Who Is Responsible When AI Gets It Wrong in a Refugee Crisis? (2026)

Who Is Responsible When AI Gets It Wrong?

The question of who is responsible when AI gets it wrong in a refugee crisis in 2026 has no single answer, which is itself the problem. A misclassified vulnerability score that blocks a household from cash assistance, a biometric mismatch that denies a registered refugee a food ration, or a language model output that misrepresents a survivor's testimony each implicates a chain of actors: the model developer, the humanitarian agency that deployed the system, the staff member who acted on the output, the host government that may have shaped the data inputs, and the donor that funded the program. In practice, the consequences fall first on the displaced person and the search for responsibility starts only after harm has occurred.

For context on the broader risk landscape, see The Risks of AI in Humanitarian Work.

The Chain of Actors

Five categories of actors typically sit between an AI system and the person it affects. The model developer designs and trains the underlying system, often a commercial provider whose terms of service disclaim liability for downstream use. The humanitarian agency procures, deploys, configures, and monitors the system. The implementing partner runs the day to day operation, often in a field office with limited technical capacity. The staff member uses the tool to make or recommend a decision. The host government and the donor shape the operating environment that determines what can be deployed and audited.

Each of these actors has plausible deniability when something goes wrong. The model developer points to the deployment context. The agency points to the model provider. The implementing partner points to the agency. The staff member points to the institution. The host government and the donor point to the operational chain. The result is a diffusion of responsibility that the displaced person cannot easily navigate.

What 2026 Accountability Frameworks Say

The dominant accountability framework in the sector is the Core Humanitarian Standard, supplemented by the IASC Operational Guidance on Data Responsibility and agency specific data protection policies. The 2025 update to the CHS Verification Scheme adds explicit reference to algorithmic decisions affecting beneficiaries. UNHCR's revised data protection policy, ICRC's handbook on data protection in humanitarian action, and OCHA's data responsibility guidelines all converge on a similar set of expectations. Agencies that deploy AI are accountable for the outcomes, regardless of where the model was built. Agencies must conduct algorithmic impact assessments before deployment. Agencies must provide accessible grievance mechanisms. And agencies must be able to reconstruct individual decisions after the fact through versioned model logs and decision documentation.

The standards are clearer than they were two years ago. Compliance remains uneven. A 2025 ALNAP review found that fewer than 20 percent of AI deployments at the surveyed agencies met the full set of expectations.

Where the Accountability Gaps Are Widest

Three gaps recur across operational reviews. The first is in the procurement layer. Agencies often deploy commercial AI tools through general purpose enterprise contracts that do not address humanitarian specific risks. The contracts allocate liability to the agency without giving the agency the technical access needed to audit the model. The second gap is in field implementation. Staff using AI tools rarely have the training to recognize when an output is wrong and the workflow does not always reward the time it takes to challenge it. The third gap is in the grievance mechanism. Affected populations rarely know that an AI system was involved in a decision that affected them, which makes it impossible to lodge a meaningful complaint.

The El Fasher protection case audits in Sudan, the Ukraine border registration reviews in 2022 and 2023, and the Gaza vulnerability scoring reviews in 2024 each surfaced versions of these gaps. None of them produced individual accountability for any specific decision. All of them produced operational improvements at the institutional level. That asymmetry, no individual accountability and gradual institutional learning, is the modal outcome in the sector.

Where the Practice Is Heading

Several developments in 2026 may narrow the gap. The first is a slow shift toward contractual clauses that require model providers to disclose training data sources, retention practices, and known performance limitations to humanitarian deployers. The second is the emergence of independent algorithmic auditors with humanitarian sector experience, who can review models and decision pipelines on behalf of agencies and donors. The third is the integration of AI specific questions into the Core Humanitarian Standard verification process, which gives agencies a stronger incentive to invest in governance infrastructure.

None of these developments resolve the underlying tension. Humanitarian work happens in environments where speed is essential, capacity is constrained, and the people most affected have the least power. AI systems amplify all three of those conditions. The honest summary in 2026 is that responsibility for AI failures in refugee crises sits with the deploying agency in principle, defaults to the field staff member in practice, and is borne by the displaced person in outcome. Closing that gap requires investment in governance infrastructure that the sector has not yet funded at scale.

Sources and Further Reading

• Core Humanitarian Standard and Verification Scheme: https://www.corehumanitarianstandard.org/
• UNHCR data protection policy: https://www.unhcr.org/data-protection
• ICRC handbook on data protection in humanitarian action: https://www.icrc.org/en/data-protection-humanitarian-action-handbook
• IASC operational guidance on data responsibility: https://interagencystandingcommittee.org/
• OCHA Centre for Humanitarian Data: https://centre.humdata.org/
• ALNAP reviews of AI in humanitarian action: https://www.alnap.org/
• CALP Network on accountability in cash programming: https://www.calpnetwork.org/

Assessments above reflect publicly documented standards, agency policies, and peer reviewed evaluations published through mid 2026.