Aileen Voit

AI is a mirror, not a crystal ball. Fed biased data, it automates the bias rather than fixing it. The question isn't only whether the system works on average, but whose lived experience shaped the training set, and whose was missing.

Backlogs make speed seductive. But speed strips the moral weight of decisions that change lives. The trade-off isn't just operational. It's ethical.

Information collected to assess one person's claim can end up accessible to the regime they fled. Migration data is among the most sensitive personal information we hold, and accountability for it rarely follows the file.

A tired caseworker rubber-stamping an AI output is not oversight. Genuine human judgement requires time, authority, and the option to disagree. Three things overworked systems rarely give.

When AI is used for targeting, prioritisation, or vulnerability scoring, it can quietly leave people out. Nobody notices. An algorithm trained on incomplete data doesn't raise a flag when it drops someone from an assistance list. It just does it.

In humanitarian contexts, that can mean someone doesn't receive food, shelter, or protection. The harm is silent and almost impossible to spot without someone actively looking. WFP has already faced scrutiny over algorithmic exclusion errors in assistance targeting.

The ICRC has put it clearly: AI systems trained on data from well-funded, digitally visible populations create blind spots that confuse "historical visibility" with actual humanitarian need. This reinforces cycles of over-recognition and neglect.

Staff paste sensitive information about the people they serve into ChatGPT, Claude, and other commercial tools every day. Case notes, interview summaries, protection assessments. These tools were not designed for humanitarian data. They were not built with the sensitivity of refugee and migrant information in mind.

Once data enters a commercial system, the organisation has lost control of it. And 69% of humanitarian AI use relies on commercial providers. That creates dependency on companies whose priorities don't match humanitarian ones.

In a sector that handles some of the most sensitive personal data in the world, this should be treated as a serious protection risk, not a convenience question.

AI tools are designed to make processes faster and more consistent. But humanitarian work — whether it's protection, MEAL, livelihoods programming, or shelter allocation — depends on context and judgement that no dataset can replace.

A protection officer who overrides a standard output because they recognise a pattern from experience. A MEAL officer who questions an AI-generated analysis because they know the data collection context. A programme manager who adjusts a targeting recommendation because they understand local dynamics the model doesn't. That is exactly the kind of expertise that keeps responses effective and people safe.

As the ICRC has written, AI systems "are never fully neutral — they are designed, developed, and deployed in a specific context, by specific actors, and with specific purposes in mind." Across all areas of humanitarian work, the stakes of forgetting that are high.

Most organisations don't know what AI tools their staff are using, what data is going into them, or what decisions are being shaped by AI outputs. This isn't a theory. It's happening right now.

Access Now's March 2026 research shows how AI enters humanitarian operations not through big decisions, but through software updates, vendor changes, and tools built into products staff already use. Cloud-based AI is quietly entering most internal systems through changes organisations never assessed or approved.

Daily AI use is highest in Kenya (65%), Sudan (60%), and Bangladesh (59%), often in settings where formal oversight is minimal. Staff adopt whatever works because they're doing three jobs at once. Without visibility, there is no governance. And without governance, there is no accountability when something goes wrong.

The EU AI Act classifies migration and asylum as high-risk domains. There is no humanitarian exemption.

AI systems used in asylum, visa, and residence permit processes face strict rules from August 2026. Penalties reach up to EUR 35 million or 7% of global turnover. For any organisation receiving EU funding or operating in EU jurisdictions, this is not abstract.

But there's a critical blind spot. The Act is strongest where states use AI in formal migration management. It says much less about humanitarian organisations using commercial AI tools in related work — protection assessments, case management, translation, programming decisions. That leaves the greatest volume of real-world risk in the hands of organisations not covered by the strongest parts of the law.