A4.4.1 — Ethical Implications of Machine Learning
Core Ethical Issues in ML
An ethical implication is a potential positive or negative consequence that a decision or action may have on well-being, justice, fairness, rights, and freedom.
| Issue | |
|---|---|
| Accountability | Who is liable when AI causes harm? |
| Algorithmic fairness | Biased training data leads to discriminatory outcomes. |
| Bias | AI inherits and amplifies biases from training data. |
| Consent | Personal data must be collected with informed consent. Data should be anonymised. |
| Environmental impact | Training AI consumes vast energy. Estimated 85+ terawatts per hour globally. |
| Privacy | Fundamental legal right: UDHR Art.12, ECHR Art.8, EU Charter Art.7. |
| Security | AI systems are targets for cyberattacks and manipulation. |
| Societal impact | AI displaces jobs and alters society. Benefits must be equitable. |
| Transparency | AI decisions must be explainable. "Black box" models erode trust. |
Bias in Training Data
- AI replicates and amplifies societal prejudices embedded in historical data.
- Underrepresented groups get worse outcomes in healthcare, criminal justice, and lending.
- Bias is often subtle, even "neutral" data reflects the perspective of whoever collected it.
- Amazon: Recruitment AI discriminated against women. Shut down.
- COMPAS (USA): Criminal sentencing AI had higher false-positive rate for Black defendants.
- Credit scoring: Historical financial data led to discrimination against minorities and low-income groups.
Ethical Concerns in Online Communication
- Misinformation: False info spread unintentionally.
- Disinformation: False info spread deliberately to deceive or manipulate.
- Deep fakes: ML-generated synthetic media making people appear to say or do things they never did. Used for fraud and reputational damage.
- Harassment bots: ML can automate abuse campaigns at scale. Detection tools may also wrongly silence legitimate voices.
- Anonymity: Enables free speech but also cyberbullying and trolling.
- Privacy: Platforms track and monetise user behaviour, often without meaningful consent.
- Bias comes from training data, not the algorithm itself. The algorithm learns what the data teaches it.
- Privacy is a legally protected fundamental right, not just a preference. Mention the UDHR or GDPR for context.