Chapter 1. Understanding the Dark Side: A Balanced Starting Point

Why AI risk is a human, technical, economic, and governance issue

1.1 AI is a capability, not a moral character

Artificial intelligence (AI) refers to a broad family of computational methods that perform tasks associated with perception, prediction, language, pattern recognition, planning, or decision support. Some systems classify images; others forecast demand, recommend content, generate text, or operate within a workflow. These systems differ greatly in design, capability, reliability, and risk. It is therefore misleading to treat “AI” as one machine with one set of properties (Russell & Norvig, 2021).

An AI system does not need malicious intentions to cause harm. A model may produce an incorrect answer, a company may use a score unfairly, a platform may optimise for engagement without considering social consequences, or a criminal may exploit a synthetic voice. Harm can emerge from technical limitations, human decisions, institutional incentives, poor data, weak oversight, or interactions among these factors. Risk management must examine the whole socio-technical system rather than the model alone (NIST, 2023; NIST, 2024).

The phrase “dark side” is useful when it names concrete problems: deception, unfairness, unsafe automation, privacy violations, security vulnerabilities, hidden costs, exclusion, and unaccountable power. It becomes unhelpful when it turns into vague fear or claims that AI is inherently evil. Evidence, context, and proportionate response matter.

1.2 Four sources of AI-related harm

• Technical failure: the system is inaccurate, brittle, poorly calibrated, insecure, or unable to recognise when it does not know.
• Misuse: a person deliberately uses AI for deception, harassment, fraud, manipulation, or other harmful activity.
• Misaligned incentives: an organisation rewards speed, engagement, cost-cutting, or profit while neglecting safety, fairness, or long-term trust.
• Systemic effects: widespread deployment changes labour markets, information ecosystems, access to services, or the distribution of power.

1.3 Risk is not the same as certainty

A risk is a possibility of harm under particular conditions. Its seriousness depends on likelihood, impact, exposure, reversibility, and who bears the consequences. A spelling suggestion has low stakes; an automated decision affecting access to employment, credit, healthcare, or public benefits has much higher stakes. The same model can be acceptable for brainstorming and inappropriate as the sole authority for a consequential decision.

A sensible risk conversation asks: What can go wrong? Who could be harmed? How would we notice? How quickly could we stop it? Who is accountable? What evidence would show that the controls work? NIST’s AI Risk Management Framework organises practical risk work around governing, mapping, measuring, and managing risk, with attention to the system’s lifecycle (NIST, 2023).

1.4 A practical principle

AI should increase human capability without quietly removing human agency. The more consequential, difficult to reverse, or unequal the outcome, the stronger the need for transparency, meaningful human review, appeal routes, monitoring, and documented responsibility. Human review must be real: a person should have time, information, authority, and competence to disagree with the system.