Language Barriers in Survivor Support: When No One Can Read Your Story

A survivor sits down to write about one of the hardest things they've ever lived through. They write it in their first language, the one that holds the words for what happened. Then the story reaches a support team that reads only English. That gap is one of the most overlooked language barriers in survivor support.

Language barriers shape who finds support and who doesn’t. Every story, question, and comment submitted to our online platform is read by a reviewer before it goes anywhere. Our team checks that the story is safe to publish and they write back with resources for whoever sent it. But none of this works if they cannot read what a survivor wrote in the first place.

More than 580,000 people have joined our online community. Survivors have shared 1,963 published stories with us so far, across 80 countries. They have written in 14 different languages, across stories, questions, messages, and comments.

In August 2026, we shipped a translation layer inside our admin dashboard. This new tool gives our review team a way to read a submission in any language without ever leaving our platform, and to reply in the language a survivor wrote in. Here’s how it works, what it protects, and where we’re still working on solutions and improvements.

Why Language Barriers Keep Survivors From Support

Most survivors never reach formal support at all. Fewer than 40 percent of women who experience violence seek help of any kind, and fewer than 10 percent of those women who seek help actually go to the police. Most turn to family and friends, not institutions. 

Language creates another barrier to support. There are more than 7,000 living languages worldwide and in the United States alone, more than 350 different languages are spoken. Even though we regard English as the world’s common language and the official language of the U.S., roughly 25.7 million people in the country have limited English proficiency, and that includes 47 percent of immigrant adults.

English only became the official language of the U.S. by executive order in March 2025. This same order revoked an executive order that was in place since 2000 telling federal agencies to improve access for people with limited English proficiency. The Justice Department described this change as “removing extensive translation services and de-prioritizing multilingualism over English proficiency.”

Language barriers run through everyday life, including access to care. About half of U.S. adults with limited English proficiency hit at least one language barrier in a healthcare setting over three years, and a third struggled to understand what a provider told them. 

These barriers also run through gender-based violence services. A 2025 global review of 195 studies found that women seeking help for gender-based violence struggled to describe what happened to them when services could not provide interpreters. The review concluded that language barriers keep pushing women out of support networks entirely.

“Survivors face many challenges when trying to get help across countries. Language barriers and the lack of temporary safe spaces make it difficult for them to access immediate support. Those who leave their jobs or escape from perpetrators often have no access to food, shelter or assistance, especially if they have no relatives in the country.” – Hkawn Mai, Women Exchange Advocacy Officer at MAP Foundation, on cross-border support for women migrant survivors in the Mekong region

Another study in Estonia found language limitations as a main barrier for displaced and stateless survivors, alongside cultural misunderstanding and mistrust of institutions. Language justice, as described by the NSVRC, is the right for survivors to communicate in the language that is most comfortable to them. 

Language access is then enacting this right by offering inclusive services. This is the standard we build toward with our community platform. A survivor should not need a second language to be understood.

14 Languages and a Small Review Team

Survivors have written in 14 distinct languages on our community platform so far, across stories, questions, messages, and comments. Most have written in English, but 149 submissions have been written in another language. This number is significant enough that our review team comes across a non-English submission regularly. 

These submissions are not only stories. Survivors also ask questions on our platform, comment on other survivors’ stories, and send messages of support. All of these different types of submissions carry the same weight, because any of them can hold a disclosure. 

We also have stories from 80 different countries, but this does not mean we have 80 different languages. Someone in Germany may write in Turkish and someone who finds us through our Spanish-language site may write in Portuguese. Knowing where a survivor lives tells us very little about the language they think and write in. 

Before we built any AI tools, volunteer translators handled non-English submissions case by case. As volume grew, that stopped being workable, since a timely review cannot depend on whether a specific volunteer is available. We now have a dedicated review team that clears submissions within 12 hours on weekdays and 24 to 36 hours on weekends.

How We Translated Survivor Stories Before Automation

Until August 2026, our reviewers handled non-English submissions by leaving our platform entirely. A reviewer stripped the identifying details out of a story, pasted short fragments of the story into an outside translation website, and put the meaning back together from those translated pieces. 

To reply, reviewers drafted in English, pasted it out for translation, and then pasted the result back in. They had no way to check how that reply read in context, as online translators can miss conceptual nuances. 

It was a careful approach, and the best way to do it at the time with the resources we had, but it came with costs, such as:

• It was slow. A single story could take dozens of trips back and forth between two browser tabs before a reviewer felt ready to act on it.

• It split a story into disconnected pieces, so a reviewer could follow the gist of one sentence and still miss the shape of what someone was writing as a whole.

• Some important contextual passages could not be pasted at all. The name of a school, workplace, or person can never leave our platform, and those are often details that pull a story together.

The more carefully a reviewer protected a survivor story, the less of it they could actually understand without greater context. But we are not the only organization stuck at this intersection of comprehension and exposure. Many times, understanding a survivor’s story not in your own language means showing their words to someone (a human interpreter) or something (online translation tool), and that exposure is inherently risky.  

A survey of 314 primary care practitioners found that confidentiality worries block the use of trained interpreters, especially in small language communities where the interpreter and the person disclosing already know each other. Another study derived from interviews with 43 people, revealed cases where the abuser served as the interpreter, and that blocking a partner from learning a local language was a deliberate tactic of control. 

On the technology side, pasting sensitive text into free AI tools is a practice that has become dangerously routine for many people. A data security report from 2025 found that 77 percent of people using AI tools at work paste data into them, 22 percent of which include personal or payment information. Furthermore, 82 percent of these pastes come from personal accounts, outside company agreements, “creating a massive blind spot for data leakage and compliance risks.”

This is exactly what our reviewers work hard to avoid. But we needed to find a way for our reviewers to read a whole story, contextual nuances included, without pasting survivor words into a tool we had no agreement with.

How Our Survivor Content Translation Tool Works

In August 2026 we shipped a translation layer inside our own admin dashboard. The tool works in four steps, following a submission from the moment it arrives to the moment a reviewer writes back. A person still reviews every story and writes every reply.

Detecting the Language at Submission

Our system checks every piece of text in a submission the moment it arrives and works out what language someone wrote it in. Stories that survivors write can also mix languages, so the system weighs how much writing sits in each one rather than trusting whichever language field comes first.

For example, a two-word English title on a story written entirely in Spanish should not decide how we file that story. Reviewers can also filter the dashboard by language, so a Spanish-speaking reviewer can pull up the submissions they are best placed to read.

Reading the Whole Story in Place

A reviewer opening a submission now sees a toggle. Turning it on renders the entire submission into English right there on the page, with the layout and the flow intact. There’s no second browser tab, no fragments, and no judgement call about what a reviewer can safely paste somewhere else. 

Handling Redacted Details Before Translation

Survivor stories often name specifics, such as a city, an employer, a school, or a person. We always redact those details, and so far 507 out of 1,963 published stories have at least one redaction.

The translation of stories still runs through an outside provider, but the difference is that our reviewers no longer visit that provider themselves. Our admin dashboard sends the text and brings the translation back on its own, in a second or two, which is what keeps a reviewer on one page.

Once a story has been redacted, our system replaces every redacted detail with a plain label before the text is sent, so a redacted city arrives as the word “City” and nothing more. The reviewer sees that same label sitting in the translated story and reads around it.

Reviewers can also translate a story before redacting it, which is often how they work out what needs redacting in the first place. In those cases the full text does go to the provider, but our contract protects everything we send. We work with the provider under a Data Processing Addendum (DPA) that bars them from storing the text, accessing it, or using it to train AI models.

Sometimes our system cannot tell exactly where a redacted detail starts and stops. In those cases, we send none of that section for translation. It shows as untranslatable and the reviewer works from the original text. We would rather lose a translation than risk sending something we did not mean to send.

Replying in the Language a Survivor Used

Reviewers can now send their reply in the language the survivor wrote in. This covers any language our translation provider supports, not only the three our website currently runs in. Someone who writes to us in Vietnamese gets a reply in Vietnamese. 

The reviewer checks a side-by-side preview before anything is sent, and our research team has already written message templates for sensitive situations using a trauma-informed approach. The new tool covers stories, questions, comments, and messages of support, and only reviewer, specialist, and admin accounts can use it.

Four Engineering Decisions That Protect Survivor Data

Each of the four decisions below shaped how our translation tool protects survivor data.

Sending as Little as We Can, and Failing Safely

A reviewer opening a new submission usually translates it first, since reading the whole story is how they work out what needs redacting. Nothing is redacted at that point, so the full text goes to the provider, and our DPA protects it. Once a reviewer marks details for redaction, our system takes those out of anything it sends after that.

Taking details out gets harder when our system cannot cleanly put a label in place of a redacted detail. Our system re-checks its own work before sending anything, and drops that section from the request if something looks off. Our dashboard flags it as untranslatable and the reviewer works from the original text instead. It costs the reviewer some time, but it means our mistakes always fall on the safe side.

Accepting Every Language a Survivor Might Use

Translating only English, Spanish, and Japanese would have been simpler, since those are the languages our website currently runs in. But survivors do not pick a language to suit our website. Our system already detects whatever language it finds, so the translation service accepts any valid language code and lets the provider handle what it can.

Paying for the Enterprise Translation Service

The free consumer translation tool and the paid enterprise API are different products with different terms. The paid one operates under a DPA. It processes content briefly and then deletes it, it does not use what we send to train models, and we keep ownership of both the original text and the translation. That distinction is the whole reason this tool can live inside our dashboard.

Testing for the Failure That Stays Hidden

A failed redaction would look just like a working one. The translation would come back, the page would look normal, and a reviewer might not notice a name or city sitting where a label should be. But by then it has already been sent over to the provider.

To catch this, we wrote tests that check what actually leaves every time we change the code. If the redacted detail is in there, the check fails and we find out before it reaches anyone. We also send made-up sentences through the live translation service and read what comes back. Translation can rearrange word order, so a label like “City” can return dropped into the wrong part of a sentence. Testing a label on its own wouldn’t show that, so we put it inside a full sentence and see what the provider does to it.

All of these decisions were guided by our AI Policy, which commits us to vetting an outside tool before we use it and sending it no more than the task requires. We wrote about where that thinking comes from in the philosophy guiding every tech decision we make. 

What Machine Translation Cannot Do for Survivors

Machine translation is what this is, and we want to be direct about its limits. Researchers have measured those limits across different languages and settings.

Accuracy can vary widely by language. Researchers ran 400 emergency department discharge statements through Google Translate in seven languages and found that meaning survived in 82.5 percent of cases overall. However, Spanish scored 94 percent while Armenian scored 55 percent, a 39-point difference on identical source text. An earlier study found 92 percent accuracy in Spanish and 81 percent in Chinese, with errors that included an instruction to stop taking a medication rendered as an instruction to continue it.

Translation systems learn from text that already exists online, so languages with less of it get worse results, and people pay for that difference. An Afghan woman lost her asylum claim after machine translation turned her first-person singular pronouns into plurals. She wanted to say that she traveled alone but the translation said "we," and judges denied the case on this discrepancy.

Accuracy is not even the hardest part. Some of these concepts do not exist across languages at all. Research with Horn of Africa refugee communities in Nairobi found that direct translations of sexual and gender-based violence terminology do not necessarily exist in those communities' lexicons. Consent had no clear equivalent, and people understood it instead through cultural markers like modest refusal. Marital rape was conceptually absent, because the community commonly equated marriage with lifelong sexual consent.

We started our newest Instagram series, called “Speak It to Defeat It,” for this reason. We’ve been working through one culture and one language at a time, looking at the words each one has, or does not have, to name these concepts. Japan was one of the first places we studied. Dr. Sachiko Kita, Director of Our Wave Japan, walked us through why the Japanese word for survivor, サバイバー, has traditionally described people who lived through natural disasters rather than abuse. She covers seven other Japanese terms related to consent, trauma, and healing.

U.S. regulation also addresses gaps within machine translation. Federal rules require that a qualified human translator review machine-translated material whenever the information is critical to someone's rights, whenever accuracy is essential, or whenever the text runs complex or non-literal. All three of those conditions describe survivor-facing communication.

So here is where we stand. A person reviews every submission and writes every reply. Translation changed what a reviewer can read, but it can never tell us whether the translation carries what a survivor meant. We know that we can lose nuance when no one on our team reads the language. Professional human translation holds a different standard, and for the most sensitive communication it remains the right one.

Why Cultural Experts Review Survivor Content

Our Wave expanded into Spanish and then Japanese in 2025, and each language required a cultural ambassador to work through the entire platform before launch. We did not use a translation file handed to a contractor. We had a native speaker deciding how the whole online experience should read for someone arriving in their language.

We translated the entire experience each time, not just the pages a survivor lands on first. A new language brings indexed search traffic. That traffic needs country-specific support resources behind it, or survivors arrive and find nothing they can use. And those resources need the rest of the experience to match, so someone gets a complete path in their own language instead of a translated landing page onto an English platform.

Now we are building a wider network of regional language and cultural reviewers. MiStory, active in more than 13 countries including Japan, Greece, the United States, and Thailand, is a primary source for that cohort. 

Our Japan partnership runs furthest ahead and shows the shape we want everywhere. Dr. Kita's team works in their own version of our review dashboard. Japanese-language submissions route to them automatically, and they review those stories themselves. We trained them on the tools and they bring in the cultural context.

That same architecture lets partner organizations review their own content with their own resources. The Dublin Rape Crisis Centre works this way with Irish stories on their WeSpeak platform.

Expanding Language Access to French and Hindi

Moving forward, expanding our platform into more languages is a priority that takes time. After analyzing submissions arriving in languages we do not yet support, French and Hindi came back as the most frequent. Those are the two that we’ve decided to target for the first half of 2027. 

One of our goals for Harbor, our forthcoming survivor healing platform, is to launch in five languages: English, Spanish, Japanese, French, and Hindi. This means that we’re not just providing five translations, we’re releasing five different versions with cultural context built in.

Language forms the first layer of this work. The layer after language encompasses what actually differs between places, including consent norms, who a survivor tends to disclose to first, and what legal options exist in a given country. We want a healing pathway that is culturally specific, personalized with AI assistance, and verified by people.

Our reasons for building tooling instead of hiring come down to a tradeoff between reviewer time and the budget of a small nonprofit. Staffing a full-time reviewer for every language a survivor might write in is not possible. Better tooling became the lever that made more languages possible at all, and that is where we continue to put in effort. 

Current Limitations of Our Translation Tool

Machine translation handles words well, but falls short in understanding full meanings of trauma narratives. Here are the current limitations of our translation tool:

• Language detection is unreliable on very short submissions: Our system sometimes labels a one-line entry as a language the survivor didn’t write in. This doesn’t affect what a reviewer reads, since they choose the language themselves.

• Story text travels to an outside provider: Reviewers often translate a story before redacting it, so the full text is sent. Once details are redacted, we strip them before sending and drop anything our system cannot cleanly label. Our DPA with the provider bars them from storing that text, accessing it, or training AI models on it. 

• We lose cultural nuance in languages nobody on our team reads: Translation gives a reviewer the words, but it cannot always tell them what those words carry.

• Building a network of cultural reviewers is slow: Finding and training the right people in each language takes longer than building software. We don’t want to rush this process, as we recognize the importance of it. We are building out this network currently.

No Survivor Should Get Less Support Because of the Language They Write In

Language remains a huge factor for survivors who need support. Most survivors never reach formal help at all. Needing a language other than the one local services are offered in can put support out of reach entirely. Our reviewers used to protect survivor stories by reading them in fragments, which kept the details safe and made the work harder at the same time. 

The translation layer we shipped in August changed that. Reviewers now read whole stories inside our own dashboard, sensitive details are replaced with labels wherever a story has been redacted, and replies go back in the language a survivor used.

Machine translation gives our reviewers the words, but it cannot tell them what those words carry in the culture where they came from. That work belongs to real people, which is why we are carefully building a network of regional cultural reviewers alongside our software. We want a survivor writing in Vietnamese or Hindi to reach the same standard of care as a survivor writing in English.

If your organization is building aligned digital infrastructure for survivors, we would love to connect. To see the resources our community has access to, visit our platform. For a full look at how we govern AI across every feature, read our AI Policy.

You can also read more about how we connect survivors to the right resources, and follow Speak It to Defeat It on our Instagram account to explore the words different cultures have, or don’t have, for naming harm.

Survivor Content Translation FAQs

How does Our Wave handle stories submitted in other languages?

Our system checks every submission when it arrives and works out what language someone wrote it in. Reviewers then read the whole submission in English inside our own dashboard, and they reply in the language the survivor used. A person reviews every submission and writes every reply.

Is it safe to machine-translate a survivor’s story?

That depends entirely on how you do it. On our platform, translation runs through a paid enterprise service under a Data Processing Addendum (DPA), which bars the provider from storing the text, accessing it, or using it to train AI models. On a story that reviewers have already redacted, our system replaces those details for plain labels before it sends anything. That arrangement differs fundamentally from pasting text into a free translation website, and it is why this tool can live inside our dashboard.

Can machine translation understand what a survivor is really saying?

Not fully, and we do not claim it can. Researchers have found that terms like consent have no direct equivalent in some languages, and that expressions of distress vary widely across cultures. Translation gives a reviewer access to the whole of what someone wrote. Human judgment does the rest, supported by a growing network of regional cultural reviewers.

What languages does Our Wave support?

Our website runs in English, Spanish, and Japanese. Survivors have written to us in 14 languages, and our replies and resource recommendations reach well beyond the three the website itself runs in. French and Hindi are the next two languages we plan to support in full.