IA · 22 June 2026 · 4 min read
Beyond the limits of computer vision: the UK's plan for AI facial age estimation reveals deep algorithmic bias
In brief: Starting next year, the UK government plans to deploy facial age estimation (FAE) systems to assess the age of undocumented asylum seekers. However, a leaked internal report detailing tests of seven algorithms reveals significant margins of error and severe algorithmic bias, particularly against individuals from Sub-Saharan Africa, raising serious questions about the use of AI in high-stakes human decisions.
by Team Mocchi's
AI at the border: the debut of facial age estimation
Beginning next year, border authorities in the United Kingdom will implement an artificial intelligence system based on facial age estimation to identify undocumented asylum seekers. This marks the world's first systematic deployment of this technology in high-stakes border control and national security environments, where algorithmic decisions have immediate and life-altering impacts on legal and human rights.
Facial age estimation analyzes facial geometry and features to calculate a probabilistic estimate of an individual's biological age. While the government intends to use this tool to streamline identification processes, the accuracy and neutrality of these systems remain highly controversial within the scientific and software development communities.
The weight of errors in high-stakes scenarios
In the context of asylum policy, distinguishing between a minor and an adult is far more than a bureaucratic formality. Minors are entitled to strict international and national legal protections, including access to education, placement in protected facilities, and a ban on being detained with adults.
If an algorithm errs by estimating a minor's age above the legal threshold of adulthood, the individual risks losing these protections instantly, potentially being placed in standard adult detention centers. In such high-stakes scenarios, the tolerance margin for software error must be virtually zero. Yet, internal testing conducted by the government itself paints a very different picture.
Ethnic and gender bias: findings from the internal report
A classified document from the UK Home Office, recently leaked and based on tests of seven different age estimation algorithms, outlines the concrete limitations of the technology. The findings reveal that even the best-performing systems suffer from severe algorithmic bias.
The primary issue is ethnic bias. The accuracy of the algorithms drops sharply when analyzing individuals from Sub-Saharan Africa, who historically represent one of the largest demographics of arriving migrants. For female Sub-Saharan Africans, the average difference between real age and the AI-calculated estimate was significantly higher than for other demographic groups. This statistical asymmetry dramatically increases the risk of a minor being incorrectly classified as an adult.
The technical root: training dataset limitations
From a software development perspective, the failure of these systems is predictable. Computer vision models are typically trained on datasets predominantly composed of Caucasian faces or, more broadly, images that do not accurately represent the demographic diversity of global populations.
When an algorithm trained on imbalanced data is deployed in real-world scenarios involving highly diverse populations, its performance degrades rapidly. This is not just a general lack of precision, but a systematic bias that penalizes specific groups of people. For AI system designers, this case serves as a crucial warning about the absolute necessity of gathering, validating, and balancing training data before deploying models to production.
From screen to physical borders: future outlook
Until now, facial age verification systems have been discussed and implemented primarily in the digital sphere—for instance, to restrict minors' access to social media or to comply with online privacy laws across various global markets. The transition of this technology from online platforms to physical state infrastructure marks a significant turning point.
Deploying automated solutions without rigorous accuracy standards or transparent appeal mechanisms threatens to undermine trust in AI technologies. For software and AI development agencies, the UK case underscores the need to redefine the boundaries of technological accountability. Automation cannot replace human judgment in complex decision-making processes, especially when a software error directly translates into a violation of fundamental human rights.