Emerging Technologies and Societal Impact
Scope
Emerging-technology items ask you to apply existing principles to systems whose risks are novel in degree or in kind, and to resist the claim that novelty suspends obligation.
Generative AI and foundation models
- Hallucination makes outputs unreliable in high-stakes advice; grounding, retrieval and evaluation sets are mitigations, not guarantees.
- Training-data provenance raises copyright, licensing and consent questions, and memorisation means models can emit personal data seen in training.
- Prompt injection and data exfiltration are security risks when a model acts on untrusted content with tool access.
- General-purpose AI obligations in the EU AI Act (documentation, copyright policy, systemic-risk evaluation and incident reporting) attach to model providers independently of downstream uses.
- Synthetic data is not automatically anonymous; generative models can leak training records, and a fidelity-fairness trade-off arises when synthetic data under-represents minorities.
Other frontiers
- Biometrics: facial recognition, emotion inference, gait and voice. The EU AI Act prohibits several uses outright; BIPA and state laws require written consent and retention limits.
- Privacy-enhancing technologies: federated learning, secure multiparty computation, homomorphic encryption and differential privacy, each with a privacy-utility budget that must be stated, not assumed.
- Ambient and IoT data: passive collection from homes, vehicles, wearables and workplaces blurs consent and erodes bystander rights.
- Neurotechnology and affective computing: inferences about inner states raise mental-privacy concerns beyond data protection.
- Quantum: harvest-now-decrypt-later risk drives post-quantum migration planning for long-lived sensitive data.
- Environmental and labour impacts: training and inference costs, water use, content-moderation labour and displacement effects are ethical issues, not externalities.
Worked example
A public benefits agency wants an LLM to triage callers and draft eligibility notes. Risks include hallucinated entitlements, disparate accuracy across dialects and disabilities, memorisation of caller details, and automation bias among caseworkers. A defensible design keeps the decision with a human, restricts the model to summarisation, logs inputs and outputs, tests on disaggregated caller groups, forbids tool access to case records, publishes an assessment, and provides an appeal route. The tool's novelty does not change the agency's existing legal duties.
Traps
Treating AI as a legal category that overrides GDPR or consumer law; assuming synthetic data is safe; assuming open weights equal ethical openness; accepting vendor accuracy claims without disaggregated evidence; ignoring energy and labour costs in impact assessments.