Personally identifiable data in humanitarian contexts is like nuclear fuel – it has immense power to do good, enabling us to reach millions with life-saving aid, but it’s inherently dangerous. Like nuclear material, once it leaks, the damage can’t be undone. And in humanitarian work, we must operate under the assumption that eventually, it will leak.
Let’s use Yemen as an example. The crisis there presents a perfect storm:
- Dwindling humanitarian funding,
- Complex identification challenges across conflict lines,
- Millions in desperate need of immediate assistance.
Artificial intelligence could offer powerful solutions for ensuring fair and efficient aid distribution, yet we must confront an impossible ethical choice in a context where consent cannot be truly informed:
Is consent ethically meaningful when the alternative is starvation, or do we compromise on our principles of informed consent to save lives?
Reality in Yemen: Too Many IDs
AI solutions are not currently being deployed in this way in Yemen, but the country’s situation provides a compelling case study for discussing these ethical dilemmas.
In Yemen, humanitarian agencies must navigate a byzantine landscape of 26 different forms of functional IDs – issued by pre-war authorities, current local administrations, warring factions, and various administrative bodies. Each claims legitimacy, and many beneficiaries hold multiple, sometimes conflicting IDs.
Traditional matching methods collapse under the weight of different Arabic name variations, inconsistent household definitions, and programs operating across conflict lines. Add to this the substantial population lacking any formal ID, and you begin to understand why AI’s promise of efficient identification is so tempting – and so ethically complex.
Deduplicating Identification Overload
Humanitarian actors in Yemen face an impossible daily task: ensuring that the same person isn’t registered multiple times across various programs, especially when they might present different forms of ID each time is a major challenge and that traditional matching methods often fall short on.
Traditional matching methods for external and internal deduplication struggle under the weight of operational realities:
- Programs operate simultaneously across conflict lines with different registration points
- Multiple implementing partners use incompatible systems
- Data collection methods vary widely
- Arabic names can have multiple valid spellings and variations
- Even basic location data becomes unreliable due to inconsistent transliterations
- Household compositions shift as families seek safety
The result? In a context where humanitarian funding has dropped, aid workers face an impossible choice: risk excluding people in desperate need through overly strict verification, or risk depleting scarce resources through double registration. When every dollar matters and each delayed decision can mean life or death, we need better solutions.
AI Opportunity for Deduplicating IDs
While not currently implemented in Yemen, recent advances in AI and machine learning could theoretically offer solutions for humanitarian contexts like:
- Probabilistic matching algorithms could compare name variations across Arabic spellings
- Machine learning models could identify household composition patterns
- Natural language processing could standardize location data despite spelling differences
- Advanced entity resolution could link different functional IDs
- Risk scoring could flag potential duplicates for review
Like successful systems in social protection and private sector programs, any AI solution would require careful human supervision along these four steps:
- AI flags potential matches based on probability scores
- Program staff review high-confidence matches
- Local staff verify cultural naming patterns
- Field teams conduct physical verification when needed
Humanitarian contexts pose unique challenges around data sovereignty, conflict sensitivity, and the urgency of life-saving assistance. Responsible AI solutions must be adapted for these realities.
Responsible AI Ethical Trap
Now imagine a humanitarian aid worker in Yemen using an AI solution to solve the identification deduplication challenge. Every day in Yemen, they greet families arriving at registration points with multiple functional IDs from different authorities, desperately needing immediate food assistance.
They would need to say something like:
“You are obviously a person in Yemen who is in a famine. I want you to give me your informed consent to share your personal information to an AI model I didn’t create, that’s on a server in America, to do something that I don’t really understand, to help me do my work. If you don’t give me that informed consent, you will not get food or lifesaving assistance.”
They – and we – face an impossible choice:
- Require consent for AI processing they (and we!) can’t meaningfully understand
- Maintain parallel non-AI systems we can’t afford
- Exclude them from life-saving assistance
The core dilemma for obtaining consent in emergency settings: do we deny life-saving assistance to those who cannot meaningfully consent to AI-powered identification systems, when the alternative is starvation?
This forces us to confront uncomfortable questions about power, privilege, and paternalism in humanitarian aid.
- Are we letting perfect be the enemy of good?
- What level of consent is truly ethical when facing urgent needs with limited resources?
NetHope emphasizes that humanitarian AI must “do no harm” while ensuring effective governance and accountability. Meanwhile, organizations like GiveDirectly are pioneering approaches to ethical AI implementation in cash assistance programs, demonstrating how tiered consent models can help balance urgent needs with ethical considerations.
The widespread availability of AI tools forces us to directly confront challenges that have always existed in humanitarian aid – but were previously theoretical or easier to ignore. When every field officer has access to powerful AI tools on their phone, abstract ethical discussions become daily operational decisions.
Moving Forward Together
The humanitarian sector can’t wait for perfect solutions while urgent needs go unmet. We must start addressing these ethical challenges now, learning and adapting as we go. Your experiences and insights matter in this conversation.
- Have you faced similar ethical dilemmas in your humanitarian work?
- How do you balance technological innovation with responsible implementation?
While we may not have all the answers today, the decisions we make now will shape how AI serves humanitarian needs tomorrow. Let’s ensure we make these decisions together, keeping both ethics and human dignity at the center of our innovation.
By Thomas Byrnes and originally published as The Ethics of AI Identity Matching in Humanitarian Aid: When Perfect Consent Meets Impossible Choices
The article touches on a few good questions (and has sparked a few interesting comments over at LinkedIn), but I think misses a few important points.
The foremost topic, that has sparked a lot of good research over the last years, is how to gather meaningful consent in the “digital era” or in other complex situations (in this case: how do you operationalize aid if explaining what you are asking consent for would take a lot of time, how to consider the power imbalances/the extreme emergency/hardship/etc). The article doesn’t really touch on this body of work and instead asks questions that don’t seem the right questions to me.
Another topic is why we are even talking about this, which includes donor reporting, fairness, fraud reduction, etc. The article seems to assume that GenAI is the only answer to this, and the existence of too many IDs the root cause of a problem that is not even clearly articulated here.
GenAI is a technology we don’t fully understand yet. At the same time, it seems to hold so many yet-to-be-proven promises that it sometimes seems to cloud our judgement. A simple exercise I always conduct is to replace GenAI, etc., by “shiny new technology”.
“Is Requiring Informed Consent to [shiny new technology] Ethical in Humanitarian Response?”
“Responsible [shiny new technology] Ethical Trap”
Are we exempt from doing a proper risk analysis before using [shiny new technology] in settings with the most vulnerable people because of the urgency on the one side and the promises of [shiny new technology] on the other?
Are we exempt from doing a proper analysis of what problems we are trying to solve, to _then_ try to see which potential solutions we might come up with that can be applied in our context, _then_ do risk analyses, including ethical risks, and _then_ arrive at an informed decision? Maybe the author did all of this, but the article doesn’t really show.
Instead it creates false dichotomies like this one: “do we deny life-saving assistance to those who cannot meaningfully consent to [shiny new technology]-powered identification systems, when the alternative is starvation?” or “Are we letting perfect be the enemy of good?”
I think a large majority would rightfully argue that this is not consent. As part of the interviews I was doing during my PhD research, this kept coming up, and it was flagged as a big problem and one that should force us to rethink our approach to data collection.
This thread resonates with me. When there’s an asymmetrical power dynamic, calling it ‘informed consent’ feels disingenuous—teams should ask if it’s protecting the person in need or shielding the organization. A better term might be ‘data use disclosure.’
To be constructive, organizations can flip the burden back onto themselves. If a tool demonstrates clear benefits over alternatives, then:
a) Only collect data directly tied to saving lives—nothing more.
b) Implement strict data expiration timelines with secure deletion protocols.
c) Prioritize tools and vendors using privacy-enhancing technologies.
d) Co-design and iterate tools with users to surface unknown risks early.
Ultimately, the key question is whether the tool provides enough value compared to other methods and whether priority risks are mitigated and communicated. While I respect the ‘do no harm’ approach championed by NetHope, I think it’s unrealistic—doctors give vaccines knowing there are risks but also that the strategy serves the greater good, with sufficient checks to mitigate issues. Organizations should approach tool design with the same mindset: informed, transparent, collaborative, and carefully weighing trade-offs.
Excellent question! I think there are some instances when rapid support is more important than gaining ethical consent which can rob valuable time. Here, we would use ‘legitimate interest’.
For more complex matters, I think its worth considering the idea of social licenses which strive to establish ongoing trust and acceptance through community engagement with a smaller group of people who can take time to understand and consider complex topics, and participate in determining how data is accessed and reused for the whole community.
Makes me wants to create a decision matrix that plots risk and complexity, and advises on the right type of consent for each use case.
Thanks for highlighting this significant ethical issue. It goes without saying that compromising on our data collection principles, even to save lives, is not justifiable as it undermines the very ethical framework we claim to stand for. As much as the situation may present a stark ethical dilemma, standing firm on principles like informed consent is crucial for maintaining the integrity of humanitarian efforts. It’s unfortunate that this situation happens more frequently than we would like to acknowledge.
Though the AI dimension is novel, the questions regarding what constitutes ethical data management – and ethical for whom – are familiar. This particular article provides a good opportunity to invite reflection not only on the form and function of AI in participant registration, but also the paradigm of practice in which we tend to invert the responsibility hierarchy to place much of the compliance burden on communities, whilst prioritising the convenience of humanitarians (who theoretically have the means and resources to absorb the heavy lifting). It’s valid to ask how AI will be scaffolded onto that which already exists, but I’d encourage us to also question whether standards they will build upon are truly reflective of “best practice”, or if our imagination has just been limited by that which came before.
What I don’t really understand here (or from scanning the article) is the lifesaving element? What seems to be being suggested is a de-duplication which should make aid delivery more efficient and maybe more fair. (I know it’s an example, but it’s a good one to discuss.)
Those are probably good things to do. I can imagine how at the response level, they might create some cost savings which save some lives. But from what I can see it is a fairly long causal chain between “analysing profile information” to “life is saved” . If the benefit is efficiency savings, then there are a lot of ways to run aid programmes more efficiently that don’t involve compromising people’s privacy/violating their consent.
What am I missing?
thank you for facilitating this excellent discussion. as others have already said, information and power asymmetries make “informed” consent nearly impossible in any such context/ scenario. also, it doesn’t make sense to put the onus/ burden on the recipients to “opt out” or express reservations to such a mechanism- while individual privacy rights need to have their place in legal frameworks.
scaling the exercise of such rights is virtually impossible in the humanitarian context as it requires a separate mammoth machinery on its own- channels to express “opt out” or reservations or withdraw consent, people to process, systems to record such reservations/ withdrawal etc. making this an “industry” in and of itself.
what humanitarian organisations need to look at moving forward is not just less invasive or data minimized measures for deduplication- but really rethinking deduplication from ground up- programme conceptualisation implementation in particular, and also how inter-agency collaboration could create centralised hubs/ blockchain (or other mechanism) based systems that would allow authentication of unique recipients, and thus avert duplication.
Powerful observation. I think it will take strong and collective efforts by the development professionals to stop this unethical practices.