When warning everyone is the same as warning no one
How I used behavioural science to solve warning fatigue in a consulting brief presented by the UK Met Office.
Earlier this year, my team and I were given a brief by the UK Met Office: improve the effectiveness of severe weather warnings.
The problem
The UK's severe weather warning system is built on a sophisticated risk modelling framework. Variables like hazard, vulnerability, and exposure are independently scored and combined across 8 weather types, namely, rain, wind, snow, ice, fog, thunderstorms, lightning, and extreme heat. Each of these produces its own matrix. Then, these are all compressed into one, publicly visible matrix with just 2 axes: likelihood and impact.
This compression is, of course, necessary. Nobody wants to parse 8 separate hazard models before deciding whether to bring an umbrella. However, this does come at a cost. Users often complain that weather alerts feel irrelevant. This may be the case because the underlying matrix is not personally relevant to many users. A quick glance at the app gives everyone the same signal, regardless of whether it actually applies to them.
Moreover, yellow weather warnings constitute the majority of all warnings issued. According to user feedback, these are the ones most consistently ignored.
Target behaviour: when a yellow warning is issued, recipients 1) pay attention to the warning, and 2) understand its personal relevance and take appropriate precautions.
Diagnosing the problem
We used the COM-B model to locate where the system was breaking down. Based on our analysis, we designed our intervention to target two capability problems: limited attention and high cognitive load, and difficulty interpreting risk and relevance.
In other words, the high frequency of alerts, paired with low personal relevance often leads to cognitive overload and habituation. In psychological literature, this is known as "warning fatigue" and the "cry wolf effect". As a result, people stop reacting to warnings when they receive too many that don't seem personally relevant to them.
Warning fatigue: a state of exhaustion produced by repeated alert exposure, leading people to desensitise, ignore, or opt out of communications entirely.
Cry wolf effect: when a high volume of warnings is followed by no adverse outcome, people learn that warnings overstate risk. This learned discounting reduces emotional arousal and perceived urgency the next time a warning fires, even when the risk is genuine.
The intervention: personalisation as a filter
Our intervention was a personalised warning notification system where users can input information about their daily routines, timings, and commute routes. At onboarding, users answer a short survey about their routine: "Do you commute daily?", "Do you live somewhere flood-prone?", "Do you spend time outdoors?", and so on. Their answers route them into relevant, back-end "packs" such as a "Driver", "Storm", "Outdoor". Each of these packs filter yellow warnings against custom likelihood and impact thresholds that the user set for themselves. For instance, a commuter who leaves home at 7:30 AM would receive a targeted alert at 7:15 AM flagging fog and ice risk specifically on their route. Red and amber warnings remain unchanged.
Why this works: self-determination theory
The theoretical backbone here is self-determination theory. When users actively choose what they receive, they retain a sense of autonomy over the interaction. Essentially, this turns the yellow weather warning system into something they have personally opted into and shaped for themselves. This makes the warnings more personally relevant to users because information that matches a person's own stated preferences is more likely to be processed via System 2, i.e., conscious, deliberate cognition, rather than be filtered out through the automatic and heuristic-based shortcuts of System 1. In other words, people pay more attention to things they have personally chosen to pay attention to.
We also used the EAST framework to ensure our intervention is structurally sound.
- Easy: push notifications make the behaviour easy by removing the need for proactive checking.
- Attractive: personalisation makes it attractive by matching content to individual preferences.
- Social: descriptive norms increase opt-in and compliance. For example, "78% of commuters in your area report concerns regarding flooding".
- Timely: synced, routine-based delivery time makes the information actionable right when it is required.
Potential limitations
To ensure we paint a holistic picture of this problem and its solution, we mapped the intervention's downstream effects using a ripple map. Alongside the intended benefits of reduced cognitive load, higher relevance, and lower habituation, we flagged potential risks.
Firstly, the personalisation filter could create a false sense of security if it filters out warnings users didn't anticipate needing. Nevertheless, we attempted to account for this drawback by only applying the personalisation mechanism to yellow warnings and leaving the red and amber ones unchanged. Secondly, the higher amount of personally relevant information this mechanism would need may raise concerns around privacy and data security.
The key takeaway
Simplification and personalisation solve very different problems. Conflating the two seems to emerge as a common design mistake. The Met Office's existing matrix serves as an excellent simplification as it compresses an enormous amount of risk modelling into something that is legible by users who pay enough attention. However, legibility for the average recipient is not the same as legibility for each, specific recipient. Once a warning system reaches the frequency and generality that the UK's has, the next level may be better-targeted delivery.