The canonical principles, one by one
Lesson 2 of 5 in Core Ethics Principles and Trustworthy AI.
Distil the 84-plus documents and seven principles remain standing. Learn these seven properly — what each demands, how each fails, and what operationalizing each looks like — and you can decode any ethics charter, any trustworthy-AI framework, and most statutes on sight, because they are all remixes of this list.
One habit to build now: for every principle, ask "what would violating this look like, and who would catch it?" A principle you cannot imagine being violated — or being caught — is decoration, not governance.
Fairness — comparable people get comparable treatment
The demand: AI systems should not create or reinforce unjust discrimination — across groups (group fairness) or between similar individuals (individual fairness).
How it fails: biased training labels (Amazon’s recruiting tool), proxy variables (health cost as a proxy for health need in the Obermeyer study, which halved Black patients’ access to care programs), and unequal error rates (Gender Shades). Note the two harm shapes from the harms module: allocative (resources withheld) and representational (groups demeaned or stereotyped).
Operationalized as: disaggregated performance testing, bias audits (NYC Local Law 144 made them mandatory for hiring tools), fairness metrics — chosen deliberately, because lesson three proves they cannot all hold at once.
Transparency — no hidden AI, no hidden logic
The demand: people should know that AI is being used (existence disclosure), what it can and cannot do (capability disclosure), what data and design choices shaped it (documentation transparency), and when content is synthetic (labeling).
How it fails: silent deployment — the Dutch childcare-benefits families never knew an algorithm had flagged them; candidates screened out by tools they were never told existed.
Operationalized as: EU AI Act Article 50 disclosure duties, model and system cards, public registers of high-risk systems, AI-content labels. Distinguish it from explainability: transparency is about making information available; explainability is about making a specific decision intelligible. A fully documented system can still be unexplainable, and vice versa.
Explainability — decisions you can interrogate
The demand: an affected person, an auditor, and a regulator should each be able to get a useful-to-them account of why the system produced this output. Explanation is audience-relative: the data scientist wants feature attributions; the loan applicant wants "what would have changed the outcome?"
How it fails: the black-box problem — deep models don’t carry human-readable reasons. Post-hoc techniques approximate them: LIME and SHAP estimate which inputs mattered; counterfactual explanations state the smallest change that flips the decision ("€5,000 more annual income and you would have been approved").
Operationalized as: explainability requirements for high-risk systems (EU AI Act Arts 13, 86), the long-running GDPR "right to explanation" debate over Article 22, and — crucially — matching the explanation form to the audience, not the developer’s convenience.
Accountability — a human answers for the system
The demand: for every AI-influenced outcome, some identifiable person or organization is answerable — able to explain, correct, compensate, and be sanctioned. Four load-bearing components: answerability (someone must respond), auditability (records exist to reconstruct what happened), traceability (logs link outcome to model version, data, and configuration), and redress (affected people can contest and get remedy).
How it fails: the "many hands" problem — provider blames deployer, deployer blames vendor, everyone blames the model. Moffatt v. Air Canada (2024) is the antidote in miniature: the airline argued its chatbot was "a separate legal entity responsible for its own actions"; the tribunal held the airline liable for the refund policy its bot invented.
Operationalized as: role-based duty chains (provider/deployer), logging obligations, liability regimes — fault-based (prove negligence) versus strict (harm suffices) — and complaint mechanisms.
Privacy and data governance — data used lawfully, minimally, and under control
The demand: AI must respect the people inside its data: collect no more than needed (data minimization), use it only for stated purposes (purpose limitation), with valid legal basis and meaningful consent where required.
How it fails: scraping faces without consent (Clearview AI — fined and banned across multiple EU states), models memorizing and regurgitating personal data, systems inferring sensitive traits (pregnancy, orientation, health) from innocuous signals.
Operationalized as: GDPR obligations and DPIAs, plus privacy-enhancing technologies worth knowing conceptually: differential privacy (calibrated noise so no individual’s presence is detectable), federated learning (train where the data lives; share model updates, not data), synthetic data (train on generated stand-ins). Each helps; none is magic — synthetic data can leak its source distribution, and federated updates can be attacked. PETs shift risk; governance still has to manage what remains.
Safety, robustness, and security — works as intended, resists attack, fails gracefully
The demand: three related properties. Safety: the system does not endanger life, health, property, or rights in normal use. Robustness: it keeps working under messy, shifting, out-of-distribution conditions — accuracy on the test set is not accuracy on the world. Security: it resists deliberate attack — data poisoning, adversarial examples, prompt injection, model theft.
How it fails: hallucinated legal citations filed in court (Mata v. Avianca), medical advice degrading on populations unlike the training data, chatbots jailbroken into revealing dangerous instructions.
Operationalized as: pre-deployment testing and red-teaming, fail-safe design (degrade to human review, never to silent failure), post-market monitoring, incident reporting. The engineering habit to demand: graceful degradation — a trustworthy system knows when it doesn’t know.
Human oversight — humans retain meaningful control
The demand: AI supports human decision-making without displacing human responsibility. The standard taxonomy: human-in-the-loop (a person approves each decision), human-on-the-loop (the system acts; a person monitors and can intervene), human-out-of-the-loop (no real-time human role — acceptable only where speed makes review impossible and other controls compensate).
How it fails: nominal oversight — the rubber stamp. Automation bias, alert fatigue, and throughput pressure hollow out review until the human is a liability shield, not a control. You met this in the autonomy spectrum; it is the most common gap between paper compliance and reality.
Operationalized as: EU AI Act Article 14 (oversight by people with the competence, authority, and time to intervene — including the ability not to use the system), oversight design patterns (blind second reads, measured disagreement rates, workload caps), and the concept of meaningful human control from the autonomous-weapons debate.
Key terms: explainability, transparency, human oversight, automation bias, black box
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.