A4.4 Ethical considerations

The model was trained on data that does not represent all patients — it skews heavily toward a specific demographic (male, over 50). This means the training data is unrepresentative.

Algorithmic bias occurs when a model learns patterns from skewed or unrepresentative training data, causing it to perform unequally across different groups. A model trained mainly on older male data may produce worse recommendations for women or younger patients.

This directly causes discriminatory healthcare outcomes — patients underrepresented in the training data receive less accurate or appropriate treatment recommendations. This is both an ethical fairness issue and a patient safety issue.

The scenario most directly illustrates algorithmic bias leading to unequal treatment outcomes across demographic groups, making option A the correct choice.

The question asks which ethical concern is most directly illustrated by training a healthcare ML model predominantly on data from one demographic group.

Data poisoning involves a malicious actor deliberately corrupting training data. The scenario describes a bias arising from naturally skewed data collection — not a deliberate attack. This option is eliminated.

Informed consent refers to individuals agreeing to how their personal data is used. Developers do not give consent — patients do. This option misapplies the concept and is eliminated.

While energy consumption is a real ethical concern in ML, the scenario says nothing about computational resources or environmental impact. This is not relevant here and is eliminated.

The remaining option — algorithmic bias leading to unequal treatment outcomes across demographic groups — directly matches the scenario of skewed training data producing unfair outcomes for underrepresented patients.

Deploying facial recognition in public spaces affects all citizens who pass through those areas, whether or not they are suspects. This raises questions about privacy and surveillance.

Privacy is recognized as a fundamental human right (e.g., UN Universal Declaration, Article 12). Citizens in public transport stations have not given informed consent to have their biometric data captured and processed by an ML system.

Continuous, automated biometric surveillance of an entire population — even in public — is qualitatively different from traditional policing. It creates a chilling effect on freedom of movement and expression, even for innocent people.

The most significant and directly relevant ethical concern is that the system erodes privacy rights and enables mass surveillance without individual consent — matching option A.

The question asks for the key ethical concern about deploying ML facial recognition cameras in all public transport stations.

While energy consumption is an ML ethics topic, running cameras on public transport does not represent a notably large or unique energy cost. This is not the primary ethical concern here and is eliminated.

Facial recognition systems are commonly trained using supervised learning. This option is factually incorrect and is eliminated.

The scenario describes a tool to assist law enforcement, not replace it entirely. This is an exaggerated and unsupported claim that does not represent the stated deployment context.

The correct answer is that mass surveillance without consent erodes a fundamental right to privacy — the central ethical issue in deploying facial recognition technology at scale in public spaces.