The alert reached people’s phones. But did everyone understand it?

A mobile phone does not ask which language you speak.

It asks where you are.

If you are in an area threatened by flooding, flash floods, landslides or another imminent hazard, an alert may take over your screen and interrupt whatever you are doing. An extreme alert — the highest severity level — will sound even when the device is set to silent, and the message will remain on screen until the user dismisses it. There is no need to download an app, register for the service or provide an address in advance.

The technology has found you.
The meaning still has some distance to travel.

Since 1 October 2025, Defesa Civil Alerta, Brazil’s public emergency alert system, has operated nationwide. It uses Cell Broadcast technology to send messages to compatible devices connected to 4G or 5G networks within an area identified as at risk. Civil defence agencies decide what the message will say and when it will be sent.

It is an important advance. A siren does not follow each person across a city; a mobile phone usually does. The system allows an emergency warning to find people, rather than requiring people to seek out an official source.

But geographical coverage alone does not guarantee comprehension.

At present, alerts are sent only in Portuguese. Official information indicates that English and Spanish versions are planned for the future, based on the language configured in the device’s operating system. No implementation timetable has been announced.

The prospect matters.

So does the gap.

Language is a geography of its own

Every warning system draws a map.

The first shows mobile towers, municipalities, at-risk areas and compatible phones. It is the map of transmission.

The second shows languages, cultural frames of reference, literacy levels, accessibility needs, trust in institutions and familiarity with the vocabulary being used. It is the map of comprehension.

The two maps do not necessarily overlap.

The technical reach of Defesa Civil Alerta has its own limits. Not every device can receive Cell Broadcast messages, and the system depends on users having compatible phones connected to the appropriate mobile networks. Brazilian authorities explicitly recognise that it does not reach 100 per cent of devices.

That is an infrastructure problem.

Language poses a different one.

Brazil’s 2022 Census identified 295 Indigenous languages spoken or used at home by Indigenous people aged two or over. Nearly 475,000 people reported speaking at least one of them.

This does not mean that every nationwide alert must immediately be translated into hundreds of languages. It establishes something more fundamental: Brazil is already linguistically diverse before we even consider tourists, refugees, immigrants and other residents whose first language is not Portuguese.

English and Spanish would narrow the gap.

They would not close it.

Some people can communicate in Portuguese in everyday situations yet still struggle to interpret a technical instruction within seconds and under pressure. Being able to hold a conversation in a language is not the same as quickly understanding expressions such as “mass movement”, “flood threshold”, “self-rescue zone” or “proceed to a safe location”.

During an emergency, that difference is no longer academic.

An imperfect translation on an institutional webpage may cause confusion. In a disaster alert, it may change what someone decides to do: whether to stay, leave, return, seek shelter or take a route that is no longer safe.

Risk communication does not merely describe danger.
It forms part of the response to it.

Danger needs editing too

In 2026, researchers from Cemaden, Brazil’s national disaster-monitoring and early-warning centre, and the National Institute for Space Research examined the risk communication capabilities of Brazilian municipalities.

The study combined an analysis of records from 1,993 municipalities with 60 in-depth interviews and qualitative fieldwork in Cataguases, Minas Gerais, and Nova Friburgo, Rio de Janeiro. The researchers identified shortages of funding, equipment and staff, difficulties in responding to misinformation, and problems with the language used to communicate risk. Terms that may appear familiar to specialists — including “extreme events” — were not always understood by the public as intended.

That finding reframes the problem.

It is not enough to ask how many languages a message is available in. We must first ask whether the message was designed to be understood in any of them.

An emergency alert is a tiny and unforgiving editorial form. A few words must explain what is happening, where it is happening, how urgent it is and what people should do. There is no room for lengthy context, convenient ambiguity or ornamental elegance. Every word must compete with fear, noise, distrust and false information — all while time is running out.

The message must be brief without being incomplete.

Direct without becoming imprecise.

Urgent without causing panic.

Locally relevant without sacrificing institutional consistency.

That requires writing, editing and testing. Translation may come later in the workflow. In a genuinely multilingual system, however, it must shape the message from the outset.

Localising an alert is not a matter of replacing Portuguese words with foreign equivalents. It requires determining whether the instruction remains actionable in the target context.

“Seek shelter” is of limited use when the public does not know which shelter is meant.

“Avoid at-risk areas” assumes those areas can be recognised.

“Follow the authorities’ instructions” depends on people knowing which authorities, through which channels, and how to recognise an authentic message.

Sometimes a sentence is accurately translated and operationally useless.

Translating the siren

This is where artificial intelligence becomes attractive.

Alerts must be produced quickly. Different events require different messages. A system may need to operate at scale, in the middle of the night, under pressure and across local authorities with unequal resources. Keeping translators available for every language combination, in every municipality and at every hour would be difficult and, in many contexts, impracticable.

AI offers speed, availability and the capacity to operate at scale.

In the United States, the National Weather Service has developed a programme that uses artificial intelligence to translate weather information. Its workflow brings together language models, professional translators and bilingual meteorologists. Human-verified translations are fed back into the system to improve later output. When time is critical, some requests can also go directly to the model, which draws on validated translation memories and generates new segments when necessary.

During a Spanish-language pilot involving National Hurricane Center products, NOAA reported that AI assistance cut the time required to translate some materials from approximately one hour to less than ten minutes.

Ten minutes can make a tangible difference when a storm is approaching.

But speed does not automatically turn a translation into a trustworthy warning.

The American programme is instructive precisely because it does not treat AI as a machine into which English goes and public safety emerges in another language. It draws on professional translators, bilingual subject-matter experts, terminology work, recurrent evaluation and opportunities for public feedback.

The technology accelerates the workflow.
It does not replace the workflow.

That distinction matters because recent research on AI and climate communication presents a less triumphant picture than the fluency of the models might suggest. Generative systems can support multilingual outreach and improve presentation while still creating problems of accuracy, context, completeness, uncertainty and equity. The research base also remains heavily weighted towards English-language, text-based communication.

A longitudinal study of machine-translated climate communication subtitles similarly found substantial improvements in readability and acceptability, but less progress in functional equivalence. Nuanced meaning remained a persistent source of error.

A machine can produce a sentence that is perfectly natural and dangerously wrong.

It can turn a recommendation into an order, weaken a prohibition, remove a condition that limited an instruction, or choose a technically valid term that is unfamiliar to the very community expected to act on it.

In ordinary texts, these are translation errors.

In an alert, they can prompt the wrong action.

The authorised version you cannot read

The National Weather Service programme also exposes a paradox.

Its materials warn that automatically processed translations may contain errors or omissions and that, in cases of doubt, the English text remains the authorised version.

From an institutional perspective, the disclaimer is understandable. An organisation must determine which text prevails when versions diverge.

From the perspective of the person who depends on the translation, it is less satisfactory.

If that person cannot understand the authorised version, saying that English prevails does not correct the alert they received. The disclaimer establishes which text controls; it cannot restore meaning at the moment a decision must be made.

This is where editorial governance meets public urgency.

A literary translation can be corrected in a new edition. A news article can carry a correction. In software, an unsuitable string can be replaced in the next update.

A disaster alert has a brutally short shelf life.

By the time a correction arrives, the water may already be there.

Emergency messages therefore cannot depend on disclaimers issued after translation. Appropriate linguistic safeguards must be built into the infrastructure before an emergency occurs.

That means writing clear, controlled source texts; developing consistent terminology for different hazards; preparing message templates in advance; validating instructions with specialists and communities; recording approved versions; testing whether people understand what they are expected to do; and establishing when an automated translation may be used, when it requires review and when it should not be published.

Translation does not begin when an alert enters a model.

It begins when someone decides how the danger will be described.

Who approves the alert?

In Brazil, state and municipal civil defence agencies are responsible for the content and timing of alerts. To use the system, they must meet requirements that include training and contingency planning.

That allocation of responsibility makes sense: those closest to the area often understand the risk best.

Translation, however, creates another chain of decisions.

Who will define the terminology in English and Spanish?

Will there be a national repository of pre-approved messages?

Will local authorities be allowed to adapt them?

Will an AI system generate versions in real time, or work from templates that have already been approved?

Will professional translators, meteorologists, geologists, civil defence personnel and community representatives take part in validation?

How will discrepancies between the Portuguese version and the others be handled?

Which version would be treated as authoritative in any later investigation?

The Anatel and Ministry of Integration and Regional Development pages consulted for this article mention the planned introduction of English and Spanish, but do not explain how translations will be produced, validated or updated, or who will be responsible for each language version.

The absence of that information does not mean that no planning exists.

It does mean that the linguistic architecture is not yet publicly visible with the same clarity as the transmission architecture.

It should be.

In high-risk systems, trust does not depend only on whether the technology works. It also depends on knowing how it was authorised to work, who can correct its output and who is accountable for the decisions embedded in it.

“Human review” does not resolve everything if we do not know what, exactly, the human reviewed.

One person may correct the grammar without verifying the instruction.

A translator may compare the languages without knowing how the hazard is developing.

A subject-matter expert may verify the technical content without being able to judge whether the translated message is natural and usable.

A local official may understand the territory without having the authority to reject the system’s output.

The task requires different kinds of expertise and an explicit chain of approval — not to add bureaucracy during an emergency, but to prevent improvisation when the emergency arrives.

Different forms of expertise may be distributed across the workflow.

Final accountability cannot be.

It should rest with the public authority that authorises and issues the alert, supported by a documented chain of linguistic, technical and local validation.

AI should not translate a disaster alone

This is not an argument for excluding artificial intelligence.

It is an argument for putting it in the right place.

AI can produce initial versions, retrieve validated terminology, adapt previously approved messages, identify inconsistencies between languages and help small teams serve larger audiences. It may be especially valuable when the time available is measured in minutes.

But the system must operate within limits established before the event.

The greater the possible consequence of a sentence, the less freedom the model should have to improvise.

Critical language — evacuation orders, prohibitions on returning, evacuation instructions, dam-failure warnings and directions to emergency shelters — may require controlled, pre-approved templates. Variable details such as the location, time and nature of the hazard can be inserted into those templates without recreating the entire message for every event.

The best use of AI may not be to generate alerts freely in many languages.

It may be to help qualified people build, test, retrieve and update messages in which every critical word has already been considered.

That sounds less spectacular than asking a model to “translate now”.

It is also closer to safety engineering.

The alert does not end with delivery

The expansion of Defesa Civil Alerta is an achievement in public infrastructure. The system has increased Brazil’s capacity to reach people directly in threatened areas and reduced dependence on prior registration or constant monitoring of the news.

The next challenge is harder to see.

It does not appear in coverage maps or message counts. It begins after delivery, when a warning must still be understood, trusted and acted on.

The United Nations Office for Disaster Risk Reduction has stressed that effective warnings must be understandable, actionable, localised and developed with attention to the languages, channels and voices of communities. Translation alone is only part of that work.

The 2025 global assessment of multi-hazard early warning systems expresses the same principle more simply: warnings are effective only when they are received, understood, trusted and acted on.

In a changing climate, early warning systems are also part of adaptation infrastructure. Their effectiveness cannot be measured solely by the number of devices reached.

A message is not complete when it leaves the server.

It is not complete when it appears on a screen.

It is not complete even when it has been translated.

It becomes complete when a person understands the risk, recognises the instruction, trusts its source and is able to act in time.

Transmission is a technical process.
Comprehension is the public service.

And between the two lies a linguistic border that no siren can cross alone.

Postscript: when the alert came close to home

1 November 2026, São Luís, Maranhão.

Two days after this article was published, the question in its title appeared on mobile phones across São Luís, the city where I live.

At 10 a.m., before the controlled demolition of the Marcelino Machado Bridge over the Estreito dos Mosquitos, residents received an audible and visual Defesa Civil Alerta notification. Classified on screen as an “extreme alert”, it announced that the demolition would take place shortly and instructed people to move away from the area and follow the guidance of the Civil Defence and the Maranhão Military Fire Brigade.

The message was brief, direct and actionable. The sound, however, surprised many people and quickly became part of the event itself, prompting comments and humour online. That reaction does not establish that the alert failed. It reveals something more useful: delivery and reception are different stages of the same public service.

The episode also clarifies why the linguistic question begins before translation. Even in a shared language, a warning must define the hazard, locate the risk, calibrate urgency and state the required action in terms a heterogeneous public can use. Multilingual versions add another border; they do not create the first.

The event does not change the argument of this article. It brings it home. The alert crossed the network as intended. Its public meaning was then formed in the interval between interruption and interpretation — precisely where communication becomes service.