “friendly” environments, because if they can’t predict the response with enough confidence, the safest, rational option is to stay quiet. This is why “psychological safety” as a label sometimes misleads, because the word “safety” imports connotations of threat and protection that aren’t quite right. In practice, we’re often not primarily managing danger: we’re navigating uncertainty. And that helps us realise that ambiguity is one of the biggest destroyers of psychological safety.

Before we speak up with an idea, a question, challenging something, or admitting an error, we attempt to estimate the risk of a negative outcome. And to make this calculation, we need data. That data might take the form of past experiences with the people in the group, for example what we’ve seen happened to others when they took interpersonal risks, or even what we’ve heard happened to them. The data might also be our own past experiences in previous workplaces or our childhoods, and many other things too. In the absence of direct contextual information such as how this person has responded to me in the past, I’ll fall back on more distant, less directly connected information, such as how my last boss behaved, instead. And whilst I may use that less relevant data for my risk calculus, as a result I have much less confidence in the prediction I’ve made.

Ambiguity is one of the biggest killers of psychological safety.

We may, of course, be highly confident of a negative outcome. For example, I remember an old boss I had who would always lay into people for making mistakes or not following procedure. The benefit for me (if you could call it that) in this case was that I had a high degree of confidence in how he’d react to me raising a concern – which meant that I was able to plan and prepare for that response. In this situation, at least I had some agency. 

A potentially worse situation to be in is where we have a boss who may either react very well or very badly, but we can’t predict which – because in this case, we can’t even attempt the risk calculus in the first place. The hot-and-cold manager isn’t just unpleasant, they are actively creating the worst possible conditions for interpersonal risk-taking, because their behaviour resists modelling. In these conditions, we can’t build a prediction and our tacit calculus can’t be run. Our default behaviour under unresolvable ambiguity is very likely, don’t take the risk. Ambiguity removes agency because it removes the inputs to decision-making.

Of course, there is a distinction between our confidence in prediction and how accurate that prediction is – someone can be confident about how they’ll be received, and still be wrong. But the point stands – for our prediction to be accurate, we need good data to work from. 

This is why telling people “This is a safe space” rarely works. The declaration of a “safe space” doesn’t actually generate meaningful data, and data is what our mental calculus runs on. If anything, the need to announce it might itself imply that if you have to say it’s safe, it probably isn’t. What we need isn’t an assertion of safety; it’s enough consistent observational data to form our own accurate prediction.

Ambiguity aversion

In 1961, economist Daniel Ellsberg (later famous for the Pentagon Papers) described a thought experiment that illuminates exactly this. Imagine an urn containing 90 balls: 30 red, and 60 that are some unknown mix of black and yellow. We have to bet on which colour we’ll pull out. Most people choose red even though red is outnumbered two to one. Why? Because with red, we know the odds. With black or yellow, we don’t. Given a choice, we’d typically rather know the odds, even if they aren’t that great, than face uncertainty about what they are.

This is ambiguity aversion, which is distinct from, and additional to, loss aversion. It isn’t just that people dislike bad outcomes. We dislike situations where we can’t even form a reasonable estimate of what the outcome might be. The hot-and-cold manager is an Ellsberg urn. The employee isn’t just facing a potentially bad outcome: they’re facing unpredictable probabilities, which, combined with loss aversion, can trigger a more paralysing form of risk avoidance than even a known-negative environment (the “consistent asshole” manager) would. 

Essentially, in some cases, we may actually prefer a predictable, slightly negative outcome, over an uncertain, potentially positive one. 

The power gradient complication

The data available for making predictions, and the cost of getting them wrong, aren’t the same for everyone, so it’s worth considering how this interacts with power. 

Someone with institutional standing, tenure, and a personal relationship with leadership likely has a good amount of power in an organisation, and they have also accumulated a rich predictive dataset. Someone newer, more junior, from a different cultural background, or without access to informal networks is operating with far less data – not necessarily because the environment is hostile to them, but because they have fewer reference points to make the calculation. More importantly, for that person who also likely has less overall power in the organisation, the stakes of getting the calculation wrong are often higher. Predictability is unevenly distributed, and that uneven distribution maps closely onto existing power gradients. Accessing predictability, it turns out, is a privilege. And like most privileges, it’s often invisible to those who have it.

This is why making norms explicit and adopting low-context communication, where we don’t assume too implicit shared understanding, aren’t just tools for building team performance, they’re fundamentally redistributive. They help transfer predictive capacity from those who already have it to those who don’t.

The 2-Gate Model of Psychological Safety: Increasing interpersonal predictability

If we think about the elements of our mental process when we make these calculations, we effectively weigh up “how much data do we have?”, and “what does that data tell us?”  

This framework thus suggests two distinct targets for intervention, which we’ll call “gates” (we’ll see why in a moment):

Gate 1 is ambiguity reduction – making the implicit, explicit. This is about how much data we have and our confidence in it. It lets me know whether I can form a confident enough (or indeed, any) prediction about the result of this interpersonal risk

Gate 2 is shifting the calculation positive. This is the actual calculation that I run, which lets me know whether that prediction leans positive enough to clear my bar for loss aversion.

If Gate 1 doesn’t open, we don’t even get to Gate 2. And both gates typically have to open in order for us to take the risk.

For Gate 1, our interventions are all about increasing predictability – giving us more opportunities to gather data, and make implicit norms and expectations more explicit. They include artifacts and practices like co-created social contracts and team charters, well defined group norms, definitions of “done”, PACE, Ladders of Participation, Lean Coffee, the Andon Cord. These things don’t necessarily change whether the environment is kind or punishing (we could feasibly have a social contract that says “be horrible to each other”!), but they do make it legible, and more predictable. That predictability is the first precondition for psychological safety.

Gate 2 is making those predictions as positive as possible, reducing interpersonal costs, and increasing the benefits of speaking up. This is categorically harder. We can design in opportunities for better legibility but we can’t enforce positive responses from everyone to interpersonal risk taking.  What we are aiming for is repeated demonstrations over time that it is in fact safe for people to take interpersonal risks. In practice, this means not just having a social contract, but making sure the agreements are as supportive as possible of interpersonal risk taking, and that we act decisively if it is breached.  It also includes following through when issues are raised so people see that they matter, being mindful of how we respond, appreciating diverging perspectives and contributions, praising people for admitting mistakes, reframing work as learning, and many more. Every time these occur, they provide a helpful, positive social proof data point that updates the next prediction.

Of course, silence produces less data; less data increases ambiguity; and more ambiguity produces more silence. That’s a reinforcing feedback loop, and like all reinforcing loops it accelerates in whichever direction it’s already moving. High psychological safety environments tend, over time, to become safer, because interactions produce data, data reduces ambiguity, reduced ambiguity encourages more interactions. Low psychological safety environments tend to become less safe for exactly the same reason in reverse. This is why deliberate intervention matters: left alone, the system will amplify whatever condition it’s already in.

It also helps explain why Gate 1 and Gate 2 interventions have different characters. Gate 1 interventions are largely structural: explicit norms, and defined expectations and processes. Structural interventions, once in place, tend to be reasonably durable. Gate 2 interventions are more like changing a flow over time, through consistent demonstrated behaviour, which accumulates slowly and erratically and can be eroded quickly. Both matter; but they operate at different speeds and with different leverage.

Thinking in a new way about psychological safety

This matters practically because it suggests two distinct failure modes in low psychological safety environments with different remedies:

The first is failing to open Gate 1. This is Ambiguity Failure: we don’t have enough data to perform the calculation. This failure to have enough (or any) data may occur because we’re in a new team, working under unfamiliar conditions or are in an environment consisting of highly unpredictable characters. 

The second is failing to open Gate 2 – Valence Failure: we have enough data, but the result is negative or neutral (valence is a psychological term that describes the attraction or repulsion one feels towards something).

Valence Failure isn’t only about clearly negative predictions. Even a neutral predicted response (where the cost and benefit are equal) tends, under loss aversion, to register on the negative side of the bar, i.e. a neutral result is rarely safe enough. The prediction has to lean significantly and confidently positively to clear the bar (and that bar is set differently for each of us, by history, culture, and neurodiversity as well as the immediate context).

A lot of psychological safety interventions treat Ambiguity Failure as if it were Valence Failure – they assume that the problem is that people expect negative responses, when in reality, maybe they just can’t predict the outcome at all. We should therefore remember to ask: “is this an ambiguity problem or a valence problem?”, because the interventions for the two are very different. 

This framework proposes something the psychological safety and voice literatures to date have not explicitly done: connect the calculus by which people decide to speak up with the structural conditions that make speaking up possible, and names predictability as the missing mechanism connecting them.

References:

Berger, J., Cohen, B.P. and Zelditch, M., 1972. Status characteristics and social interaction. American Sociological Review, 37(3), pp.241-255.

Ellsberg, D., 1961. Risk, ambiguity, and the Savage axioms. The quarterly journal of economics, 75(4), pp.643-669.

Detert, J.R. and Burris, E.R., 2007. Leadership behavior and employee voice: is the door really open?. Academy of management journal, 50(4), pp.869-884.