In our recent publication, led by Dr Patrick Curran, we examined this problem through a review of 39 climate risk and vulnerability assessments from cities around the world, using sea-level rise as a case study. We looked at how uncertainty was represented in the assessment itself, and then how that uncertainty was used when making policy recommendations or planning decisions. 

To do this, we used the Levels of Uncertainty framework. It distinguishes between:

• Level 1: a sufficiently well-defined future that can be represented using a single value or deterministic model

• Level 2: uncertainty that can be characterised statistically

• Level 3: multiple plausible futures that can be represented using scenarios

• Level 4a: deep uncertainty, where only a range of plausible outcomes can reasonably be specified

• Level 4b: situations where even the range of possible futures cannot be adequately characterised.

This distinction matters because these different forms of uncertainty require different decision approaches. A Level 2 problem may reasonably be addressed using probabilistic analysis. A Level 3 problem lends itself to scenario analysis. At Level 4, the emphasis shifts toward robustness and adaptation: testing whether a decision performs adequately across plausible futures, identifying when it may fail, and allowing the strategy to change as conditions evolve.

The figure below from the Unified Sea Level Rise Projection for Southeast Florida provides a useful example.

Within the same assessment, sea-level rise is represented in four different ways. A single recommended projection represents Level 1 uncertainty. The IPCC median provides a statistical representation consistent with Level 2. The NOAA scenarios represent multiple plausible futures at Level 3. The shaded range between projections represents Level 4a uncertainty. 

There is nothing inherently wrong with using different representations for different purposes. The problem arises when the uncertainty recognised in the analysis is removed in the decision.

For infrastructure decisions, the Southeast Florida guidance ultimately directs decision-makers to select a projection value between the IPCC Median and NOAA High curves. In our classification, that decision therefore returns to Level 1, despite the assessment having presented the problem as Level 4a. 

We saw the same pattern across assessments and plans globally. Just over half of the assessments, 54%, correctly characterised sea-level rise as deeply uncertain. However, when those assessments translated into decisions, 71% relied on a single projection. In other words, many assessments recognised uncertainty in the analysis, but then effectively removed it when a decision had to be made. Only 8.3% of decisions used approaches that adequately accounted for deep uncertainty.

The methodological problem is therefore more specific than simply “failing to acknowledge uncertainty”. In many cases, the uncertainty is present in the science and even in the assessment, but is progressively simplified as it moves toward the decision.

That is important because climate change contains uncertainties that cannot just be eliminated through better data or more sophisticated modelling. Future sea-level rise depends on interacting uncertainties in emissions, ice-sheet behaviour, socioeconomic development, technology and policy. Some of these uncertainties will narrow over time. Others remain irreducible over the planning horizon.

Planning approaches that nevertheless select a single expected future are often described as “predict-and-act”: estimate what the future will look like, optimise the decision for that future, and proceed on the assumption that reality will remain sufficiently close to the projection.

There are many examples of the problems this creates. Melbourne’s water supply was modelled as having an adequate buffer through to 2020 under what was described as “severe climate change”, yet in the three years following the modelling, water supply failed to exceed half the long-term average and later reached record lows. Dresden demolished inner-city housing based on projected population decline just as population began recovering. The Channel Tunnel was planned around projected first-year ridership of roughly 16 million passengers, while actual ridership was around 7 million.

The common problem is not that forecasts sometimes turn out to be wrong. Any model can be wrong. The problem is designing a decision whose success depends heavily on one particular representation of an uncertain future.

One reason this happens in climate planning appears to be the way scenarios themselves are described and used. RCPs, SSPs and similar climate scenarios are exploratory tools. Their purpose is to describe alternative plausible futures and allow analysts to investigate how systems and decisions perform under those conditions.

Yet 87.5% of the documents in our review that referenced RCPs described them as “projections”.

That terminology matters because it changes the implied analytical task. If an RCP is treated as a projection, the natural response is to ask which projection is most likely and then design around it. If it is treated as a scenario, the question becomes: what happens to this decision if the world develops this way?

The recent change in climate scenarios provides a useful example. As I discussed previously, SSP5-8.5 has been replaced as the upper-end marker in the new CMIP7 scenario architecture because assumptions underpinning that pathway are increasingly considered less plausible. For adaptation planning, however, the physical conditions associated with SSP5-8.5 have not ceased to matter. A particular amount of sea-level rise may still occur under another pathway, although perhaps later. An increment reached around 2100 under one scenario could, for example, be reached one or two decades later under another. This distinction matters for infrastructure with a 50-, 80-, or 100-year life. 

This is why adaptation analysis is often more useful when structured around physical conditions, thresholds and decision failure points, rather than scenario labels and fixed years. A road may become regularly impassable once relative sea level reaches a particular increment. A stormwater system may fail when rainfall intensity exceeds a particular threshold. An insurance market may begin withdrawing when losses occur with sufficient frequency. The analytical task is then to determine whether those conditions are plausible, when they might emerge, how much warning we are likely to have, and what action should follow. If the timing changes as climate scenarios are updated, the decision framework can be updated without discarding the underlying analysis. 

The consequences of getting this wrong run in both directions. Designing around a future that proves too severe can lead to premature investment, opportunity costs and unnecessary disruption. Designing around a future that proves too mild can leave infrastructure and communities exposed to conditions outside the design envelope. Selecting the highest scenario does not solve the methodological problem. It effectively substitutes one deterministic future for another.

Methods for dealing with this type of uncertainty are well established. Decision Making under Deep Uncertainty includes approaches such as Robust Decision Making and Dynamic Adaptive Policy Pathways. These methods test how strategies perform across multiple plausible futures, identify vulnerability or failure thresholds, and allow decisions to change as new information becomes available. The two cases in our review that did incorporate these ideas illustrate both robustness-based and adaptive approaches.

Yet only 8.3% of the decisions we reviewed adequately incorporated deep uncertainty.

The practical implication is fairly simple. Climate adaptation should not begin by asking which projection is correct. It should begin by asking which uncertainties materially affect the decision, under what conditions the preferred option stops working, and what flexibility is available if those conditions emerge.

That keeps the representation of uncertainty consistent with the decision problem rather than progressively removing it as the analysis moves toward action.

Paper: Curran, P., Hardaway, K., & Logan, T.M. (2026). Lost in Projection: Uncertainty is Misrepresented in Climate Risk and Vulnerability Assessments. Risk Analysis, 46, e70331. https://doi.org/10.1111/risa.70331