1.3 Responsible AI principles

Where responsible-AI principles come from

Over the past several years a striking consensus has emerged across dozens of published AI ethics frameworks — from the OECD AI Principles and the UNESCO Recommendation to the NIST AI RMF's "trustworthy AI" characteristics and countless corporate codes. Although the wording varies, the same core principles recur. The AIGP expects you to know these principles, what each means operationally (how it shows up as a real control or practice), and — crucially — the tensions between them, because responsible AI is rarely about a single virtue in isolation. Do not memorize any one organization's exact list; instead learn the underlying ideas, because different frameworks group and name them differently while pointing at the same handful of concerns. The principles below are the ones that appear most consistently across the major sources the exam draws on.

The core principles

Fairness and non-discrimination

AI should treat people equitably and avoid unjustified, harmful bias against individuals or groups — especially on protected characteristics such as race, sex, age, or disability. Operationally this means testing training data and outputs for disparate impact, choosing appropriate fairness metrics, and remediating skew. A key subtlety: there are multiple, mathematically incompatible definitions of fairness (e.g., equal false-positive rates versus equal predictive value), so "fair" requires an explicit, context-specific choice.

Transparency and explainability

People affected by AI should be able to know that AI is being used and to obtain a meaningful explanation of how it works or why a particular decision was made. Transparency is broad disclosure (documentation, notices, model cards); explainability is the narrower technical ability to describe the factors behind an output. Operational practices include disclosure notices, model documentation, and interpretability techniques. Deep learning's opacity makes this hard.

Accountability

Someone must be answerable for an AI system's outcomes; responsibility cannot evaporate into "the algorithm did it." Operationally, accountability means assigning clear ownership, keeping documentation and audit trails, conducting impact assessments, and enabling redress. It underpins governance structures (roles, sign-offs, records) and is the principle most directly tied to legal liability.

Safety and robustness

AI should perform reliably, remain secure against manipulation, and avoid causing physical, psychological, or economic harm — including under unexpected or adversarial conditions. Robustness means the system degrades gracefully and resists attacks such as data poisoning or adversarial examples. Operational practices include rigorous testing, red-teaming, monitoring for drift, and fallback mechanisms.

Privacy and data governance

Because AI is data-hungry, it must respect privacy and sound data governance: a lawful basis for data, data minimization, security, and controls over how personal data is used in training and inference. This principle links AI governance back to established privacy law (such as the GDPR) and to security controls protecting both the data and the model.

Human oversight and human-centered values

Humans should remain meaningfully in control of consequential AI, and systems should serve human autonomy and well-being. This is often operationalized along a spectrum — human-in-the-loop (a person approves each decision), human-on-the-loop (a person monitors and can intervene), and human-in-command (humans set overall bounds). The EU AI Act mandates human oversight for high-risk systems.

Contestability and redress

Individuals affected by an AI-driven decision should be able to challenge it and seek a remedy. Operationally this means providing an appeal or review channel, a human re-examination path, and correction of errors — closely tied to transparency (you cannot contest what you cannot understand) and accountability (someone must receive and act on the challenge).

A compact way to remember the set:

The tensions between principles

The reason governance is hard — and the reason the AIGP tests judgment rather than recitation — is that these principles frequently pull against each other:

• Accuracy vs. explainability. The highest-performing deep models are the least interpretable; demanding full explainability may force a simpler, less accurate model.
• Fairness vs. accuracy (and fairness vs. fairness). Correcting disparate impact can lower overall accuracy, and satisfying one fairness metric can mathematically violate another.
• Transparency vs. privacy and security. Disclosing how a model works, or publishing training data, can leak personal data or expose the system to gaming and attack.
• Personalization vs. privacy. More useful, tailored AI usually requires collecting more personal data.
• Innovation/speed vs. oversight. Heavy human-in-the-loop review adds friction and cost the business may resist.

There is no universal formula for resolving these trade-offs; responsible AI requires context-specific balancing, documented reasoning, and stakeholder input. On the exam, a scenario that seems to make one principle absolute usually has a "best" answer that acknowledges the trade-off and reaches a proportionate, well-documented decision. Recognizing that principles coexist in tension — and that governance's job is to balance them deliberately rather than maximize any single one — is the central insight of this domain.