AI Ethics in Humanitarian Action: When an Algorithm Influences a Human Life

AI ethics in humanitarian action means converting human rights, fairness, privacy, safety, transparency and accountability into operational controls. Ethical humanitarian AI uses representative data, meaningful human oversight, challenge and correction routes, and a named person who remains accountable for consequential decisions.

Published 2026-09-14 · By Shahzad Asghar

I have spent much of my career working around humanitarian data, information systems and operational decision-making. That experience has left me with a fairly simple view of artificial intelligence.

The important question is not whether AI can perform a task. The question is what happens when it is wrong.

That distinction matters in humanitarian work. An inaccurate product recommendation may waste somebody's money. An inaccurate recommendation in a protection, identity, health, eligibility or assistance process can affect somebody's rights, safety or access to services.

AI ethics in humanitarian action means converting principles such as human rights, fairness, privacy, safety, transparency and accountability into controls that govern real systems. It requires representative data, meaningful human oversight, a way to challenge errors and a named person who remains accountable for consequential decisions.

This is why AI ethics takes on a different meaning in humanitarian settings.

What is AI ethics?

AI ethics refers to the principles and controls used to make sure artificial intelligence systems respect human rights, privacy, fairness, safety, transparency and human accountability.

The Ethics of Artificial Intelligence therefore goes well beyond asking whether a model is accurate.

It asks who provided the data, who may be missing from that data, what decision the system can influence, who reviews the result, what happens when the system is wrong and whether an affected person can challenge the outcome.

UNESCO's Recommendation on the Ethics of Artificial Intelligence puts human rights and human dignity at the centre of its approach. Its principles include proportionality, safety, privacy, accountability, transparency, human oversight and non-discrimination.

These principles make sense almost anywhere. In humanitarian operations, however, they need to become operational controls. A policy document alone does not protect anyone. That implementation step is where AI ethics becomes AI governance.

AI ethics risks and the controls they require

Why AI ethics is different in humanitarian work

Humanitarian organizations often work with people who have limited choices.

A refugee cannot always choose another registration authority. A displaced family may not have several assistance providers to select from. A person seeking protection may have little understanding of how information about them moves between organizations, governments and technology providers.

There is therefore an imbalance of power before AI enters the process. AI can widen that imbalance.

Consider a model intended to identify households that may need assistance first.

On paper, this sounds reasonable. The organization has limited resources. There are thousands of cases. Data could help staff identify patterns faster.

Now consider what the model does not know.

Perhaps women heading households are under-represented in historical records. Perhaps people living outside formal settlements appear less frequently in the database. Perhaps one nationality has better documentation than another. Perhaps historic assistance decisions reflected funding restrictions rather than actual vulnerability.

The model may accurately reproduce those historical patterns. That does not make the result fair.

This is one of the most important distinctions in AI ethics: statistical accuracy and ethical acceptability are not the same thing.

The data problem comes before the model problem

Discussion about the Ethics of Artificial Intelligence often starts with algorithms. In humanitarian operations, I would start earlier. I would start with the data.

The ICRC has raised this issue directly. Its analysis of why data selection matters in humanitarian AI explains why training and evaluation data must be examined for their composition, quality, origin and relevance to humanitarian realities.

This should concern any organization using historical operational data.

A database is not a neutral description of a population.

It is a record produced by policies, access conditions, registration rules, staffing levels, funding decisions, security restrictions and previous programme choices.

If those records become training or reference data for AI, the assumptions embedded in the original process can reappear in a new system. Sometimes they become harder to see because the output now carries the authority of an algorithm.

This is why humanitarian data protection and AI governance need to begin before model selection.

Human oversight must mean more than approving a screen

Almost every responsible AI policy now refers to human oversight. The phrase is easy to write. Implementing it is harder.

A staff member clicking "approve" after reading an AI recommendation does not automatically constitute meaningful human oversight.

The person needs enough information, authority and time to disagree with the system.

This becomes especially important when staff process hundreds or thousands of cases. Once an AI recommendation appears on screen, people can develop a tendency to accept it unless something looks obviously wrong.

The human then becomes part of the automation rather than an independent decision-maker.

Meaningful oversight requires a clear division of responsibility. The AI can classify, summarize, translate, identify patterns or recommend. A named person remains accountable for decisions affecting rights, protection or access to assistance.

UNESCO explicitly states that AI systems should not displace ultimate human responsibility and accountability. That is a useful test for humanitarian organizations.

If nobody can clearly answer, "Who is accountable for this decision?", the AI system should not be in production.

Case study: UNHCR CareLine

A useful recent AI ethics example comes from UNHCR.

On 8 July 2026, UNHCR published information about CareLine, an AI-enabled system developed by its Regional Bureau for Asia and the Pacific with UNHCR Innovation.

The problem was straightforward.

UNHCR receives thousands of emails every month from refugees and asylum-seekers seeking information, assistance and updates. Staff need to review, classify and route those messages before the relevant team can respond.

CareLine helps identify possible duplicate requests, categorize enquiries and support triage.

The distinction between assistance and authority is important. According to UNHCR, CareLine supports the administrative process while all decisions affecting individuals continue to be made by UNHCR personnel under established protection standards and accountability frameworks.

That separation is a practical example of ethical AI system design.

The AI does not need authority over the refugee's case to provide value. It can reduce repetitive administrative work, help route information, identify duplicate correspondence and support staff. The consequential decision stays with a human.

This pattern should receive far more attention in discussions about responsible AI.

Organizations often ask, "What decision can we automate?"

A better starting question is, "Which parts of this process can AI assist without transferring accountability to the system?"

The second question usually produces safer architecture.

Where humanitarian AI can go wrong

There are several recurring areas where humanitarian organizations should be cautious.

Automated eligibility decisions

If an AI system decides whether a family qualifies for cash assistance, food, shelter or another service, errors have immediate consequences. Even a model with very high aggregate accuracy can produce unacceptable outcomes for smaller groups.

Refugee status and protection decisions

AI can help summarize documents, retrieve policy material, compare information or support administrative work. Allowing a system to determine whether an individual requires international protection is a very different proposition.

An analysis in the International Review of the Red Cross on AI, human rights and ethics in humanitarian action warns that AI is generally an inappropriate substitute for human decision-making in highly sensitive matters, including asylum decisions.

Biometric and identity systems

These systems may improve identity assurance and reduce duplicate records, but they also create long-term privacy and security questions. Biometric information cannot be reset like a password.

Generative AI advice

A chatbot answering questions about assistance programmes may sound confident even when its answer is incorrect. In an ordinary customer-service setting that may cause inconvenience. In a humanitarian setting, incorrect information about registration, documentation, deadlines, legal procedures or available services may have serious consequences.

The correct design response is not simply a disclaimer saying, "AI can make mistakes."

The system needs approved information sources, controlled retrieval, testing, logging, escalation paths and a clear point where a person takes over. These controls are also central to AI security and assurance.

AI ethics is a management responsibility

There is a tendency to assign responsible AI to the technology team. That is a mistake.

Technology teams can test security, architecture, data flows, access controls and model performance. They cannot decide by themselves whether an AI use is acceptable within a protection mandate.

That decision belongs to the organization.

Programme owners, protection specialists, legal advisers, data protection officers, cybersecurity teams, procurement colleagues and technology teams each hold part of the responsibility. Senior management owns the final risk decision.

The ICRC reached a similar conclusion in its institutional policy on artificial intelligence. Its approach anchors AI use in the humanitarian mandate and Fundamental Principles rather than treating AI policy as a standalone technology matter.

That is the right level at which to address the issue. AI ethics should sit inside institutional governance.

A practical test before approving humanitarian AI

Before approving an AI use case, I would expect an organization to answer seven questions.

1. Problem: What humanitarian problem are we solving, and why does it require AI?

1. Data: What information enters the system, and do we have a lawful and legitimate reason to use it for this purpose?
2. Harm: Which people or groups could receive a worse outcome if the system is wrong?
3. Authority: What decision can the AI influence, and what decision must remain with a person?
4. Transparency: Can staff and affected people understand how the system is being used?
5. Redress: How can an incorrect output be challenged, corrected and traced?
6. Accountability: Who is personally accountable for accepting the remaining risk?

If the team cannot answer these questions clearly, the project is not ready for deployment. The AI use-case intake form and AI risk-tiering template provide a practical place to record those answers.

Notice that none of these questions asks which model has the highest benchmark score. That comes later.

Privacy and cybersecurity are part of AI ethics

There is another issue that is sometimes separated from ethical discussions: security.

For humanitarian organizations, I do not believe AI ethics can be separated from cybersecurity and data protection.

A system may be fair and transparent but still be unacceptable if sensitive refugee information is sent to an external model provider without proper controls.

Teams need to understand where prompts are processed, how long information is retained, whether provider systems use that information for further model development, which administrators can access logs, where data is stored and whether sensitive information can leave approved environments.

The same applies to retrieval-based AI systems. Connecting a language model to internal documents does not automatically make it safe.

The system may expose information across access boundaries that existed in the original repositories. A user who could not open a protection file directly should not suddenly receive its contents because an AI assistant retrieved a paragraph from it.

Access control must follow the data, not merely the application. The risks and evidence requirements are examined further in LLM security for public institutions.

Responsible AI procurement

Much humanitarian AI will come from external providers. Procurement therefore becomes part of AI governance.

Contracts should address data ownership, retention, model changes, audit rights, incident reporting, subcontractors, hosting locations, security testing, service termination and deletion of organizational data.

Organizations should also plan for provider dependency. An AI service can change its model, pricing, terms, geographic availability or technical behaviour after deployment.

The humanitarian organization remains accountable even when the technology comes from somebody else.

A vendor contract transfers work. It does not transfer institutional responsibility.

What good AI ethics looks like in practice

Good AI ethics is usually less dramatic than people expect.

It can mean deciding not to automate a particular decision.

It can mean removing sensitive fields before data reaches a model.

It can mean testing results separately for different groups instead of relying on one overall accuracy figure.

It can mean showing staff the source behind an AI-generated answer.

It can mean keeping a person in control of protection decisions.

It can mean allowing communities to report when an AI-supported service gives them incorrect information.

It can mean shutting down a use case when the organization cannot reduce the risk to an acceptable level.

The ICRC's AI policy takes a responsible, human-centred approach to using AI in support of its humanitarian mission. UNHCR's AI Approach similarly places human rights, transparency, participation and human oversight at the centre of the AI lifecycle.

This is where the Ethics of Artificial Intelligence becomes practical: not in the principle itself, but in the decision made because of it.

Frequently asked questions

What is AI ethics in simple terms?

AI ethics is the set of principles and controls used to make sure artificial intelligence is developed and used in ways that respect people, rights, privacy, fairness, safety and accountability.

Why is AI ethics important in humanitarian organizations?

Humanitarian organizations often work with sensitive personal data and make decisions affecting people in vulnerable situations. An incorrect, discriminatory or insecure AI system can therefore cause direct harm or restrict access to protection and assistance.

What are examples of ethical issues in artificial intelligence?

Common examples include biased outcomes, use of personal information without an appropriate basis, opaque automated decisions, incorrect AI-generated information, weak cybersecurity, surveillance, inadequate human oversight and the inability to challenge an automated outcome.

Should AI make decisions about humanitarian assistance?

AI can support analysis and prioritization, but decisions that materially affect a person's rights, protection or access to essential assistance require meaningful human responsibility and an appropriate review process. The level of automation should depend on the potential harm if the system is wrong.

What is the difference between AI ethics and AI governance?

AI ethics defines the principles an organization expects its AI systems to respect. AI governance turns those principles into authority, policies, approval processes, risk assessments, technical requirements, monitoring, audit and accountability.

Can generative AI be used safely in humanitarian work?

Yes, for suitable use cases and with appropriate controls. Lower-risk applications may include summarization, translation, knowledge retrieval and administrative assistance. Systems using sensitive information or influencing protection and assistance decisions require stronger controls, testing, monitoring and human review.

Who is responsible when an AI system makes a mistake?

The organization deploying the system remains responsible for how it is used. Responsibility should not be assigned to an algorithm. Organizations need named owners for the service, the data, the operational process and the risk decision.

The question humanitarian leaders should ask

Artificial intelligence will become part of humanitarian operations. In many areas, it already has.

The choice facing leaders is therefore not between using AI and avoiding AI. The choice is what authority we give it.

AI can help people process information faster, identify patterns that humans may miss, communicate across languages and reduce administrative workload.

But efficiency cannot be the only measure.

Humanitarian organizations exist to serve people facing displacement, conflict, disaster and other forms of vulnerability. Technology has to operate within that mandate.

For me, that is the practical meaning of AI ethics.

Before asking whether an AI system works, ask who could be harmed when it does not.

Before asking how much work it can automate, ask which responsibility must remain human.

And before putting an AI system into production, make sure somebody is willing to put their name against the decision to use it.

That is where responsible artificial intelligence starts.

Sources and further reading

• UNESCO Recommendation on the Ethics of Artificial Intelligence
• UNHCR CareLine: Using AI responsibly to help answer refugees' queries
• UNHCR AI Approach
• ICRC: Why data selection matters in humanitarian action
• ICRC policy on artificial intelligence
• International Review of the Red Cross: AI for humanitarian action, human rights and ethics

← All articles