The Voices Our Data Systems Were Built to Silence

For decades, we've relied on structured forms — checkboxes, dropdown menus, pre-coded response categories. Tools designed for analysts, not for the people living through crises.

I've managed needs assessments, response monitoring, and output reporting across Bangladesh, Ethiopia, Afghanistan, and Nigeria. The experience is the same everywhere: forms capture whether aid was received. They don't adequately capture what a mother actually needs, in her own words, with her own emphasis.

Data forms, by design, flatten context. They translate lived experience into categories someone in an office predetermined before going to the field. The voices the humanitarian agenda was built to uplift have been filtered through our tools before they ever reached a decision-maker.

We acknowledge we did the best we could with available resources. But "best we could" still meant: pre-coded forms, translated by intermediaries, interpreted by analysts, aggregated into dashboards that decision-makers read months later. The most vulnerable — women, children, displaced communities, people with disabilities — are represented as data points, not as people with context, priorities, and agency.

Every humanitarian data form is an act of pre-judgment. Someone in a capital city decides which questions matter, which response options exist, which categories are worth tracking. The beneficiary's job is to fit their reality into those boxes.

In Cox's Bazar, I coordinated data across 1,100+ radio listening groups in refugee camps. Our structured surveys still couldn't capture what displaced Rohingya families actually prioritised. The forms asked what we wanted to know. Not what they needed to tell us.

Accountability to Affected Populations has been a humanitarian commitment for over a decade. The principle is clear: affected people should participate in decisions that impact their lives. But look at how we actually collect data from them.

We design forms in English. Translate them — often imperfectly — into local languages. Train enumerators to ask questions in a specific sequence. Offer pre-coded response options. Record answers in categories built for aggregation and dashboards.

At every step, the beneficiary's voice is compressed. Their priorities filtered through our framework. Their context stripped to fit our schema. The people closest to a crisis have always had the answers. Our tools just weren't built to listen.

Even qualitative methods — the approach we trust to preserve nuance — pass through layers of interpretation. An enumerator translates. A researcher codes themes. An analyst writes findings. The original intent of the person who spoke has been reshaped at least three times before it informs a decision.

I've conducted several Key Informant Interviews in my humanitarian career, and during the COVID-19 pandemic, I led secondary data analysis using the DEEP platform with several steps of workflow designed to reduce cognitive bias. The rigour was real. But the original voices of affected populations were still mediated through documents written about them, not by them.

We did the best we could. And the results mattered — they informed decisions that affected millions of people. But the interface was always the bottleneck to evidence generation — not the data, not the analysis, not the people.

The question we need to ask ourselves is uncomfortable: in a sector built on the principle of centering affected populations, why have our data tools been structurally designed to exclude their direct input?

Voice-native data collection inverts the power dynamic entirely. It doesn't ask what we want to know. It asks: what do you need us to hear?

With voice data, a beneficiary speaks — in her language, with her priorities, with her emphasis — and AI captures that as structured, analysable data without stripping the context. The original recording remains as the auditable source of truth. She can verify it, correct it, update it. That's accountability to affected populations — not as a reporting checkbox, but as system architecture.

Modern voice AI doesn't just transcribe. It extracts entities, classifies urgency, detects sentiment, and maps speech to analytical frameworks — while retaining the original recording as the auditable source. The person's own voice becomes the data. Not an intermediary's interpretation of what they said.

This is what truly inclusive evidence generation looks like: voice as the default input. Not a supplement. Not an accessibility feature. The primary way affected populations contribute to the humanitarian evidence base. In their language. In their words. In their framing.

First — voice as the default input method. Not an alternative. The primary interface for how affected communities contribute to humanitarian evidence.

Second — AI-powered structuring that preserves context. Extract entities, classify urgency, map to analytical frameworks — while retaining the original recording as the auditable source of truth.

Third — multilingual by design. Google's WAXAL covers 21 African languages. Meta's Omnilingual ASR supports 1,600+. The infrastructure is arriving. Humanitarian systems need to integrate it now, not wait for perfection.

Fourth — beneficiary-owned feedback loops. When a person's spoken testimony is the data, they can verify it, correct it, update it. That's accountability to affected populations built into the system architecture.

Fifth — real-time evidence for real-time decisions. Voice collapses the collect-clean-analyse-report cycle into seconds. Decision-makers receive evidence while it's still actionable — not weeks after the situation has moved.

Voice data doesn't just improve collection. It restores the agency we've been designing out of our evidence systems for decades. After a decade of building platforms that run on forms, I'm now building the ones that run on voice.