No more guessing: New AI decodes public emotion during disasters

During a health crisis or natural disaster, it's hard to know what people are actually most worried about. A research team built CrisisKD, a new AI system that reads social media posts and identifies both the subject people mention — like "vaccine shortages" or "closed roads" — and the emotion attached to it, such as fear or anger. Governments and aid groups could use this to respond more precisely instead of guessing.

The system uses a two-step trick called knowledge distillation. First, a powerful, expensive AI reads thousands of tweets and generates detailed labels, explaining which words point to problems and what emotions they carry. Then, a smaller and much cheaper AI learns from those examples. The result is a model that costs far less to run and still understands the nuances of crisis conversations.

On a hand-checked test of 500 tweets, the smaller model's emotion accuracy improved by 17 points, and sentiment accuracy by 6.5 points, compared with the same AI before training. It also gained ground on a public benchmark for extracting specific aspects and the feelings tied to them. The team is releasing the dataset of over 50,000 labels and the training code for free, so others can build on the work.

The real payoff is practical: an affordable AI that monitors public mood during emergencies could help disaster teams spot rising panic, food shortages, or anger at slow aid delivery. Officials could then adjust their messaging or send resources where emotions are hottest. The catch is that AI still misses sarcasm and unusual slang, so human oversight remains essential — but as a radar for public sentiment, this is a significant step forward.