AI and the Do No Harm Principle: Can They Coexist? (2026)

AI and the Do No Harm Principle in 2026

The question of whether AI and the Do No Harm principle can coexist in humanitarian work is more than rhetorical. Do No Harm, developed by Mary Anderson and codified through years of humanitarian practice, asks aid actors to recognize that intervention is not neutral and that programs can inadvertently fuel conflict, inequity, or dependency. AI systems introduce a new layer of intervention, often invisible to affected populations, that can cause harm at speeds and scales the original framing did not anticipate. Reconciling the two is the central ethical task of the sector in 2026.

The short answer is that they can coexist, but only under conditions that most humanitarian agencies have not yet built. Coexistence requires extending the Do No Harm framework to cover algorithmic decisions, treating the people affected by those decisions as participants rather than data points, and accepting that some AI deployments will fail the test and should not proceed.

What Do No Harm Originally Asked

The original framework asks aid actors to map the connectors and dividers in a community, to assess how aid resources and implicit messages affect those dynamics, and to redesign programs that exacerbate harm. The framework was built for material aid: food, shelter, cash, and services. It assumed that the unit of analysis was a program and that the unit of harm was a community.

AI systems break both assumptions. The unit of intervention is often individual: a vulnerability score, a routing decision, a chatbot interaction. The unit of harm can be invisible: a denied benefit that the affected person never knew was algorithmically determined, a misclassification that propagates through subsequent decisions, a re identification that exposes someone to a risk they could not anticipate. Mapping connectors and dividers in this environment requires tools the original framework did not provide.

How the Sector Is Extending the Framework

The 2026 practice has begun to extend Do No Harm to cover algorithmic decisions. Algorithmic impact assessments now sit alongside conflict sensitivity assessments in the deployment process at the agencies that have moved fastest. The IASC operational guidance on data responsibility, the ICRC handbook on data protection, and CALP Network guidance on responsible cash and digital programming all reference Do No Harm directly. The OECD AI Principles and the UNESCO Recommendation on the Ethics of AI provide reference points that some humanitarian agencies have begun to incorporate.

The extension is uneven. The Core Humanitarian Standard 2025 update references algorithmic decisions affecting beneficiaries but does not yet provide a full audit methodology. For the broader risk context, see The Risks of AI in Humanitarian Work.

Where Coexistence Breaks

Three deployment patterns recur in cases where Do No Harm and AI are hardest to reconcile. The first is opaque vulnerability scoring used to ration scarce assistance. When the model cannot be explained to affected populations and the decision cannot be contested, the deployment violates the spirit of Do No Harm regardless of how technically sophisticated the model is. The second is biometric registration in contexts where the host government has incentives that diverge from refugee protection. The technical security of the biometric system does not address the political risk. The third is AI mediated protection intake without adequate clinical and ethical oversight. Survivors disclosing sensitive information to chatbots may not understand what is being recorded or shared.

In each of these patterns, the test is not whether AI improves operational efficiency. It is whether the affected population is better off in expected value after accounting for the harms the deployment introduces. That test is rarely run formally in 2026.

What Coexistence Looks Like in Practice

The agencies that are reconciling the two principles share a small set of practices. They map the connectors and dividers that the AI deployment will affect, including the dividers introduced by the deployment itself. They engage affected populations in the design and review of the system. They publish algorithmic impact assessments and revise the deployment based on the findings. They accept that some deployments will fail the test, and they walk away from those deployments even when the operational case for them is strong. They invest in the governance infrastructure that makes all of this possible.

The honest summary in 2026 is that AI and Do No Harm can coexist when the institutional conditions for governance are in place and cannot coexist when they are not. The sector's central task is to build those conditions before the next generation of AI deployments scales further. For context on the governance gap that defines the moment, see What Is the Humanitarian AI Paradox.

Sources and Further Reading

• Core Humanitarian Standard and Verification Scheme: https://www.corehumanitarianstandard.org/
• IASC operational guidance on data responsibility: https://interagencystandingcommittee.org/
• ICRC handbook on data protection in humanitarian action: https://www.icrc.org/en/data-protection-humanitarian-action-handbook
• OECD AI Principles: https://oecd.ai/en/ai-principles
• UNESCO Recommendation on the Ethics of Artificial Intelligence: https://www.unesco.org/en/artificial-intelligence/recommendation-ethics
• CALP Network on responsible cash and digital programming: https://www.calpnetwork.org/
• ALNAP guidance on responsible AI in humanitarian action: https://www.alnap.org/

Standards and frameworks above reflect publicly documented agency policies and peer reviewed research published through mid 2026.