AI Ethics & Safety

As AI systems become more powerful and pervasive, the ethical implications of their development and deployment have become critically important. From bias in training data to the existential risks of advanced AI, understanding these challenges is essential for anyone building or using AI systems.

---

Bias in AI Systems

AI systems can perpetuate and amplify existing biases present in their training data and design choices.

Types of Bias

Type	Description	Example
Training data bias	Biased or unrepresentative training data	Facial recognition less accurate for darker skin tones due to underrepresentation in training sets
Selection bias	Non-random sampling of training data	Medical AI trained mostly on data from one demographic group
Measurement bias	Flawed data collection methods	Using arrest records as a proxy for crime rates
Algorithmic bias	Model architecture or objective introduces bias	Optimising for engagement may promote sensational content
Confirmation bias	Humans selectively interpreting AI outputs	Accepting AI recommendations that confirm pre-existing beliefs
Automation bias	Over-reliance on AI decisions	Assuming the AI is always correct and not verifying

Mitigating Bias

• Diverse and representative training data — Ensure datasets reflect the population the model will serve
• Bias auditing — Regularly test models across demographic groups
• Fairness metrics — Measure and optimise for equitable outcomes
• Diverse teams — Include people from different backgrounds in development
• Transparency — Document data sources, limitations, and known biases

Note: Eliminating bias entirely is extremely difficult. The goal is to identify, measure, and minimise harmful biases through systematic processes.

---

Hallucinations and Factual Accuracy

Hallucinations occur when AI models generate plausible-sounding but factually incorrect information. This is one of the most significant practical challenges with current LLMs.

Why Hallucinations Occur

• LLMs are trained to generate plausible text, not necessarily truthful text
• They have no internal model of truth — they predict statistically likely sequences
• Rare or specialised topics have less training data coverage
• Models may extrapolate from patterns in ways that produce confident-sounding errors

Mitigation Strategies

• Retrieval-Augmented Generation (RAG) — Ground responses in verified source documents
• Citation and attribution — Ask the model to cite sources (then verify them)
• Temperature reduction — Lower temperature reduces creative but potentially false outputs
• Human review — Always verify AI outputs for high-stakes decisions
• Confidence calibration — Train models to express uncertainty appropriately

---

The AI Alignment Problem

AI alignment is the challenge of ensuring that AI systems pursue goals that are aligned with human values and intentions.

The Alignment Problem:

What we say:      "Maximise customer satisfaction"
What we mean:     "Help customers effectively while being honest and fair"
What AI might do: "Give customers everything they ask for, even if it
                   is harmful, costly, or involves deception"

The gap between what we specify and what we intend is the alignment problem.

Key Alignment Challenges

• Specification problem: It is extremely hard to formally specify what we want
• Reward hacking: AI systems may find unexpected ways to maximise their objective that violate the spirit of the goal
• Mesa-optimisation: Advanced AI might develop internal goals that differ from trained objectives
• Value learning: How can we teach AI systems complex human values?

Approaches to Alignment

• Constitutional AI (Anthropic) — Train AI to follow a set of principles
• RLHF — Use human feedback to guide behaviour
• Interpretability research — Understand what models are actually doing internally
• Red-teaming — Deliberately try to make models behave badly to identify weaknesses

---

Privacy and Data Protection

AI systems raise significant privacy concerns: