Decision Making Under Deep Uncertainty

In 2008, TEPCO's own engineers modelled what a repeat of the 869 AD Jōgan earthquake would do to the Fukushima Daiichi nuclear plant. The result was clear: tsunami run-ups exceeding 15 metres against a seawall designed for 5.7. TEPCO sat on the study for three years. On 7 March 2011, four days before a 14-metre wave destroyed backup power and triggered triple meltdown, they finally reported the findings to regulators. The information existed. The decision making under deep uncertainty failed.

Decision making under deep uncertainty means deciding when probabilities are too weak or too contested to justify the weight placed on them.

Sufficient certainty to act comes from judgment about assumptions, not from better forecasts. TEPCO had information. Their own 2008 modelling confirmed the risk. Chubu Electric had already hardened a neighbouring plant against the same scenario, proving that mitigation was feasible. TEPCO lacked the willingness to surface an assumption and ask how significant it was.

What decision making under deep uncertainty demands

TEPCO's probabilistic risk assessment had estimated a wave exceeding 10 metres as a once-per-100,000-to-1,000,000-year event. That number became the decision. A probability estimate, treated as settled fact, replaced every question about what the plant was actually vulnerable to. That estimate was somebody's work product. The engineers who produced it had no incentive to revisit it, and the regulators who accepted it had every incentive to leave it alone.

Click to expand

When their own engineers produced evidence that contradicted the estimate, the estimate won. The decision model had replaced the judgment it was supposed to inform. The Carnegie Endowment analysis published a year after the disaster made the point that should have been obvious beforehand: the accident was preventable: the existing information was sufficient, if anyone had been willing to ask what it meant for the assumptions underpinning the plant's design.

I have watched this happen dozens of times. Not always with consequences measured in meltdowns, but always with the same logic: the model exists, so the thinking is done. Deep uncertainty is the condition where that substitution breaks down most visibly, because the probabilities are too weak to carry the weight placed on them. The response is to stop trusting the model and return to what actually works: surface the assumptions that carry the decision, then monitor what matters after you act.

Planning for a century without a forecast

The Netherlands offers the sharpest contrast to TEPCO's failure, and the best large-scale example of decision-making under uncertainty I have encountered anywhere. Sea level rise projections for 2100 range from just over one metre to the IPCC's non-excluded scenario of two metres by 2100 and five metres by 2150. The Dutch response was not to pick a number and build to it.

The Delta Programme, established by the Delta Act in 2012, uses Dynamic Adaptive Policy Pathways: 14 adaptation strategies with monitoring triggers that signal when to switch from one pathway to another. The Delta Fund commits over one billion euros annually, with €29 billion allocated to 2050. They are planning infrastructure to 2100 and beyond, not because they know what will happen, but because they have structured their decisions to work across multiple futures.

The approach works because it separates two questions that most organisations conflate. What do we need to do now? And what will we need to do if conditions change? Most decision-makers try to answer both with a single forecast. The Dutch answer them separately. The first gets funded and built. The second gets monitored, with defined triggers: measured sea level data, storm surge records, subsidence rates. When a trigger fires, the programme shifts to the next pathway without waiting for political consensus. Government agencies and corporations spend comparable sums on risk registers that have never triggered a single change in direction. Their apparatus is designed to produce a number and file it. When context shifts, nobody notices until the seawall is underwater.

Roger Estall and I built the Universal Decision-Making Method around this same logic. The Dutch proved what it looks like when an entire country takes it seriously.

When your forecasts disagree by a factor of nine

In August 2021, the United States Bureau of Reclamation declared the first-ever shortage on the Colorado River. Traditional climate models had been producing contradictory projections for years: annual flows ranged between 5 million and 45 million acre-feet. A factor of nine. At that point, probabilistic forecasting is not uncertain. It is useless for making the operational decisions that determine whether 40 million people have water.

For post-2026 river operations, the Bureau tested operating strategies against 8,400 hydrologic scenarios, looking for strategies that performed acceptably across multiple futures rather than optimally for one predicted future. The annual basin deficit stands at 3.6 million acre-feet and growing. Lower Basin states committed to cutting 1.25 million acre-feet in 2024. Carly Jerla of the Bureau put it plainly: "We need something else" beyond the interim policies of the past, because the old approach was built on assumptions nobody had examined.

This is what decision making under deep uncertainty looks like when someone finally acts on it. The Bureau arrived at what I have argued for decades, applied at continental scale. I grade more cases in the scenario planning examples piece, and the same discipline holds: the useful ones left an owned trigger behind.

Decision making under deep uncertainty is where the method proves itself

The organisations that handle deep uncertainty well follow the same pattern. They surface the assumptions embedded in their plans and judge which ones matter. They reach sufficient certainty to act, then monitor for the moment those assumptions break down. The decision-making process under uncertainty walks through each step in sequence.

The ones that handle it badly build a model, treat its output as a fact, and stop thinking. TEPCO is the extreme case, but not the unusual one. I have seen the same substitution in boardrooms where a Monte Carlo simulation replaced a conversation about what actually mattered, and in government agencies where a risk register became proof that someone had thought about uncertainty. The example cases across Apollo 13, Blockbuster and Grenfell show these three failure modes side by side.

In Deciding, we described this as vulnerability: the exposure created by the Decider's own choices, not by external events. An earthquake does not make a nuclear plant vulnerable. The decision to build the seawall at 5.7 metres, and to suppress internal evidence that 5.7 was insufficient, creates the vulnerability. A drought does not make a water system vulnerable. The decision to plan for one predicted future, ignoring models that disagree by a factor of nine, creates it.

Every organisation faces deep uncertainty of some kind, whether the uncertainty sits in market conditions or environmental systems. The question is not whether your Context will change but whether you have designed your decision to detect that change when it comes. This is what separates genuine decision making under deep uncertainty from forecasting dressed up as due diligence. Your decisions, not the world, determine what you are exposed to. The organisations that survive understand that making decisions with uncertainty is a permanent condition: they monitor for change rather than pretending their forecast was the last word.

---

Grant Purdy is the co-author, with Roger Estall, of Deciding (2020), and the architect of the Universal Decision-Making Method.