Skip to content

Insights Innovation

A forecast is not an evacuation

Prediction has become the easy half. The hard half is the chain of human decisions that has to fire before the water arrives.

Rakchham, Himachal Pradesh Photograph: Rowan Heuvel / Unsplash

Give a district magistrate a model that is right eight times in ten about a cloudburst six hours out. Then watch what happens.

The model is not the constraint. The constraint is that the magistrate must now decide, on a probability, to move several thousand people in the dark, at a cost that is certain, against a hazard that is not. If they move people and nothing happens, that is a visible failure with a name attached. If they do not move people and something happens, that is a tragedy, and tragedies are more often attributed to the weather.

The incentive runs the wrong way, and no amount of model skill corrects it.

What AI has genuinely changed

It would be wrong to be dismissive. The last few years have produced real, verifiable gains:

  • Nowcasting. Machine-learning models on radar and satellite inputs have narrowed the window on short-horizon convective events in ways that conventional numerical methods struggled with. In a region where lead times are measured in tens of minutes, this is not marginal.
  • Change detection. Automated analysis of satellite imagery flags slope deformation, glacial lake growth and ground displacement at a cadence and area no human analyst could cover.
  • Damage assessment. Post-event imagery classification compresses the first assessment from weeks to days, which moves relief money faster.
  • Pattern discovery. Models applied to long hazard records find relationships in the data that were not in anyone’s hypothesis.

Each of these is a real improvement. Each of them stops at the same wall.

The wall

Every one of those capabilities ends by producing information. None of them produce a decision, and none of them produce movement.

Between the output and the outcome sit a series of human links, and the chain is only as good as its weakest:

Someone has to be monitoring the output. Someone has to be authorised to act on it without waiting for confirmation from a level above. Someone has to issue the instruction in a form a household understands. Someone has to be trusted enough that the household complies. And somewhere, there has to be a place to go that is better than the house.

A probabilistic forecast handed into a chain with no clear owner does not become an evacuation. It becomes an email.

The thing we keep getting wrong about uncertainty

Technical systems express hazard as probability. Administrative systems need a binary: act or do not act. Somewhere in between, a human has to convert one into the other, and that conversion is where the accountability lands.

Most protocols leave the threshold to judgement in the moment. This is the worst possible arrangement, because it loads the decision onto an individual at the point of maximum stress and minimum information, and it guarantees inconsistency between districts facing identical data.

The alternative is to agree the threshold in advance, in daylight, with everyone present. Forecast-based financing, where a defined trigger automatically releases money and authorises pre-agreed action, exists precisely to solve this. It converts the question from “should I act tonight” to “was the trigger met”, which is a question with an answer.

It also provides political cover, which is not a cynical consideration. It is the mechanism by which a reasonable official can afford to be wrong.

Technology that does not need a network

A sobering constraint for anyone designing Himalayan systems: the hazard frequently takes the infrastructure the system runs on. Power fails, towers fall, roads close, and the valley is on its own.

This should discipline design more than it does. It argues for:

  • Protocols that degrade to something a person can execute on paper
  • Local sirens and physical signals with independent power
  • Community volunteers with defined authority, not only defined tasks
  • Radio, which remains the only broadcast medium that reliably survives

A system that is excellent when connected and inert when not is a system that is excellent except during disasters.

The Silkyara lesson

In November 2023, forty-one workers were trapped in the collapsed Silkyara-Barkot tunnel in Uttarkashi for seventeen days. They came out alive.

What is instructive about the rescue is what finally worked. Advanced machinery was brought in. The large-diameter auger drilling through the debris broke. The final stretch was completed by rat-hole miners working manually through the last metres of rubble.

The lesson is not that technology failed. The machinery got them most of the way. The lesson is that a response capability built entirely around one high-technology pathway has no fallback when that pathway jams, and that the fallback, when it was needed, was a skill set the formal system does not maintain, train or pay for.

Redundancy is not inefficiency. In disaster response it is the whole design.

What we are asking at WSDM 2026

The Innovation pillar at the summit is deliberately not a technology showcase. #TechForResilience exists so that working tools can be put in front of the people who would have to operate them, and the test we want applied to every one of them is the same:

  1. What decision does this change, and who makes it?
  2. What happens to it when the power and the network are gone?
  3. Who is accountable when it is wrong, and have they agreed to that in advance?

A tool that cannot answer those three is a demonstration, not a capability.


The Innovation pillar runs across all three days of WSDM 2026, alongside the #TechForResilience exhibition. Abstracts on early warning, AI for hazard modelling and decision-support systems are invited.

Call for abstracts

Abstract submissions open on 15 October 2026. Abstract submissions are open.

In response to requests from many institutions, the abstract submission deadline has been extended to 24 November 2026. Take this opportunity to present your work.

Fees apply only to abstract presentations. Registration to attend is free and mandatory.