User Journey Through Humanitarian Identification Systems

The past decade has witnessed an immense digital transformation in the humanitarian sector. Among the most consequential aspects of this shift is the emergence of digital identity systems, which have become central to how organizations register, track, and serve displaced populations. As digital identity systems are integrated into humanitarian practices, they become a site of contestation and reflection on the role of technology and the manner in which it mediates humanitarian engagement with beneficiaries.

This brief examines the possible implications of ongoing AI advancements for humanitarian digital ID systems at each stage of the humanitarian journey, offering a synthesis of an often-complex process alongside informed speculation on emergent areas for humanitarian attention.

Rapid advancements in AI have also altered the operational ecosystem of humanitarians. AI tools are being explored to strengthen fraud detection, improve biometric accuracy, optimize registration processes, and protect the integrity of digital ID databases. As machine learning and AI models mature, researchers and humanitarian actors are looking for novel ways to apply them to increasingly complex humanitarian challenges. Some of these applications correspond directly to new AI-enabled challenges such as synthetic identity generation and falsified biometrics, which could disrupt the reliability of humanitarian operations if not properly addressed.

At each stage of the humanitarian journey, individuals interacting with systems encounter different iterations of both the identification process and the possible applications of AI. By considering the interaction, verification/authentication needs the documentation required, and the potential role of AI, this brief distills both a complex administrative process and a rapidly evolving technology to ground future humanitarian discussion regarding this critical topic.

Humanitarian journey

1. Initial registration and intake

The interaction

When people first arrive or present themselves to humanitarian service providers, they usually encounter a registration process. Registration may begin with group preregistration to collect core data on households traveling together, primarily used to organize population movements and facilitate initial assistance distribution. The subsequent initial registration process for many humanitarian organizations, especially non-governmental organizations (NGOs), typically consists of collecting the individual’s name and names of their family members. However, individuals seeking “refugee status” with UN agencies are asked for more information and undergo an initial registration interview in which their personal biographical data is collected.

Verification/Authentication

Registration staff record personal data including names, family composition, place of origin, and demographic information. In many cases, individuals present any existing identity documents from their country of origin, though refugees often lack such documentation due to the circumstances of their flight.

Documentation required

Preexisting identity documents (e.g., national ID cards, passports, birth certificates), if available. However, the lack of documentation does not prevent registration, as humanitarian agencies recognize that displaced persons may have fled without documents.

Potential role of AI

Adversarial applications

AI-generated fake documentation
During initial registration when individuals provide documentation to prove identity, AI-generated forged documents pose significant risks. Humanitarian operational needs require processing a wide range of identification documents from different origins, making verification difficult. Many AI models are trained on legitimate documents, meaning they can generate counterfeits with precision. Such AI-generated documentation can capture both subtle markers and anti-counterfeit measures like shadows or signatures. 1

For example, a recently exposed service called “OnlyFake” allowed anyone to automatically generate convincing IDs and offered to produce large quantities of fake ID documents. Though the service claimed to use neural networks for photo and signature generation, the exact AI techniques utilized remain vague. 2 Regardless, the implications of these services are clear: counterfeit identification documents are not only easier than ever to produce with generative adversarial networks and other AI tools, but also remarkably convincing. 3

Complementary applications

AI-enhanced document authentication
AI-driven identity verification uses machine learning, computer vision, and natural language processing to assess documents. AI tools can rapidly validate documents and detect sophisticated forgeries, 4 ensuring document authenticity. Automated verification systems can verify identities by cross-referencing information from diverse sources and analyzing document authenticity, which speeds up onboarding while minimizing errors.

Critical considerations

Facial recognition systems show significant demographic bias; studies find error rates of 34.7% for darker-skinned women, compared to 0.8% for light-skinned men. 5 AI systems trained on images of young white males are most biased against women of color. 6 Algorithmic bias in automated decision-making may amplify historical biases where past biased decisions are recognized as norms and applied at much greater automated scale.

2. Biometric enrollment

The interaction

Biometric enrollment is increasingly common within humanitarian systems. Notable examples include the United Nations High Commission for Refugees’s (UNHCR) Population Registration and Identity Management Ecosystem (PRIMES) and Biometric Identity Management System (BIMS), the World Food Programme’s (WFP) SCOPE, and the International Organization for Migration’s (IOM) Biometric Registration and Verification System (BRAVE). Enrollment commonly takes place during or immediately following initial registration and/or interview, and generally includes fingerprints and/or iris scans.

Verification/Authentication

• Biometric technology captures physiological characteristics used to uniquely identify individuals. This typically involves fingerprinting, iris scans, and faceprint capture that can be used for facial recognition. Biometrics ensure that identities registered in the system are globally unique and cannot be duplicated. Some humanitarian organizations argue this can mitigate multiple registration and beneficiary fraud.

• Individuals are assigned a unique identifying number in the registration system.

Documentation required

• Consent is obtained from individuals to collect biometric data and share basic personal data with partners and humanitarian organizations for the purposes of providing access to protection, assistance, and services—though there are concerns around “informed consent.” 7
• For children, consent is typically obtained from accompanying parents or guardians.

Potential role of AI

Adversarial applications

Morphed biometrics
“Facial morphing,” a method where two different faces are digitally merged into a new single image, has already proven able to successfully trick both human observers and facial recognition systems. 8 If misused in identity documents, a single morphed face can biometrically authenticate the identities of two individuals, allowing one ID document to be shared between them. One especially infamous example of facial morphing in 2018 saw German activists blend their photo with that of an EU official and successfully obtain a legitimate passport using the morphed image. 9 In humanitarian settings, morphed biometrics may be used by non-registered individuals to pass as a registered person and fraudulently access benefits and services. It should be noted that there is minimal incentive for two individuals to intentionally morph their images, as only one would actually receive the allotted assistance. Facial morphing thus would be used only in an adversarial manner, raising the risk that vetted and deserving individuals may subsequently be denied their assistance.

Morphed biometrics undermine the data integrity of the digital ID records into which they are integrated, skewing critical data that humanitarians rely on to inform their operations. Accurate understanding of population movement and demographic details regarding the number of affected people and potential regions or populations of priority forms the bedrock of data input for humanitarian needs assessments and response planning analytics. 10 Subsequent verification exercises are also pivotal in maintaining operational transparency and allow humanitarian organizations to adjust resource allocation strategies based on updated population figures. 11

The risks of morphed facial biometrics is especially pronounced in population tracking efforts and border-crossing monitoring, where it could enable non-registered individuals access to asylum services without undergoing the required screening. This raises potential counterterrorism and public safety concerns, as unregistered individuals and their backgrounds remain completely unknown to humanitarian organizations and host countries alike. Blanket assumptions of malicious use of facial morphing are likely to overlook the use of facial morphing as a reaction to ever-increasing hostile immigration policies and narrowing pathways to legal asylum. However, even in scenarios where non-registered individuals do not have malicious intent, their use of facial morphing is likely to directly impact registered and verified individuals whose identities they have morphed with, and who may bear long-term consequences or denial when they come to claim asylum.
For humanitarians, the spillover of morphed biometrics into migration processes can disrupt resettlement efforts and make it more difficult to ensure individuals can relocate as part of resettlement efforts, asylum processes, and legal migration routes.

AI-generated synthetic biometrics
Deepfake imagery and video can mimic a person’s facial appearance, expressions, and even voice in real time. 12 One study used machine learning algorithms to create a fake fingerprint—dubbed “DeepMasterPrints”—which exploits a system’s partial match tolerances to enable successful access to touch-based authentication systems. 13 These emerging studies illustrate a possibility that, in the future, false biometrics produced by sophisticated AI models could threaten humanitarian digital ID systems during identity verification processes. Hypothetically, false biometrics could introduce fraud into beneficiary authentication into aid registration and distribution, allowing for the theft or channeling of humanitarian aid to unauthorized actors.

Typically, digital ID systems focus on the delivery point at which services are accessed. However, humanitarians also utilize digital ID systems within their supply chain verification, to track the movement and access of vast amounts of assistance and limit fraud in the supply chain such as aid diversion—when aid is diverted from intended recipients (e.g., beneficiaries, NGOs, governments, etc.)—and aid theft—when actors gain unauthorized access to resources. Given that aid diversion and theft are already challenges for humanitarians, the entrance of AI-generated deepfakes furnishes well-placed supply chain actors with new tools to facilitate aid redirection and circumvent humanitarian controls and oversight.

Complementary applications

AI-enhanced liveness detection
One element of distinguishing real biometrics from fake is to verify beyond the biometric sample itself. Enhanced liveness detection uses an AI model to analyze biological cues such as eye blinking, facial muscle movement, and skin texture responsiveness to prompts to verify that a person undergoing biometric authentication is physically present. 14 These capabilities prevent common methods used to bypass remote biometric verification, including using printed photos and video replays, and potentially the kind of static morphed or AI-generated biometrics discussed earlier. 15

Moves to expand humanitarian reach, while reducing the need for personnel, have prompted interest in supporting the capacity of beneficiaries to register and update their biometrics remotely. Though this may help support access and continued services, it also introduces challenges regarding oversight. The implementation of liveness checks has already proven useful in humanitarian operations where biometric digital identification is used to access aid remotely. In 2023, a UNHCR operation in Iraq distributed EyePay Phones to enhance cash distribution activities. The mobile phones, produced by the payment authentication company IrisGuard, use biometric matching in the form of iris scans to verify an individual’s eligibility to receive aid based on their digital identity records. Once verified through IrisGuard, individuals could receive financial aid through their EyePay Phones. 16

AI-adaptive biometrics
Humanitarian organizations must manage digital identification records over long durations of time and often in less-than-ideal conditions. Humanitarians therefore engage in continuous registration operations and verification exercises, activities aimed at ensuring biometric and personal data accuracy within digital identification records. These humanitarian activities can be expensive and are often wrought with hidden costs. 17

The integration of AI-adaptive biometric systems could address the challenge of changes in biometrics by adapting biometric information on record to age-based changes (e.g., aging) and enhancing poor-quality data. 18 AI-adaptive biometric systems work by training models on diverse age ranges or even synthetically aged images, which enable them to predict how aging may alter a person’s face and recognize the same person despite a gap of years. 19 Successful usage could lessen the need for humanitarian organizations to conduct expensive verification or reenrollment exercises.

Handling poor-quality data
AI adaptive biometrics can also help with the handling of poor-quality data, a common occurrence in adverse field conditions. During humanitarian aid operations, low light, outdoor glare, dirt, and dust can impact samples. Biometric matching using poor-quality samples can cause recognition failures, which in the humanitarian sector can result in disruptions accessing much-needed aid. Refugees in the UNHCR Kakuma camp in Kenya have repeatedly voiced frustrations over their biometric data not being consistently recognized, leading to delays in accessing cash aid. 20

Advanced AI enhancement techniques can tolerate such data challenges. Machine learning models like deep neural networks and generative adversarial networks (GANS) have been used to successfully deblur and reconstruct fingerprint images that were initially too unclear to match. 21 These techniques use AI to compare degraded fingerprint samples to their knowledge of fingerprint patterns, then intelligently fill in or sharpen the ridges to create a clearer version for identification. 22 Similarly, AI-driven image restoration can be applied to facial images to remove noise and increase resolution, improving success-matching rates in low-quality datasets. 23

From an operational standpoint, these capabilities can mean fewer false negatives and fewer false matches. In humanitarian contexts where adaptability often translates to inclusivity, AI adaptability techniques can help ensure successful biometric recognition, regardless of factors that may affect data quality. These systems can enhance the reliability and flexibility of identity verification over time—attributes critical to humanitarian operations and the realities of data collection in the field.

Critical considerations

Despite massive improvements in deep learning techniques, facial recognition technology operates through probabilities and never provides definitive results, with accuracy deficits particularly affecting recognition of faces in profile, twins, or individuals wearing face masks. 24 Historical patterns of discrimination have made inroads into the design and implementation of facial recognition technologies from their earliest moments, and biased AI systems may go undetected due to the opacity of machine learning algorithms operating as “black boxes.” In humanitarian contexts where facial recognition technologies are the sole means for identification and identity verification, inaccuracies in such systems may lead to misidentification of individuals with darker skin types, and if identification is a precondition for accessing humanitarian assistance, biased AI could result in exclusion from life-saving aid.

3. Document issuance

The interaction

Following registration and biometric enrollment, individuals are issued identity documentation that demonstrates their identity and refugee or asylum-seeker status.

Verification/Authentication

• Issued documentation links the individual’s biographical information to their biometric data and unique registration number.
• Government-recognized identity documentation or UNHCR-issued identity cards can serve as proof of legal identity, facilitating freedom of movement, access to banking, and ability to register for services.

Documentation issued

• Depending on the context: refugee certificates, identity cards attesting to refugee status, residence cards, or UNHCR-issued documentation with either government logos, joint government-UNHCR logos, or UNHCR logo only.
• Ration cards or entitlement cards for food and assistance distribution, often issued to women in households.

Potential role of AI

Adversarial applications

Spillover effects of fake documentation
In addition to fraud, the potential broader effects of AI-generated documentation are broader systemic degradation. AI-generated fake documentation that entered the system during initial registration would now be encoded into official humanitarian identity documentation, compromising the integrity of the entire identification system. If fraudulent family groupings were created using AI-generated documents showing nonrelated persons as a family, these false relationships would now be memorialized in official documentation. If humanitarians suspect a high number of fake or fraudulent documentation within their digital ID databases, it may be necessary to undertake a costly reenrollment or additional verification exercise to ensure the veracity of the evidence-based foundation on which response plans are built. 25

Complementary applications

Secure documentation design
AI systems can support the production of secure identity documents by incorporating advanced security features and enabling rapid verification of document authenticity through automated systems. AI could be used to design more sophisticated anti-counterfeiting features that are difficult to replicate, while also providing verification tools that can quickly authenticate legitimately issued documents. By integrating AI-driven policies into document management system (DMS) platforms, organizations can proactively enforce security protocols and reduce compliance risks at scale.

Data management and quality assurance
AI technologies can support the management and verification of identity documentation by analyzing data to ensure accuracy and detect inconsistencies in documentation records. AI systems could flag anomalies or inconsistencies between registration data, biometric records, and documentation to be issued, helping identify potential fraud or errors before official documents are produced.

Critical considerations

Regional differences and local law can present complex challenges for biometric systems. Establishing model requirements through guidance from institutions like the National Institute of Standards and Technology (NIST) or alignment with existing humanitarian data standards can help address issues with substandard processes and data. Even though UN agencies are exempt from most legal and regulatory frameworks that govern data protection, humanitarian principles of protection suggest that the use of AI in documentation issuance must maintain data protection and ensure that vulnerable populations maintain control over their personal information. 26

4. Vulnerability and needs assessment

The interaction

During or after registration, individuals with specific protection needs are identified and referred for further assessment by protection specialists. Multi-sector rapid needs assessments, such as the Inter-Agency Standing Committee’s (IASC) cluster structure, the Multi-cluster/sector Initial Rapid Needs Assessment (MIRA) method, and Humanitarian Needs Overview (HNO), are conducted using standardized tools to document the scope of needs and identify priority needs of affected populations.

Verification/Authentication

• Assessment methodologies include household interviews, community-level assessments, and secondary data review to understand conditions and vulnerabilities.
• Cross-checking of information through physical verification, dwelling inspections, and neighbor interviews.
• Assessment teams involve UNHCR, WFP, and other relevant partners, including non-UN NGOs.

Documentation required

• Registration or ration card number to link assessment data to the master recipient database.
• Supporting documentation for specific claims where available (e.g., medical records, birth certificates).
• Signed consent forms for assessment interviews.

Potential role of AI

Adversarial applications

Manipulation of vulnerability assessments
If AI-generated documentation or morphed biometrics have successfully entered the system, fraudulent registrants could present themselves as vulnerable individuals during needs assessment, diverting resources from those genuinely in need. If AI-enabled fraud were to occur at scale, humanitarian organizations could have difficulty accurately identifying and prioritizing the most vulnerable.

Complementary applications

Predictive analytics for needs identification
Machine learning algorithms trained on survey data can recognize patterns of poverty and need in various data sources to prioritize aid to those in greatest need, with research showing this approach can improve targeting performance and increase aggregate social welfare in humanitarian cash transfer programs. AI could be operationalized in needs assessment by integration into within existing structures and processes, such as IASC’s cluster structure, the MIRA method, and HNO. The WFP’s HungerMap uses predictive models to anticipate food insecurity in conflict zones. Traditional data collection in these areas is often risky or near-impossible, but by analyzing trends like market prices and satellite images, HungerMap helps prioritize aid to the regions most at risk. In theory such a tool could help prevent hunger crises from further escalating. 27 Importantly though, one study finds “where AI brings automation, speed, and insights, traditional methods offer contextual understanding, strategic decision-making, and community engagement.” 28

Predictive analytics for anticipatory action
Predictive analytics building on data-driven machine learning can provide essential insights about potential risks to affected populations before a crisis unfolds, supporting anticipatory approaches to humanitarian action. AI systems could identify emerging vulnerabilities—such as food insecurity, protection risks, or health crises—before they become acute, enabling proactive intervention. For instance, UNHCR’s Project Jetson initiative uses predictive analytics to forecast forced displacement in Somalia. The project draws on many data sources to train its ML algorithm, including market prices, river water levels, rainfall patterns, remittance data, and data collected by the UNHCR.

Automated risk scoring
AI enhances assessment processes by analyzing vast amounts of data to detect suspicious connections, assign risk scores, and identify unusual patterns of behavior, allowing for dynamic, tailored approaches with real-time refresh of information. This could help prioritize assessments and allocate specialized resources to the most complex or high-risk cases.

Critical considerations

Algorithmic bias in automated decision-making may have two causes: biased historical training data and improper algorithm design. A key issue is that of bias amplification, where past biased decisions are recognized as norms and applied at a much greater—and crucially—automated scale. Individuals should not be subject to decisions that significant affect them based solely on automated processing without explicit consent, and automated decision-making systems require ongoing audits, diverse development teams, and multidisciplinary oversight including ethicists and social scientists.

The “do no harm” principle necessitates that humanitarian actors consider potential ways their actions or omissions may inadvertently cause harm or create new risks for populations they intend to serve. 29 Humanitarian innovation could potentially introduce unnecessary risks to vulnerable populations when technical failures in newly introduced systems lead to delays, disruption, or cancellation of aid distribution. Any AI system used in vulnerability screening must be carefully designed to avoid excluding genuinely vulnerable people due to algorithmic errors or biases.

5. Service delivery and verification

The interaction

At distribution points or when accessing services, individuals must verify their identity against their registration record to access assistance.

Verification/Authentication

Traditional methods

• Presentation of ration cards or entitlement documents, with card numbers verified against master recipient databases.
• Physical verification through interviews and cross-checking of information used in eligibility determination.

Biometric verification

• Biometric verification at point of distribution using fingerprint or iris scanning to confirm the individual’s claimed identity and ensure only eligible individuals access assistance.
• Iris-enabled point-of-sale devices at distribution locations, such as camp supermarkets or ATMs, allow access without requiring vouchers, cards, or PINs.
• For cash assistance programs, fingerprint verification may be required monthly to keep cards active.

Digital/Token-based systems

• Emerging systems use digital tokens on smartphones or QR code cards that can be verified at distribution points without sharing personal identifying information beyond the initial registration.
• Vendors verify beneficiaries by sending one-time password (OTP) codes to beneficiaries’ mobile phones.

Documentation required

• Physical ration/entitlement card or UNHCR-issued identity documentation.
• For cash-based interventions using systems like CashAssist, registration data from the central database is used to create payment instructions to financial service providers.
• Biometric verification (fingerprint, iris scan) at point of service.
• In digital systems, QR codes or OTP codes rather than physical documents.

Potential role of AI

Adversarial applications

Biometric spoofing at distribution
Deepfake imagery and video can mimic a person’s facial appearance in real time, and false biometrics could introduce fraud during beneficiary authentication at aid distribution points. Individuals could potentially use AI-generated biometric representations to access aid meant for registered beneficiaries, especially in remote or less-supervised distribution contexts.

Supply chain fraud
Typically focus is on delivery points where services are accessed, but humanitarians also utilize digital ID systems within supply chain verification. Given that aid diversion and theft is already a challenge, the entrance of AI-generated deepfakes furnishes well-placed supply chain actors with new tools to facilitate aid redirection and circumvent humanitarian controls and oversight. AI-enabled fraud at this level could involve falsified verification of deliveries, phantom beneficiaries, or manipulation of distribution records.

Complementary applications

Enhanced liveness detection
Liveness detection could reduce the risk of impersonation at in-person aid collection points by thwarting instances of individuals attempting to fraudulently access aid by presenting photos of registered individuals or deepfake videos to biometric authenticators. Combining automated identity verification with biometric technology can block spoofing attempts that use 3D masks, deepfakes, or printed photos, with systems performing real-time checks and assigning dynamic risk scores.

Real-time fraud detection
AI and machine learning are well-suited for fraud detection because of the amount of digital information and ease of analyzing both text and data, with the US Treasury Department using machine learning to prevent or recover over $4 billion in fraud in fiscal year 2024. 30 Deep learning techniques including convolutional neural networks and recurrent neural networks are utilized in fraud detection, with machine learning models able to distinguish between legitimate and fraudulent transactions by recognizing subtle patterns that might be missed by traditional rule-based systems.

AI-enhanced deduplication and network analysis
Biometrics facilitate deduplication by ensuring each biometric data point corresponds to a single registrant. AI-based techniques provide alternative and additive deduplication methods, some reducing the need for biometrics in identification practices. Techniques such as graph analytics offer more intensive data deduplication capabilities. These algorithms map the relationships between individuals, devices, and other identifiers to identify suspicious patterns and connections. 31 In peer-to-peer (P2P) transaction systems, graph algorithms like Louvain community detection have been successfully used to uncover clusters of fraudulent accounts that share common identifiers such as IP addresses and credit cards. 32 To put it simply, community detection algorithms work to identify clusters of entities that interact with each other or share data within a group more often with entities outside the group. 33 In humanitarian contexts, these techniques can spot patterns where multiple identities are associated with the same device or financial account, regardless of biometric uniqueness, thus helping to detect fraud in cash aid operations.

Anomaly detection in distribution
The role these algorithms can play in humanitarian operations extends beyond beneficiary fraud to the broader humanitarian supply chain. Fraud in the supply chain can occur when aid is diverted from intended recipients (e.g., beneficiaries, NGOs, governments, etc.), or when actors gain unauthorized access to resources. Tracking fraud higher on the supply chain can be complex, as humanitarian crisis responses often involve a plethora of entities such as host governments, local organizations, and international actors. 34 The misattribution of donor funds within humanitarian clusters may go unnoticed with traditional methods but could be more easily identified when using advanced algorithms to identify suspicious transactional behavior. Though advanced techniques such as graph analysis and community detection do not require a machine learning element, training an AI model based on the findings of these algorithms can vastly improve fraud detection capabilities. 35

Adaptive verification based on risk
Fraud detection systems using supervised AI learning models offer the benefit of proactive detection, automation, and quantitative performance metrics. 36 Features based on trends extracted via graphing and community algorithms could be used to train AI models to successfully classify fraud risk entities within a humanitarian operation. For instance, if an organization uses algorithms to identify fraud-indicating patterns, these patterns can be introduced to a supervised learning model which could then automate the process on a large scale. The considerations—from digital IDs to NGOs—that are classified as high risk by these models can then be investigated by humanitarian staff.

Critical considerations

Automation bias refers to the tendency to depend excessively on automated systems, which can lead to incorrect automated information overriding correct human decisions. Stress and time pressures on humanitarian staff can exacerbate the issue, as personnel outsource decision-making to automated systems. Distribution staff who are already under pressure may over-rely on AI verification systems even when human judgment suggests a problem, potentially excluding legitimate beneficiaries or allowing fraud to proceed.

When AI systems based on deep learning techniques operate with opacity as “black boxes,” their designers and developers are often unable to understand and sufficiently explain how machines have reached certain decisions, making it more challenging to identify biases in the algorithms and potentially leading to exclusion from assistance for certain groups.

6. Continuous registration and updates

The interaction

Registration data is regularly updated and verified over time through continuous registration activities. Continuous registration can be conducted through both in-person and remote services, such as UNHCR’s Kiosk Automated Services and Information (KASI). Continuous registration enables responsiveness to changing circumstances (e.g., marital status, family composition, address, etc.), and ensures that processes are in place to update individual records and reissue identity documentation, as appropriate. Continuous registration functions to strengthen identity data over time, verifying information and updating data changes throughout the period during which the individual is of concern to UNHCR. 37

Verification/Authentication

• Regular updates to the master recipient database to account for births, deaths, changes in family composition, or lost/stolen documents.
• Reissuance of entitlement documents provides an opportunity to revalidate registration information and assess needs.
  The regular updating and recording of new data elements strengthens identity records and increases confidence in refugee identity documentation.

Documentation required

• Evidence supporting claims for updates (medical records for deaths, birth certificates for new family members)
• Presentation of existing registration documents for verification and updating
• Reenrollment of biometric data if needed

Potential role of AI

Adversarial adaptations

Data integrity threats over time
Morphed biometrics compromise the integrity of digital identification records and distort crucial operational data that humanitarian organizations depend on. If humanitarians suspect high penetration of morphed biometrics within databases, costly reenrollment or additional verification exercises may be necessary to ensure the veracity of evidence-based response plans. The cumulative effect of undetected fraud over time could significantly compromise population data used for needs assessments, resource allocation, and strategic planning.

Evolving fraud techniques
As AI technologies advance, new methods of generating synthetic biometrics and fake documentation are likely to emerge or methods may become more sophisticated, requiring humanitarian organizations to continuously update their detection capabilities else they could face an escalating fraud challenge during registration updates and verification exercises.

Complementary applications

AI-adaptive biometrics for aging
Humanitarian organizations face a major challenge in maintaining integrity within digital ID databases is that people’s biometric traits, namely facial features, change over time. As time passes, children in refugee camps grow into adults and adults age further, these changes to physical appearance warrant periodic updating of biometric records. 38 The average length of displacement for refugees has grown to 17 years, and the numbers of refugees repatriating has declined to the lowest level in over 30 years. 39

Without updating to account for physical change, traditional means of biometric matching may begin to fail, threatening an individual’s access to aid. AI solutions such as age-invariant recognition or “synthetic aging” have been proposed to mitigate the issue. These AI-adaptive biometric systems work by training models on diverse age ranges or even synthetically aged images, which enable them to predict how aging may alter a person’s face and recognize the same person despite a gap of years. 40 Successful usage could lessen the need for humanitarian organizations to conduct expensive verification or reenrollment exercises. Further, the ability to include a buffer for changes in biometrics could reduce false negatives caused either by difficulty in reading biometrics or a lack of up-to-date information, allowing refugees and other vulnerable populations to maintain access to aid through their digital identification records and limiting the chance of access disruptions.

Continuous quality enhancement
Image resolution plays a significant role in shaping accuracy of age estimation models, with low-quality images losing critical features like wrinkles or face shape, though this can be addressed by normalizing resolution through image upsampling methods to improve classification accuracies. AI systems could continuously enhance the quality of biometric records in databases, improving match rates over time without requiring reenrollment.

Perpetual monitoring and real-time updates
Perpetual “Know Your Customer” (KYC) requirements represent an innovative ongoing approach that moves away from traditional periodic reviews in favor of real-time refresh of customer information. 41 New data and insights regularly updated in automated systems and real-time AI monitoring quickly identify and respond to emerging threats. Rather than waiting for scheduled verification exercises, AI could enable continuous monitoring of registration data quality, flagging records requiring updates based on predicted aging patterns, detecting inconsistencies that emerge over time, and identifying potential fraud indicators.

Automated anomaly detection
AI could identify unusual patterns in registration updates—such as multiple claims of deaths in a family unit, inconsistent aging patterns, or documentation presented that doesn’t match expected templates—triggering manual review before changes are finalized in the database.

Quantum-resistant security 42
While blockchain and other encryption technologies are selected for their robustness and security, experts warn that quantum computing poses a serious threat to current cryptographic systems. While they are currently unavailable, McKinsey estimates that 5,000 quantum computers will be operational by 2030. 43 This figure illustrates the need for the humanitarian sector, like all sectors, to begin planning for the security implications of accessible quantum computing. 44

AI technology is emerging as a powerful tool to detect quantum attacks and adapt systems for quantum safety, 45 with machine learning models trained on network and system data able to recognize signs of quantum-powered attacks through network sweeping and flagging unusual bursts in authentication requests or cryptographic failures. 46 As humanitarian organizations increasingly rely on encrypted identity systems for long-term data management, AI-enhanced quantum resistance will become critical.

Iterative system improvement
The iterative nature of audits in AI-based decision-making systems is fitting given their technical nature, responding to the dynamic nature of “learning” systems where the control mechanism is not fixed in time, and helping respond to “black box” issues. AI systems used in continuous registration can themselves be continuously improved, with regular audits identifying biases or errors that can be corrected before they cause systematic problems.

Critical considerations

Data augmentation techniques that alter images with rotations and brightness adjustments can create more refined datasets, with regular data analysis schedules critical for identifying numerous areas of improvement in recognition systems. However, organizations must balance the benefits of AI-enhanced data quality with data protection principles, ensuring that continuous processing of biometric data complies with privacy regulations and informed consent requirements.

AI identity verification is not a silver bullet that alone can solve fraud and privacy problems. Human interaction plays an important role in training systems and solving complex cases that automated systems struggle with, requiring human judgment drawing on social sciences, law, and ethics to develop standards that no optimization algorithm can resolve. The continuous registration process must maintain meaningful human oversight, particularly for consequential decisions about changes to family composition, vulnerability status, or eligibility for assistance.

Cross-cutting considerations: AI in humanitarian identity systems

The integration of AI into humanitarian identification systems presents a fundamental duality. AI offers transformative potential to enhance accuracy, efficiency, and security—from improving biometric matching despite aging or poor environmental conditions to detecting sophisticated fraud patterns at scale. Yet these same capabilities introduce new vulnerabilities through adversarial AI applications and risk amplifying existing biases at unprecedented levels.

Algorithmic bias and equity

Despite massive improvements in deep-learning techniques, federal testing shows that most facial recognition algorithms perform poorly at identifying people besides white men, with law enforcement agencies using automated facial recognition disproportionately arresting Black people due to factors including lack of Black faces in training datasets and tendency of officers’ own biases to magnify these issues. 47

In humanitarian contexts where facial recognition technologies are the sole means of identification and identity verification, inaccuracies in such systems may lead to misidentification of individuals with darker skin types, and if identification is a precondition for accessing humanitarian assistance, biased AI could result in exclusion from life-saving aid.

Transparency and explainability

Facial recognition and AI systems operate as “black boxes,” making it difficult to understand their decision-making processes. Emerging explainability techniques could help identify whether factors considered in decisions reflect bias and could go some way towards reducing the opaque nature of machine-based systems. Some proponents argue that this could enable more accountability than in human decision-making. Article 22 of the GDPR grants data subjects the right to an explanation of automated outcomes, and organizations require Data Protection Impact Assessments where AI processing poses high risk to rights and freedoms.

The “do no harm” imperative

Humanitarian actors must adhere to the “do no harm” principle, which necessitates evaluating how their actions or inaction might unintentionally cause harm or introduce new risks to the populations they aim to assist. Specifically, humanitarian innovation carries the risk of introducing unnecessary dangers to vulnerable groups if technical failures in new systems cause delays, interruptions, or the cancellation of aid distribution. Therefore, any deployment of AI must undergo rigorous harm assessment, focusing particularly on vulnerable populations who are more susceptible to being adversely affected by system failures or inherent biases.

The essential role of human judgment

AI identity verification, while a valuable tool, is not a complete solution for fraud and privacy concerns. Human interaction remains vital for training AI systems and resolving complex cases that automated systems cannot handle, necessitating human judgment informed by social sciences, law, and ethics to establish standards beyond the scope of optimization algorithms. Consequently, the successful deployment of AI in humanitarian settings demands meaningful human oversight, especially for high-stakes decisions that determine access to critical, life-saving assistance.

Ultimately, AI should augment rather than replace human judgment in humanitarian identification systems, serving as a tool to enhance—not diminish—the protection and dignity of displaced and vulnerable populations.