Machine learning creates ethical issues when systems unfairly disadvantage people, misuse personal data, obscure responsibility, or produce decisions that cannot be explained. The four central concerns are bias, privacy, accountability, and transparency.
The Reasoning
A model learns patterns from training data. Ethical harm can originate in data selection, model design, deployment, or the use of outputs.
| Ethical issue | Mechanism and example |
|---|---|
| Bias | Unrepresentative or historically biased data can create systematically unfair outcomes. A recruitment model trained on previous appointments may discriminate against underrepresented groups because past decisions contained human prejudice. |
| Privacy | Models may use medical records, locations, or online activity. Collection without informed consent, secondary use, or insecure storage can violate data privacy. Data minimization, access controls, and anonymization reduce risk. |
| Accountability | Responsibility can be unclear when an automated decision causes harm. Organizations need named responsibility, human oversight, audits, and an appeals process. |
| Transparency | A black box model may not reveal how inputs produced an output. Limited explainability makes errors and discrimination harder to detect or challenge. |
These issues interact: an opaque model can conceal bias, while weak accountability can allow privacy violations to continue. Ethical design therefore requires representative data, fairness testing, secure handling, explainable outputs, and meaningful human review.
A common misconception is that machine learning is automatically objective because algorithms are mathematical. Models reflect human choices and patterns in their training data, so they can reproduce or amplify existing inequalities.
IB Exam Technique
For A4.4 Ethical considerations, do more than identify an issue. Explain its mechanism, identify affected stakeholders, state a specific consequence, and propose a realistic mitigation. For an evaluate question, balance benefits against the severity, likelihood, and manageability of risks, and reach a supported conclusion. Examiners reward applied analysis, not a generic list of terms.
