This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
A facial recognition system integrated into the Thai government's 'Pao Tang' wallet application caused significant accessibility issues for citizens attempting to register for financial aid programs. Users, particularly the elderly, reported frequent failures in identity verification, forcing them to physically visit bank branches or ATMs to resolve the issue. The system's brittleness regarding appearance changes, such as wearing glasses or tying hair, contributed to the exclusion of eligible participants.
A Thai wallet app failed to recognize people’s faces, resulting in citizens and disproportionately elders unable to sign up for Thai government’s cash handout and co-pay programs or having to wait in long queues at local ATMs for authentication.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
Accuracy and effectiveness of AI decisions and actions are dependent on group membership, where decisions in AI system design and biased training data lead to unequal outcomes, reduced benefits, increased effort, and alienation of users.
Also classified as
The AI system failed to perform reliably under varying real-world conditions, showing extreme sensitivity to lighting, hairstyles, and glasses.
The facial recognition system exhibited unequal performance, disproportionately failing for elderly citizens and individuals with minor appearance variations, thereby restricting their access to public benefits.
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The accessibility issues and verification failures were caused by the lack of robustness in the deployed facial recognition AI system.
Added categories peak at 2: Minor (Psychological Harm, Indirect)
A domestic deployment of a brittle facial recognition system for Thai financial aid programs caused minor disruptions to public service delivery and unequal access for elderly citizens. It represents a low-level national security concern primarily affecting domestic administrative functions and societal trust in automated government systems.
The facial recognition system operates as a deterministic tool that takes direct biometric inputs (live face scans and ID photos) and outputs a verification decision without maintaining an independent state or planning.
Risk Level 2. High Risk: The system utilizes biometric identification and is deployed by public authorities to determine eligibility and manage access to essential public services and financial aid.
The system's primary goal is to perform facial recognition to verify the identity of users registering for government financial aid programs.
The reports provide consistent, detailed accounts of the facial recognition failures across multiple government campaigns, supported by social media complaints, news coverage, and official bank instructions.
The issues could be interpreted as user error or poor lighting rather than systemic AI failure, though bank-level failures suggest a core lack of system robustness.
Technology
Strict Facial Matching Algorithm
Algorithm failed to match faces with minor changes like hair or glasses.
Unstable App Updates
App updates reset verification status, forcing users to scan again.
Poor Camera Integration
App struggled to capture clear images under variable lighting.
Data Inputs
Worn or Faded ID Cards
Physical damage or fading on ID cards prevented successful scanning.
Glare and Reflections on Card
Light reflections on the ID card surface blocked verification.
Outdated ID Photos
Older photos in government database did not match current faces.
Human Factors
Varying User Hairstyles
Users wearing glasses or changing hair failed the match.
Suboptimal Scanning Environment
Users scanned in dark rooms or against busy backgrounds.
Incorrect Scanning Posture
Users did not hold still or align their face within the frame.
Process and Methods
Inflexible Matching Criteria
The system lacked tolerance for normal variations in user appearance.
Inefficient Fallback Options
Failing users were forced to visit physical bank branches.
Management
Inadequate Pre-release Testing
App updates were deployed without resolving facial scanning bugs.
Poor Support Infrastructure
Bank staff and ATMs were overwhelmed by verification failures.
Regulatory Environment
Strict Anti-Fraud Policies
Government rules mandated high-security biometric checks.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
