During the Sri Lanka cyclone that killed 410 people, ChatGPT usage spiked 17x. Nobody checked whether the answers were right. OpenAI is now building disaster response tools on that unverified foundation.
image from grok
OpenAI detected a 17x surge in ChatGPT usage during Sri Lanka's Cyclone Ditwah (410 deaths, 233,000 displaced) and subsequently announced disaster response as a pillar of its OpenAI for Countries program. The company held an AI Jam in Bangkok with disaster management leaders but has not published any accuracy audits or shared what information users actually received during the crisis—despite explicitly not knowing whether AI outputs contradicted local emergency services or WHO guidance. The core concern is asymmetric stakes: a hallucinated code suggestion is an inconvenience, while a hallucinated evacuation route during a storm that killed 410 people is something else entirely.
When Cyclone Ditwah hit Sri Lanka on November 28, 2025, killing 410 people and displacing nearly 233,000, people turned to ChatGPT. According to OpenAI's own data, cyclone-related messages jumped 17 times above baseline. In Thailand, a separate storm that same month drove a 3.2-times increase in ChatGPT usage. Asia is the world's most disaster-prone region, and it is increasingly an AI usage hotspot during crises.
What nobody has asked — including OpenAI — is whether the information people got was accurate.
On March 29, 2026, OpenAI held an AI Jam in Bangkok, bringing together 50 disaster management leaders from 13 countries alongside representatives from the Gates Foundation, the Asian Disaster Preparedness Center (ADPC), and the nonprofit DataKind. Participants built custom GPTs for situation reporting, needs assessment, and public communication. Sandy Kunvatanagarn, OpenAI's head of public policy, said the session was "aimed at closing the gap between what AI can do and how it's actually used in the field."
The gap that deserves more attention is a different one.
OpenAI announced disaster response as a planned pillar of its OpenAI for Countries program at Davos on January 21, 2026 — a month and a half after Cyclone Ditwah had already struck Sri Lanka. Reuters reported at the time that OpenAI was hoping to work with governments on disaster planning, among other areas. The cyclones that hit the region in late 2025 were not the origin of the idea. They became the justification for it.
The usage figures OpenAI cites — a 17-times spike during Ditwah, a 3.2-times jump in Thailand — come from internal data the company has not shared publicly or subjected to external audit. OpenAI says people were using ChatGPT to access information during the crisis. It has not published what that information was, whether it was correct, or whether it contradicted guidance from local emergency services or the WHO. The company did not respond when asked what accuracy checks it runs on disaster-related outputs.
This matters because the stakes are asymmetric. A hallucinated code suggestion is an inconvenience. A hallucinated evacuation route during a storm that killed 410 people is something else.
The World Bank estimates disasters have cost ASEAN countries more than $11 billion in previous years. Asia accounts for roughly 75 percent of people affected by disasters globally. The scale is real. OpenAI is not wrong that AI could play a role. The question is whether anyone is checking whether it currently does — and the company's own communications suggest that check has not happened.
The blog post announcing the Bangkok event quotes four people: two OpenAI employees, a Gates Foundation official, and an ADPC director. No academic researchers specializing in disaster management, no independent evaluators of AI accuracy, no representatives from the UN Office for Disaster Risk Reduction. A company announcing a new deployment of its product in a high-stakes domain, with no independent voices willing to go on record, is a pattern worth flagging.
OpenAI says it is exploring a second phase focused on pilot deployments and deeper technical collaboration with participating organizations. What "deeper technical collaboration" means, who funds it, and whether it includes independent evaluation are questions the blog post does not answer. The company's Davos presentation listed 11 countries signed up for OpenAI for Countries, with each deal structured differently. The disaster response work appears to be part of that program, but OpenAI did not specify how it is funded or what success metrics will be used.
The cyclones that hit South and Southeast Asia in late 2025 were real humanitarian crises. The people who died, the communities displaced, the infrastructure destroyed — those are not framing devices. When OpenAI uses them as evidence of AI demand in disaster response, it is making a claim about what people needed and what they got. That claim deserves scrutiny that the company's own communications do not provide.
The Bangkok event produced working prototypes and expressions of intent. Whether any of it will function correctly when the next cyclone hits, who will pay for it, and whether anyone will measure whether it helped or hurt are questions left open.
Sources: OpenAI blog post | Reuters — Davos announcement, January 21, 2026 | WHO — Cyclone Ditwah situation report
Story entered the newsroom
Research completed — 0 sources registered. The OpenAI blog post is LIVE and accessible — the earlier 403 was a bot-check artifact, not a real 404. The story was prematurely killed based on a fa
Draft (704 words)
Reporter revised draft (704 words)
Approved for publication
Headline selected: OpenAI sees 17x usage spike during Sri Lankan disaster. Now it's building a program. Accuracy check: later.
Published (704 words)
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Artificial Intelligence · 3h 50m ago · 3 min read
Artificial Intelligence · 4h 11m ago · 3 min read
