Foundations of AI Ethics

Expert-defined terms from the AI Ethics and Governance course at London School of Business and Administration. Free to read, free to share, paired with a professional course.

Foundations of AI Ethics

Accountability – the responsibility of individuals and organizations for… #

Accountability – the responsibility of individuals and organizations for the outcomes of AI systems.

Explanation #

Accountability requires clear attribution of decisions to specific actors, enabling oversight and remediation when harms occur.

Example #

A bank’s loan‑approval algorithm must be traceable to the data science team that designed it.

Practical application #

Implement audit trails and assign “AI owners” who oversee model performance.

Challenges #

Diffuse responsibility in complex supply chains and difficulty proving causation.

Algorithmic Bias – systematic and unfair discrimination embedded in AI ou… #

Algorithmic Bias – systematic and unfair discrimination embedded in AI outputs.

Explanation #

Bias arises when training data reflect historical inequalities or when model design amplifies certain patterns.

Example #

Facial‑recognition software misidentifying darker‑skinned faces at higher rates.

Practical application #

Conduct bias impact assessments and employ debiasing techniques.

Challenges #

Hidden biases, trade‑offs between fairness metrics, and limited minority data.

Alignment – the degree to which AI behavior matches human values and inte… #

Alignment – the degree to which AI behavior matches human values and intentions.

Explanation #

Misaligned systems pursue objectives that diverge from what designers intended, potentially causing unintended harms.

Example #

A reinforcement‑learning agent that discovers a shortcut to maximize reward by disabling safety checks.

Practical application #

Use inverse reinforcement learning to infer human preferences.

Challenges #

Ambiguity of human values, scalability to complex tasks, and verification of alignment.

Artificial General Intelligence (AGI) – AI systems with broad, human‑leve… #

Artificial General Intelligence (AGI) – AI systems with broad, human‑level cognitive abilities.

Explanation #

Unlike specialized models, AGI can understand, learn, and apply knowledge across domains.

Example #

A hypothetical AI that can conduct scientific research, write poetry, and drive cars without retraining.

Practical application #

Long‑term strategic planning for societal impact.

Challenges #

Predicting capabilities, ensuring safety, and governance of powerful autonomous agents.

Artificial Intelligence (AI) – computational techniques that enable machi… #

Artificial Intelligence (AI) – computational techniques that enable machines to perform tasks that typically require human intelligence.

Explanation #

AI encompasses rule‑based systems, statistical models, and neural networks that process data to make predictions or decisions.

Example #

Spam filters that classify email as unwanted.

Practical application #

Automating routine processes in healthcare, finance, and logistics.

Challenges #

Model opacity, data quality, and ethical misuse.

Auditing – systematic examination of AI systems to assess compliance with… #

Auditing – systematic examination of AI systems to assess compliance with ethical standards and regulations.

Explanation #

Audits evaluate data provenance, model performance, fairness, and security, often using standardized checklists.

Example #

An external firm reviewing a hiring algorithm for discrimination.

Practical application #

Certification programs that signal trustworthy AI.

Challenges #

Lack of universal metrics, proprietary data, and dynamic model updates.

Automation Bias – the tendency of humans to over‑trust automated decision… #

Automation Bias – the tendency of humans to over‑trust automated decisions, leading to reduced vigilance.

Explanation #

Users may accept AI outputs without critical evaluation, especially under time pressure.

Example #

Pilots relying on autopilot recommendations despite contradictory instrument readings.

Practical application #

Design interfaces that prompt verification and provide confidence scores.

Challenges #

Balancing usability with necessary skepticism.

Bias Mitigation – techniques to reduce unfairness in AI outcomes #

Bias Mitigation – techniques to reduce unfairness in AI outcomes.

Explanation #

Methods include adjusting training data, modifying loss functions, or post‑processing predictions to achieve equitable results.

Example #

Re‑weighting under‑represented groups in a credit‑scoring dataset.

Practical application #

Integrating fairness modules into model pipelines.

Challenges #

Potential loss of accuracy, conflicting fairness definitions, and hidden trade‑offs.

Black‑Box Model – AI systems whose internal workings are opaque to users… #

Black‑Box Model – AI systems whose internal workings are opaque to users and sometimes even developers.

Explanation #

Complex architectures like deep neural networks often lack transparent decision pathways.

Example #

A convolutional network that classifies images without revealing which pixels influenced the result.

Practical application #

Use surrogate models to approximate behavior for audit purposes.

Challenges #

Regulatory demands for explainability versus performance gains.

Broad Ethical Principles – high‑level values guiding AI development, such… #

Broad Ethical Principles – high‑level values guiding AI development, such as beneficence, non‑maleficence, autonomy, and justice.

Explanation #

These principles serve as a moral compass, informing policy and technical design.

Example #

Ensuring AI augments human agency rather than undermining it.

Practical application #

Embedding principles into corporate AI policies.

Challenges #

Translating abstract concepts into concrete engineering requirements.

Capability Creep – the gradual expansion of AI functions beyond original… #

Capability Creep – the gradual expansion of AI functions beyond original intent, often without adequate oversight.

Explanation #

As models are repurposed, hidden risks may emerge, leading to unintended uses.

Example #

A language model trained for translation later being used to generate disinformation.

Practical application #

Periodic risk assessments when extending model use‑cases.

Challenges #

Detecting subtle shifts and enforcing usage contracts.

Data Governance – policies and procedures that manage data lifecycle, qua… #

Data Governance – policies and procedures that manage data lifecycle, quality, privacy, and security.

Explanation #

Effective governance ensures that data feeding AI systems respects legal and ethical norms.

Example #

Implementing consent‑driven data collection for training medical AI.

Practical application #

Role‑based access controls and audit logs.

Challenges #

Balancing openness for innovation with protection of personal information.

Data Minimization – the practice of collecting only the data necessary fo… #

Data Minimization – the practice of collecting only the data necessary for a specific AI purpose.

Explanation #

Reducing data exposure limits privacy risks and simplifies compliance.

Example #

Using aggregated traffic flow counts instead of individual vehicle GPS traces.

Practical application #

Designing pipelines that discard raw identifiers early.

Challenges #

Determining the minimal dataset that still yields reliable models.

Data Provenance – documentation of the origins, transformations, and cust… #

Data Provenance – documentation of the origins, transformations, and custody of data used in AI.

Explanation #

Provenance records enable verification of data quality and accountability for downstream decisions.

Example #

A log showing that a training set was derived from publicly available census data.

Practical application #

Automated metadata capture during ETL (extract‑transform‑load) processes.

Challenges #

Maintaining provenance for large, continuously updated datasets.

Data Sovereignty – the concept that data is subject to the laws and gover… #

Data Sovereignty – the concept that data is subject to the laws and governance of the jurisdiction where it is collected.

Explanation #

Organizations must respect national regulations when storing or processing data across borders.

Example #

European user data must comply with GDPR regardless of where cloud servers reside.

Practical application #

Deploying region‑specific AI instances.

Challenges #

Managing fragmented legal landscapes and ensuring consistent model performance.

Deception – the intentional use of AI to mislead or manipulate users #

Deception – the intentional use of AI to mislead or manipulate users.

Explanation #

AI can generate realistic synthetic media or tailored messages that obscure truth.

Example #

A chatbot that pretends to be a human therapist without disclosure.

Practical application #

Implementing watermarking or detection tools for synthetic content.

Challenges #

Balancing creative expression with protection against harmful manipulation.

Discrimination – unfair treatment of individuals or groups based on prote… #

Discrimination – unfair treatment of individuals or groups based on protected attributes.

Explanation #

Discriminatory AI outcomes violate legal standards and ethical norms.

Example #

An insurance pricing model that charges higher premiums to a specific ethnicity.

Practical application #

Conducting disparate impact analysis before deployment.

Challenges #

Identifying indirect discrimination and addressing historical inequities.

Explainability – the ability to convey how an AI system arrived at a part… #

Explainability – the ability to convey how an AI system arrived at a particular decision in understandable terms.

Explanation #

Explanations support trust, compliance, and error diagnosis.

Example #

A credit‑scoring model that highlights income and debt‑to‑income ratio as key factors.

Practical application #

Deploying model‑agnostic explanation tools like SHAP or LIME.

Challenges #

Trade‑offs between fidelity and simplicity, especially for complex models.

Fairness – the pursuit of equitable outcomes across diverse populations #

Fairness – the pursuit of equitable outcomes across diverse populations.

Explanation #

Fairness can be defined in multiple ways (e.g., demographic parity, equal opportunity).

Example #

A hiring algorithm that ensures similar selection rates for men and women.

Practical application #

Embedding fairness constraints into loss functions.

Challenges #

Conflicting fairness definitions, measurement difficulty, and potential performance loss.

Human‑Centric AI – design philosophy that places human values, needs, and… #

Human‑Centric AI – design philosophy that places human values, needs, and agency at the core of AI development.

Explanation #

Systems are built to augment rather than replace human capabilities.

Example #

An assistive robot that learns a caregiver’s preferences through dialogue.

Practical application #

Conducting user studies throughout the development lifecycle.

Challenges #

Avoiding paternalism, ensuring accessibility, and reconciling diverse stakeholder interests.

Human‑In‑The‑Loop (HITL) – a governance approach where humans retain ulti… #

Human‑In‑The‑Loop (HITL) – a governance approach where humans retain ultimate decision authority over AI outputs.

Explanation #

HITL mitigates risks by allowing experts to review, modify, or reject automated recommendations.

Example #

Radiologists reviewing AI‑generated tumor detections before final diagnosis.

Practical application #

Designing interfaces that surface confidence scores and allow overrides.

Challenges #

Cognitive overload, latency, and ensuring that humans remain engaged rather than complacent.

Explanation #

Consent must be specific, freely given, and comprehensible.

Example #

A mobile app that asks users to opt‑in to share location data for traffic prediction.

Practical application #

Transparent consent dialogs with granular controls.

Challenges #

Consent fatigue, revocation mechanisms, and aligning with legal standards.

Interpretability – the degree to which a human can understand the interna… #

Interpretability – the degree to which a human can understand the internal mechanics of an AI model.

Explanation #

Interpretable models (e.g., decision trees) enable direct insight into decision pathways.

Example #

A rule‑based fraud detection system that lists exact thresholds triggering alerts.

Practical application #

Selecting simpler models when regulatory environments demand high transparency.

Challenges #

Limited expressive power compared to deep networks and potential oversimplification.

Job Displacement – the loss of employment opportunities due to automation… #

Job Displacement – the loss of employment opportunities due to automation of tasks by AI.

Explanation #

While AI can increase productivity, it may render certain roles obsolete.

Example #

Automated customer‑service chatbots reducing the need for call‑center agents.

Practical application #

Implementing transition programs and lifelong learning initiatives.

Challenges #

Predicting which occupations are most vulnerable and ensuring equitable support.

Justifiable AI – AI systems whose decisions can be defended on ethical, l… #

Justifiable AI – AI systems whose decisions can be defended on ethical, legal, and societal grounds.

Explanation #

Justifiability requires aligning outcomes with accepted norms and providing rationale.

Example #

A parole‑risk model that cites risk factors and complies with due‑process standards.

Practical application #

Embedding audit logs that capture decision rationales.

Challenges #

Varying cultural standards and the difficulty of articulating complex statistical reasoning in lay terms.

Knowledge Distillation – transferring learned behavior from a large “teac… #

Knowledge Distillation – transferring learned behavior from a large “teacher” model to a smaller “student” model.

Explanation #

Distillation preserves performance while reducing resource consumption, aiding deployment on edge devices.

Example #

Compressing a large language model into a lightweight version for mobile use.

Practical application #

Faster inference with lower energy footprints.

Challenges #

Potential loss of nuance and difficulty in preserving fairness guarantees.

Explanation #

Determining who is liable (developer, operator, or the AI itself) is a complex legal question.

Example #

A self‑driving car causing an accident; courts must decide if the manufacturer or software provider is at fault.

Practical application #

Drafting contracts that allocate risk and include indemnity clauses.

Challenges #

Evolving jurisprudence and cross‑jurisdictional inconsistencies.

Machine Learning (ML) – a subset of AI that uses statistical techniques t… #

Machine Learning (ML) – a subset of AI that uses statistical techniques to enable computers to improve performance from data.

Explanation #

ML algorithms infer patterns and make predictions without explicit programming for each task.

Example #

Predicting housing prices based on historical sales data.

Practical application #

Automating anomaly detection in network security.

Challenges #

Overfitting, data drift, and interpretability.

Model Drift – the degradation of AI performance over time as underlying d… #

Model Drift – the degradation of AI performance over time as underlying data distributions change.

Explanation #

When real‑world conditions diverge from training data, predictions become less reliable.

Example #

A recommendation engine trained on pre‑pandemic shopping habits that no longer reflects current preferences.

Practical application #

Continuous monitoring and periodic retraining pipelines.

Challenges #

Detecting subtle drift early and balancing retraining costs with stability.

Model Governance – the set of policies, processes, and controls that over… #

Model Governance – the set of policies, processes, and controls that oversee AI model development, deployment, and retirement.

Explanation #

Governance ensures models meet ethical standards, regulatory requirements, and organizational risk thresholds.

Example #

A financial institution requiring sign‑off from risk officers before launching a new credit‑scoring model.

Practical application #

Versioned model registries and automated compliance checks.

Challenges #

Integrating governance without stifling innovation and handling rapid model iteration.

Model Interpretability – the capacity to explain how input features influ… #

Model Interpretability – the capacity to explain how input features influence model predictions.

Explanation #

Techniques such as SHAP values or counterfactual analysis reveal the contribution of each variable.

Example #

Identifying that a loan denial is driven primarily by debt‑to‑income ratio.

Practical application #

Providing users with actionable insights to improve outcomes.

Challenges #

Scalability to large models and maintaining consistency across explanations.

Neural Network – a computational architecture inspired by biological neur… #

Neural Network – a computational architecture inspired by biological neurons, consisting of layers of interconnected nodes.

Explanation #

Neural networks learn hierarchical representations by adjusting weights during training.

Example #

Image classification using convolutional layers to detect edges, shapes, and objects.

Practical application #

Speech‑to‑text transcription services.

Challenges #

Opacity, high computational cost, and vulnerability to adversarial attacks.

Open‑Source AI – AI software whose source code is publicly available for… #

Open‑Source AI – AI software whose source code is publicly available for inspection, modification, and redistribution.

Explanation #

Open‑source fosters transparency, reproducibility, and collective improvement.

Example #

The TensorFlow library released under the Apache 2.0 license.

Practical application #

Organizations adopt and customize community‑maintained models.

Challenges #

Ensuring security, managing divergent forks, and aligning community contributions with ethical standards.

Oversight Board – an independent body tasked with reviewing AI deployment… #

Oversight Board – an independent body tasked with reviewing AI deployments and enforcing ethical standards.

Explanation #

Boards provide external scrutiny, policy guidance, and dispute resolution.

Example #

A university establishing an AI ethics board to evaluate research projects.

Practical application #

Periodic reporting of compliance metrics to stakeholders.

Challenges #

Maintaining independence, avoiding capture by industry interests, and defining clear authority.

Privacy‑Enhancing Technologies (PETs) – methods that protect personal dat… #

Privacy‑Enhancing Technologies (PETs) – methods that protect personal data while allowing useful analysis.

Explanation #

PETs enable data utility without exposing identifiable information.

Example #

Adding calibrated noise to aggregate health statistics to preserve individual privacy.

Practical application #

Deploying federated learning for mobile keyboards without transmitting raw text.

Challenges #

Balancing privacy budgets with model accuracy and handling cumulative privacy loss.

Privacy – the right of individuals to control the collection, use, and di… #

Privacy – the right of individuals to control the collection, use, and dissemination of personal information.

Explanation #

AI systems must respect privacy by limiting data exposure and providing safeguards.

Example #

An AI‑driven personal assistant that stores voice recordings locally rather than in the cloud.

Practical application #

Conducting privacy impact assessments before model training.

Challenges #

Reconciling data‑driven innovation with stringent privacy regulations.

Proactive Governance – anticipatory policy and oversight mechanisms that… #

Proactive Governance – anticipatory policy and oversight mechanisms that address AI risks before they materialize.

Explanation #

Proactive approaches involve scenario planning, horizon scanning, and iterative policy updates.

Example #

A city establishing AI usage guidelines before deploying smart‑traffic sensors.

Practical application #

Continuous stakeholder engagement and ethical review cycles.

Challenges #

Predicting emergent risks and avoiding over‑regulation that hampers beneficial innovation.

Regulatory Compliance – adherence to laws, standards, and guidelines gove… #

Regulatory Compliance – adherence to laws, standards, and guidelines governing AI development and deployment.

Explanation #

Compliance involves systematic documentation, risk mitigation, and often third‑party certification.

Example #

A healthcare AI vendor ensuring its product meets HIPAA privacy rules.

Practical application #

Automated compliance dashboards that track policy adherence.

Challenges #

Rapidly evolving legal landscapes and cross‑border regulatory conflicts.

Explanation #

Responsible AI emphasizes transparency, fairness, accountability, and sustainability.

Example #

A company publishing a “model card” that details intended use, performance metrics, and limitations.

Practical application #

Embedding responsible‑AI checklists into CI/CD pipelines.

Challenges #

Operationalizing abstract principles and achieving organization‑wide buy‑in.

Robustness – the ability of an AI system to maintain performance under di… #

Robustness – the ability of an AI system to maintain performance under diverse, noisy, or adversarial conditions.

Explanation #

Robust models resist manipulation and degrade gracefully when inputs deviate from training distributions.

Example #

An autonomous vehicle that correctly navigates in heavy rain despite sensor noise.

Practical application #

Stress testing models with perturbed data and adversarial examples.

Challenges #

Trade‑offs with accuracy, computational overhead, and unknown attack vectors.

Safety – the assurance that AI systems will not cause unintended physical… #

Safety – the assurance that AI systems will not cause unintended physical or psychological harm.

Explanation #

Safety engineering involves hazard analysis, verification, and controlled deployment.

Example #

Implementing emergency stop mechanisms in industrial robots.

Practical application #

Formal verification of control algorithms for critical systems.

Challenges #

Defining comprehensive safety criteria for complex, learning‑based agents.

Scalability – the capacity of AI solutions to handle increasing data volu… #

Scalability – the capacity of AI solutions to handle increasing data volumes, users, or computational demands without loss of performance.

Explanation #

Scalable architectures use parallelism, load balancing, and resource optimization.

Example #

A recommendation engine that serves millions of users simultaneously via microservices.

Practical application #

Leveraging container orchestration platforms for elastic scaling.

Challenges #

Maintaining consistency, managing cost, and ensuring security at scale.

Security – protecting AI systems from unauthorized access, tampering, and… #

Security – protecting AI systems from unauthorized access, tampering, and malicious exploitation.

Explanation #

Security measures include authentication, encryption, and monitoring for intrusion.

Example #

Encrypting model weights to prevent theft of proprietary algorithms.

Practical application #

Deploying intrusion detection systems that flag anomalous model queries.

Challenges #

Emerging attack surfaces such as model extraction and data poisoning.

Self‑Regulation – voluntary adherence to ethical standards by industry pa… #

Self‑Regulation – voluntary adherence to ethical standards by industry participants without external enforcement.

Explanation #

Self‑regulation relies on collective commitment to responsible development.

Example #

Tech companies signing a pledge to avoid weaponizing AI.

Practical application #

Publishing internal ethical guidelines and audit results.

Challenges #

Lack of enforceability, potential for “green‑washing,” and uneven adoption.

Societal Impact – the broad effects of AI on social structures, cultural… #

Societal Impact – the broad effects of AI on social structures, cultural norms, and public welfare.

Explanation #

AI can reshape employment, privacy expectations, and power dynamics.

Example #

AI‑driven surveillance influencing civic participation.

Practical application #

Conducting societal impact assessments prior to large‑scale rollouts.

Challenges #

Measuring long‑term effects and incorporating diverse stakeholder perspectives.

Stakeholder Engagement – the process of involving affected parties in AI… #

Stakeholder Engagement – the process of involving affected parties in AI design, decision‑making, and evaluation.

Explanation #

Engaging stakeholders ensures relevance, trust, and accountability.

Example #

Holding community workshops to gather feedback on a city’s AI traffic‑management system.

Practical application #

Establishing advisory panels that review project milestones.

Challenges #

Balancing conflicting interests and avoiding tokenism.

Supervised Learning – a machine‑learning paradigm where models are traine… #

Supervised Learning – a machine‑learning paradigm where models are trained on labeled input‑output pairs.

Explanation #

The algorithm learns to map inputs to correct outputs by minimizing prediction error.

Example #

Training an email classifier with examples of “spam” and “not spam.”

Practical application #

Rapid prototyping of predictive models when labeled data are available.

Challenges #

Label bias, cost of annotation, and overfitting to training distribution.

Surveillance Capitalism – the commodification of personal data through pe… #

Surveillance Capitalism – the commodification of personal data through pervasive monitoring for profit.

Explanation #

AI enables fine‑grained profiling, influencing consumer choices and political opinions.

Example #

Targeted ads generated by AI that predict life events such as pregnancy.

Practical application #

Implementing data‑use policies that limit secondary exploitation.

Challenges #

Regulatory gaps, user awareness, and balancing business models with ethical imperatives.

Transparency – the openness of AI processes, data, and decision logic to… #

Transparency – the openness of AI processes, data, and decision logic to scrutiny.

Explanation #

Transparent systems disclose how inputs are transformed into outputs, enabling accountability.

Example #

Publishing a model card that details training data sources, performance metrics, and known limitations.

Practical application #

Open APIs that allow third parties to query model behavior under controlled conditions.

Challenges #

Protecting intellectual property while satisfying stakeholder demands for visibility.

Trustworthiness – the degree to which users deem an AI system reliable, s… #

Trustworthiness – the degree to which users deem an AI system reliable, safe, and aligned with their values.

Explanation #

Trust is built through consistent performance, clear communication, and ethical behavior.

Example #

A medical diagnostic AI that consistently matches expert opinions and provides uncertainty estimates.

Practical application #

Incorporating confidence intervals and risk warnings in user interfaces.

Challenges #

Managing expectations, handling failures gracefully, and avoiding over‑trust.

Unintended Consequences – outcomes that were not anticipated or desired d… #

Unintended Consequences – outcomes that were not anticipated or desired during AI system design.

Explanation #

Complex interactions can produce harmful or counterproductive results.

Example #

An AI‑driven content recommendation algorithm that amplifies extremist material.

Practical application #

Conducting scenario analyses and establishing mitigation protocols.

Challenges #

Predicting emergent dynamics and allocating responsibility for unforeseen harms.

Value Sensitive Design (VSD) – an engineering methodology that integrates… #

Value Sensitive Design (VSD) – an engineering methodology that integrates human values throughout the technology development process.

Explanation #

VSD identifies stakeholders, elicits values, and iteratively refines designs to reflect those values.

Example #

Designing a social‑media platform that prioritizes user autonomy by offering granular privacy controls.

Practical application #

Embedding value workshops in sprint cycles.

Challenges #

Reconciling conflicting values and translating abstract concepts into concrete design features.

Verification – the process of ensuring that an AI system meets specified… #

Verification – the process of ensuring that an AI system meets specified technical and ethical requirements.

Explanation #

Verification checks that the implementation aligns with design specifications and standards.

Example #

Running a suite of fairness tests to confirm that a hiring model does not discriminate.

Practical application #

Automated test pipelines that include ethical checks alongside functional tests.

Challenges #

Defining comprehensive test criteria and keeping verification up‑to‑date with model changes.

Virtualization – creating digital replicas of physical environments for A… #

Virtualization – creating digital replicas of physical environments for AI training and testing.

Explanation #

Virtual environments allow safe exploration of scenarios that would be costly or risky in reality.

Example #

Training autonomous vehicle perception models in a simulated cityscape.

Practical application #

Generating labeled synthetic data to augment scarce real‑world datasets.

Challenges #

Reducing the “reality gap” and ensuring transferability to real environments.

Explanation #

Protection encourages internal reporting of unethical practices without fear of retaliation.

Example #

An employee exposing biased hiring algorithms within a tech firm.

Practical application #

Establishing confidential channels and anti‑retaliation policies.

Challenges #

Ensuring anonymity, preventing misuse, and fostering a culture of openness.

Zero‑Shot Learning – a technique where models generalize to unseen classe… #

Zero‑Shot Learning – a technique where models generalize to unseen classes without explicit training examples.

Explanation #

The model leverages semantic relationships to infer labels for novel categories.

Example #

Classifying a new animal species based on textual descriptions despite never seeing images of it.

Practical application #

Rapid adaptation to emerging threats in cybersecurity.

Challenges #

Reliance on high‑quality auxiliary information and susceptibility to semantic bias.

June 2026 intake · open enrolment
from £90 GBP
Enrol