Tsunami lead time, crisis-video verification and disaster visibility: the crisis communication brief for 13–17 August 2026

Coverage: 13 August 2026, 07:46 to 17 August 2026, 07:30 (Europe/Lisbon).

Three items qualify. Together, they show three ways in which useful information can fail: a near-field tsunami can leave almost no time for last-mile warning, a realistic synthetic video can survive both human and automated scrutiny, and a severe disaster can receive less cross-border attention in patterns associated with its hazard type and the ties between affected and reporting countries.

Priority	New signal	Evidence status	Test for practitioners
1	Indonesia's first tsunami warning followed the Flores earthquake by 2 minutes 16 seconds, yet modelled arrival in two near-field areas was less than two minutes later. 1	Official alert-authority timeline and sea-level observations; no published evidence yet on public receipt, local relay, evacuation or comprehension.	Which protective actions begin on strong shaking, and which still wait for a central message?
2	A new crisis-video benchmark found that generated clips which repeatedly looked real to people also produced weak or unreliable evidence from current detector families. 2	Large preprint benchmark with paired real and generated clips; not peer-reviewed, and its human and dissemination tests were controlled simulations.	Does an automated score prioritise verification work, or does it wrongly authorise publication?
3	A peer-reviewed study of 135 million articles found much smaller short-run cross-border reporting increases after floods and droughts than after earthquakes and other visually dramatic hazards. 3	Large cross-national observational study with open replication data and code; it measures news attention, not public understanding, aid or behaviour.	Which hazards in the risk register need active external distribution because organic attention is unlikely?

1. The Flores warning leaves 1 minute 44 seconds on the modelled clock

What changed. A magnitude 7.7 earthquake struck north-east of Mbay in Indonesia's East Nusa Tenggara province at 04:58:24 Western Indonesia Time on 15 August. The Meteorology, Climatology and Geophysics Agency (BMKG) issued its first tsunami warning at 05:00:39, 2 minutes 16 seconds after the earthquake. Its updated bulletin followed at 05:09:56. 1

That headline speed conceals the harder clock. BMKG's model put the first tsunami arrival in northern Manggarai and northern Ngada at 05:02:23, only 1 minute 44 seconds after the first warning was issued. Several other modelled arrivals preceded the updated bulletin. BMKG later reported measured waves at ten coastal points, beginning with 0.20 metres at Sikka at 05:03; the largest listed observation was 0.94 metres at Maurole-Ende at 05:27. The agency ended the warning at 07:30 after its stations found no further dangerous rise in sea level. 1 4

Evidence strength. BMKG's bulletin gives an unusually precise agency-side chronology: earthquake time, two warning times, modelled arrival times, observed wave times and cancellation. Those records establish what BMKG detected and issued. They do not show when a handset, siren, broadcaster, local official or coastal resident received the warning, or whether people moved before the water arrived. The widely shareable 2-minute-16-second figure therefore measures central issuance, not end-to-end warning performance.

Policy dimension. BMKG tied each threat category to an action owner. The agency classified SIAGA areas as facing modelled waves of 0.5–3 metres and WASPADA areas as below 0.5 metres. It told local governments in SIAGA areas to direct evacuation, while people in WASPADA areas were told to leave beaches and riverbanks. 1 The narrow gap between central issuance and modelled arrival makes a central bulletin unsafe as the sole planning trigger: even immediate receipt would have left less than two minutes in the first two named areas. The BMKG records do not show whether natural warning signs or pre-authorised local action worked in this event.

The cancellation also needs its own message design. BMKG ended the tsunami warning while aftershocks continued, and it told the public to remain calm, avoid damaged structures and use verified BMKG channels rather than rumours. 4 A cancellation that sounds like a general all-clear can erase the remaining instructions. Teams should name the hazard that has ended, the hazards that remain, the next update time and the route for correction in separate, scannable lines.

Operational implication. Measure four times, not one: detection, authority issuance, last-mile receipt and protective action. The first two can look exemplary while the latter two remain unknown. Exercises should also test which coastal actors can act on natural warning signs without waiting for a message, and how the central update confirms rather than delays that action.

Decision question: which local actions begin on shaking, who can start them without further approval, and where is public receipt recorded before the first modelled arrival?

2. Crisis-video detection weakens where communicators most need certainty

What changed. Researchers released RA-Bench, a preprint benchmark built around 1,830 real-video anchors and 16,056 generated clips across ten social-risk categories, including wars, disasters and public emergencies. Four open-source and five closed-source generators produced the synthetic set. The study evaluated seven conventional detectors, ten zero-shot multimodal models under several review settings and two fine-tuned multimodal detectors. No detector family generalised consistently across the benchmark. 2

The human test makes the result more operationally relevant. Twenty trained student reviewers supplied 53,550 first-stage judgements. Across those judgements, reviewers correctly identified generated clips 60.3% of the time and real clips 71.9% of the time. The researchers then isolated 633 generated clips that five reviewers had all labelled as real. On that subset, seven conventional detectors averaged an area under the curve of 47.5%, close to random ranking, and detected only 4.4% of generated clips at a 5% false-positive rate on real clips. 2

A second test simulated the last mile. The researchers applied combinations of transcoding, spatial downsampling, conversion to eight frames per second and a synthetic news badge to matched real and generated clips. Under the full transformation, the seven-detector mean AUC fell from 51.4% to 47.3%, while generated-video detection at a 5% false-positive rate fell from 5.7% to 2.8%. Every evaluated fine-tuned configuration labelled no more than 2.4% of generated clips as fake after the full transformation. 2

Evidence strength. RA-Bench is large, paired and deliberately focused on high-harm scenes. The preprint has not completed peer review. Its human reviewers were university students rather than newsroom verification specialists or affected communities. Its "social dissemination" test reproduced ordinary technical and presentation changes during circulation; it did not observe clips moving through live platforms or real crisis networks. The reported detector failures are strong evidence about this benchmark, not a universal failure rate for every tool or incident.

Policy dimension. A detector score can prioritise a clip for human verification; it cannot establish authenticity. A high-confidence "real" result cannot authorise publication, executive briefing or rebuttal when the benchmark's hardest generated clips also looked real to people. High-harm footage still needs source provenance, location and time checks, independent corroboration and a named human decision owner.

The transformed-video result changes procurement and exercises. Teams should test candidate tools on the clips they actually receive: compressed reposts, altered frame rates, overlaid logos and partial provenance. A detector that performs on original files can fail after ordinary circulation. Procurement should therefore specify acceptable false-negative performance at a fixed false-positive rate under realistic transformations, rather than accepting an overall accuracy score on a vendor's clean test set.

Operational implication. Separate three decisions: whether a clip is verified, whether it is safe to use, and whether it requires a public response. An unverified, high-harm clip may demand monitoring or a holding line before it can support a factual claim. The detector can move that clip up the queue; it should not collapse the three decisions into one label.

Decision question: if a plausible crisis video arrives compressed and stripped of provenance, which evidence beyond the detector score must exist before the team publishes the clip or rebuts the claim made with it?

3. Disaster visibility varies by hazard type and country ties

What changed. Thiemo Fetzer and Prashant Garg combined the European Commission's Europe Media Monitor with the EM-DAT disaster database. Their panel contains about 135 million articles from 466 news sources in 123 countries and 2,662 disasters between 2016 and 2023. For each outlet-country and disaster-affected-country pair, the authors estimated how the outlet country's reporting about the affected country changed during the three days after a disaster. The model compared that change with the pair's normal volume and accounted for global daily shocks. 3

Cross-border attention rose most after earthquakes, dry-mass movements and volcanic eruptions. On the study's log(1 + article count) scale, the mean estimated three-day increase was 0.0785 for earthquakes, compared with 0.0069 for floods and 0.0001 for droughts; the drought estimate was not statistically different from zero. These coefficients compare post-event reporting with the same outlet-country and affected-country pair's usual level; they are not raw article counts. Hydro-meteorological disasters received less attention than geophysical disasters after the model accounted for severity and duration. The reporting increase was significantly larger for events with at least 100 deaths than for events with 0–9 deaths. The differences for events with 10–19, 20–49 and 50–99 deaths were not statistically significant. 3

Country-pair characteristics were also associated with the pattern. Social connectedness and a measure of long-run historical relatedness were associated with steeper fatality–coverage gradients between country pairs. The authors treat those measures as correlates, not causes; their design cannot distinguish easier information flow from cultural affinity or in-group preference. 3

Evidence strength. The study is peer-reviewed, global in scale and reproducible: its balanced article-count panel, disaster windows, covariates and code are openly archived. The design measures online news attention rather than message quality, public understanding, donations, policy response or assistance. It also inherits blind spots from sparse digital news, censorship, unsupported languages and EM-DAT's weaker recording of small remote disasters. The study supports an attention-risk diagnosis; it does not prove what causes the gap or what intervention will close it.

Policy dimension. Some communication plans depend on foreign news pickup to reach partners or decision-makers. This study does not show whether extra coverage changes mobilisation, support or behaviour. It does show that teams face systematically different pickup conditions by hazard type, even after the model accounts for severity and duration. Plans for hydrological, meteorological and slow-onset hazards should therefore avoid treating organic attention as a given.

The study does not test which intervention closes the gap. A team can nevertheless make its attention assumption observable by monitoring external pickup alongside hazard and impact indicators, preparing maps and comparable trend data before the response peaks, and testing diaspora, scientific and partner networks that connect the affected place to outside audiences. Those are operational options to evaluate, not findings from the paper.

Operational implication. Add an attention assumption to hazard scenarios. A communications plan that relies on organic foreign coverage should say which outlets, languages and networks are expected to carry the story, and what the team will do when the expected pickup does not occur. The research also argues for testing those assumptions by hazard type rather than copying the media plan from the last earthquake or wildfire.

Decision question: which severe hazards in your remit are likely to be internationally under-covered, and which distribution partners can carry verified evidence before fatalities become the news hook?

Forward watchlist

Flores warning performance: any BMKG, local-government or independent after-action record that connects central issuance to channel receipt, evacuation timing, reach and public understanding.
Crisis-video verification: peer review, dataset and code release, independent replication, tests with professional verifiers and performance on real platform transformations.
Cross-border disaster attention: replications using broadcast and social channels, and studies linking attention to public understanding, funding, policy or assistance.

Editorial transparency

This edition includes one live public-warning case, one AI-information-integrity preprint and one peer-reviewed cross-national study. The selection rotates towards Indonesia and wider Asia while keeping two globally framed evidence items. It spans operational alerting, synthetic-media verification and disaster visibility.

WHO's SAPHIRE 2026 chemical-emergency simulation was eligible and directly relevant, but its public update gave only the organisers' account and no exercise report, performance data or participant findings. The United Kingdom's national wildfire alert was also eligible; the Indonesian case offered more precise event, warning, modelled-arrival and cancellation times, while avoiding another Europe-led edition. WHO's 14 August Ebola update was authoritative, but Central Africa and that outbreak featured in the previous edition, and the new communication evidence was less incremental than the selected items.

Access limits affected discovery, not the claims above. Several source indexes could not be fully enumerated: ReliefWeb's current interface required approved API access, while IFRC and PreventionWeb list pages were blocked. Search and accessible first-party detail pages provided partial coverage of those organisations. All selected sources were open in full. The BMKG record lacks last-mile and behaviour data; RA-Bench remains a preprint; the Nature study cannot establish downstream effects or causal mechanisms. The article keeps those boundaries explicit.

Direct-source access

BMKG, Gempa Bumi M7,7 – 30 km TimurLaut Mbay-Nagekeo-NTT — open HTML bulletin with issue times, modelled arrival areas, action levels and maps; direct link in citation 1 above.
BMKG, BMKG Akhiri Peringatan Dini Tsunami Pasca-Gempa M7.7 NTT — open HTML press release with observed wave times, cancellation time and continuing aftershock advice; no public-receipt or evacuation dataset linked; direct link in citation 2 above.
Liang et al., Can We Defend Against AI-Generated Video Attacks on Real-World Crisis Events? — open arXiv abstract, full experimental HTML and PDF; version 1 preprint, submitted 14 August 2026 at 16:32 Europe/Lisbon; direct link in citation 3 above.
Fetzer and Garg, Uneven patterns of cross-border media coverage following natural disasters — open-access Nature Human Behaviour article with supplementary information, peer-review file and source data; the replication archive is also open; direct article link in citation 4 above.