Chapter 4 — Responsible Ai And Ethics

What is Responsible AI?
Responsible AI means designing, building and using artificial intelligence systems in ways that are fair, safe, transparent, private and accountable. It ensures AI benefits people and society while reducing harm.

Why it matters (Class 10 perspective)
AI systems make decisions (like recommending content, approving loans, or diagnosing illness). If these systems are biased, opaque or unsafe, they can cause unfair outcomes or harm. Responsible AI helps avoid these risks and builds trust.

Key principles

• Fairness: Treat similar people similarly. Avoid biased outcomes across groups (gender, caste, region etc.).
• Transparency / Explainability: Users should understand how and why a decision was made.
• Privacy: Protect personal data used to train or run AI systems.
• Accountability: Humans must be responsible for AI decisions and be able to correct mistakes.
• Safety & Robustness: Systems should behave reliably even when conditions change or inputs are noisy.
• Human oversight: Keep humans ‘in the loop’ for important decisions.

How to build Responsible AI (basic steps)

1. Careful data collection: Check for missing groups and labeling errors.
2. Detect bias: Use metrics and visual checks to find unequal outcomes.
3. Mitigate bias: Rebalance data, change features, or apply fairness-aware algorithms.
4. Test and validate: Evaluate performance and fairness on separate test sets and for different groups.
5. Explain decisions: Provide simple explanations or reasons for outcomes to users.
6. Monitor in production: Track model behavior over time and update when needed.
7. Document and govern: Maintain records (data sources, model versions, tests) and clear responsibility chains.

Simple classroom analogy: Imagine a teacher grading exams. Responsible AI is like making sure the teacher uses the same answer key for everyone (fairness), explains grading rules (transparency), doesn’t share students’ private marks publicly (privacy), and can correct mistakes if asked (accountability).

Short summary: Responsible AI is about creating AI that is fair, explainable, safe, private and accountable. For students, it means knowing that AI decisions should be checked, explained, and improved to avoid harm.

Principles of AI Ethics

AI ethics are rules and guidelines that help designers, developers, users and organisations create and use artificial intelligence in ways that are safe, fair and trustworthy. For Class 10, the key principles are: fairness, transparency (explainability), privacy, accountability, safety and reliability, human oversight, and non‑maleficence (do no harm).

• Fairness: AI systems should treat people equally and not discriminate based on race, gender, age, caste or other protected attributes. Fairness means checking that outcomes (decisions, predictions) do not unfairly favour or harm particular groups.
• Transparency / Explainability: The behaviour of an AI system should be understandable. Users and affected people should be able to know why a decision was made or at least get a clear explanation of the main reasons.
• Privacy: Personal data used by AI should be collected and used with consent, minimised to what is necessary, and protected against misuse. Techniques like anonymisation and differential privacy help protect individuals.
• Accountability: People or organisations must be responsible for the outcomes of AI systems. If an AI system causes harm, there should be ways to find who is responsible and how to remedy the harm.
• Safety and Reliability: AI must perform correctly and consistently in the environments it is used. Systems should be tested for faults, and there must be plans for failure modes (what happens when the system fails).
• Human Oversight: Humans should be able to review, correct or stop decisions made by AI, especially in critical areas such as healthcare or justice.
• Non‑maleficence and Social Good: AI should not be used to cause harm and should aim to benefit people and society — for example improving healthcare, education and safety.

How these principles are applied (simple steps):

1. Define the purpose and limits of the AI system (what it should and should not do).
2. Collect and check data for bias; remove or correct biased data where possible.
3. Choose models that balance accuracy with explainability (e.g., simple models for explainability, complex models only when needed).
4. Test the system across different groups to measure fairness and reliability.
5. Provide clear explanations to users and keep logs for accountability.
6. Protect personal data and use privacy techniques when required.
7. Have human checkpoints for important decisions and a process to correct mistakes.

Trade‑offs and conflicts: Sometimes principles conflict—for example, a very complex model may give higher accuracy but lower explainability. Responsible AI means making such trade‑offs consciously, documenting them, and choosing the solution that best protects people’s rights and safety.

Classroom activity idea: Give students a short dataset and two models (one simple and explainable, one complex and more accurate). Ask them to compare which model they would choose for (a) medical diagnosis, (b) movie recommendation, and explain using the ethics principles.

Bias in AI means a systematic preference or error that causes a model to treat some people or groups differently from others. Bias can come from the data (e.g., unrepresentative samples), from how labels are collected (measurement bias), or from the design of algorithms. Unchecked bias may produce unfair or harmful outcomes.

Fairness means designing and using AI so outcomes are just and non-discriminatory. There are different views of fairness: group fairness (groups defined by age, gender, race should receive similar outcomes) and individual fairness (similar individuals should be treated similarly). Perfect fairness is often impossible; we aim to detect, measure, and reduce unfairness while keeping the model useful.

How bias appears: missing groups in training data, historical discrimination reflected in labels, proxy features that correlate with protected traits, or poorly chosen performance metrics. How to reduce bias: (1) Pre-processing — clean and balance data, remove or mask sensitive attributes; (2) In-processing — add fairness constraints or regularization to training; (3) Post-processing — adjust model outputs (thresholds) to equalize outcomes. Human oversight, transparency and continuous monitoring are also essential.

Important idea: fairness is measured with metrics (see formulas). If one group has systematically higher false positives or lower true positive rates than another, the model is biased. Trade-offs often exist between overall accuracy and fairness, so choices must be ethical and context-aware.

What is Privacy and Data Protection? Privacy is the right of individuals to control who knows what about them. Data protection are the practices, technologies and laws that keep personal data safe from misuse, unauthorized access, alteration or loss. In AI systems, privacy and data protection ensure that data used for training, inference and decision-making does not harm people or expose sensitive information.

Key principles

• Consent: Collect and use personal data only with clear permission (unless legally exempt).
• Purpose limitation: Use data only for the stated purpose.
• Data minimization: Collect only the minimum data needed.
• Accuracy: Keep data correct and up to date.
• Storage limitation: Retain data only as long as necessary.
• Integrity and confidentiality: Protect data from unauthorized access or corruption.
• Accountability: Organizations must demonstrate compliance with rules and safeguards.

Data lifecycle and risks

• Collection: Risk of over-collection or collecting sensitive data without need.
• Storage: Risk from weak encryption, insecure servers, or improper backups.
• Processing/Use: Risk of biased outputs, profiling or re-identification from aggregated data.
• Sharing: Risk when data is shared with third parties or across borders.
• Deletion: Risk if data is not permanently removed or backups remain accessible.

Technical protections

• Encryption: Transforming data so only authorized parties can read it (in transit and at rest).
• Pseudonymization & anonymization: Removing or replacing identifiers; anonymized data cannot be linked back to individuals.
• Differential privacy: Adding controlled noise to outputs so individual records cannot be inferred.
• Access controls & authentication: Role-based access, strong passwords and multi-factor authentication to limit who can see data.
• Secure protocols: Use HTTPS, secure APIs and audited data pipelines.
• Data governance: Policies, audits, logging and breach response plans.

Legal and ethical context

Many regions have laws (for example GDPR in the EU, national data protection laws) that give individuals rights such as access to their data, correction, deletion (right to be forgotten), and portability. Ethically, AI designers should prefer transparency, fairness and minimizing harm.

Practical tips for students

• Share only necessary personal details online and check app permissions.
• Use strong passwords and enable two-factor authentication.
• Be careful with public Wi‑Fi and avoid sending sensitive information over unsecured networks.
• Read privacy policies for apps and understand what data is collected and why.

Summary — Privacy and data protection keep personal information safe while allowing AI to deliver benefits. They combine technical tools, good practices and legal rules to reduce risks such as identity theft, surveillance, discrimination and loss of control over personal data.

What they mean: Transparency means being open about how an AI system works — its data, design, and decision process. Explainability means providing clear, understandable reasons for a specific AI decision so people can follow and trust it.

Why it matters: Transparent and explainable AI builds trust, enables accountability, helps find and fix bias or errors, and supports fair and safe use (for example in schools, hospitals, banks). Without them, decisions can seem arbitrary and harmful.

How it is achieved (simple approaches):

• Use interpretable models: linear regression or small decision trees where the rules are easy to read.
• Provide documentation: data sources, model purpose, known limitations (model cards, datasheets).
• Use post-hoc explainers: tools like LIME or SHAP produce human-friendly explanations for complex models by showing which features influenced a decision.
• Offer global vs local explanations: global explains overall model behaviour; local explains one specific decision.
• Keep a human-in-the-loop: allow people to review, override or question AI decisions.

Limitations and trade-offs: Often more complex models (deep neural networks) are less interpretable but more accurate. Designers must balance accuracy with the need for understandable decisions. Also, explanations should avoid revealing private data or be misleading.

Definition: Accountability and governance in AI mean creating clear rules, roles and processes so AI systems are developed, deployed and used responsibly. Governance defines the structures (policies, standards, boards, audits) while accountability ensures people and organizations can be held responsible for outcomes.

Why it matters: AI systems can affect safety, fairness, privacy and rights. Without governance and accountability, harms (bias, wrong decisions, privacy breaches) may go unaddressed and trust in AI falls.

Core principles:

• Transparency: Document how models are built, data sources, and decision logic where possible.
• Responsibility: Assign roles (developers, deployers, owners, auditors) with clear duties.
• Auditability: Keep logs, versioning and artifacts so systems can be reviewed and reproduced.
• Human oversight: Ensure meaningful human control especially when high risk is involved.
• Redress: Provide ways for affected people to challenge or appeal AI decisions.

Key components of effective governance:

• Policies and standards (ethical guidelines, data-use policies)
• Governance bodies (AI ethics board, compliance team)
• Risk assessment and impact assessments before deployment
• Testing, monitoring and continuous audits in production
• Incident response and remediation procedures

How it works in the AI lifecycle: Governance starts at problem definition (acceptable use), continues through data collection (consent, quality), model training (bias checks, explainability), validation (metrics, safety tests), deployment (access control, human-in-the-loop) and monitoring (drift detection, complaints handling).

Challenges: Technical opacity of models, trade-offs (accuracy vs. explainability), unclear legal frameworks, cross-border data rules, and organizational resistance.

Outcome: Good governance reduces risk, increases public trust and makes it possible to hold the right parties accountable when harms occur.

Safety in AI means designing systems so they do not cause unintended harm to people, property or the environment. Safety focuses on preventing accidents or harmful outputs even when the system encounters unusual situations.

Security is about protecting AI systems and their data from malicious actions: unauthorized access, tampering, theft or attacks that change the system's behavior.

Robustness describes an AI system's ability to keep working correctly when faced with noise, errors, changes in input, or attempted attacks. A robust system degrades gracefully instead of failing catastrophically.

How these ideas relate: they overlap. Security measures (e.g., authentication) reduce the chance that attackers cause unsafe behavior. Robustness (e.g., tolerance to noisy inputs) improves safety by preventing wrong outputs in unusual situations. Together they help build reliable, trustworthy AI.

Key principles and practices:

• Risk assessment: identify possible failures and harms before deployment.
• Testing & validation: unit tests, end-to-end tests, stress tests and adversarial tests that check behaviour on edge cases and malicious inputs.
• Fail-safe and graceful degradation: if the system is uncertain, it should alert a human or switch to a safe mode.
• Redundancy: use multiple independent sensors or models so one failure does not cause catastrophe.
• Access control & encryption: protect data and models from unauthorized users and tampering.
• Input validation & sanitization: reject or flag out-of-range or suspicious inputs.
• Monitoring & logging: detect faults or attacks in real time and allow post-incident analysis.
• Adversarial training & robust algorithms: train models to resist small, malicious changes to inputs.
• Human-in-the-loop: keep humans available for high-risk decisions.

Simple examples of failure modes:

• An image classifier that mistakes a stop sign with a sticker as a speed-limit sign (safety failure).
• A medical AI whose training data was manipulated to hide a disease pattern (security breach).
• A voice assistant that stops responding after background noise increases (lack of robustness).

Testing approaches (brief): robustness testing (add noise, distortions), adversarial testing (small deliberate perturbations), penetration testing (try to break security), and formal verification (mathematical checks for critical properties).

Outcome: Combining safety, security and robustness gives systems that operate correctly in normal conditions, resist attacks and unexpected inputs, and reduce the chance of harm.

Human-AI Interaction and Oversight describes how people and artificial intelligence (AI) systems work together, and how humans monitor, control and correct AI behaviour to ensure safe, fair and useful outcomes. Good interaction means AI assists humans effectively; good oversight ensures AI does not cause harm or make unchecked mistakes.

Key ideas:

• Roles: AI can act as an assistant (suggestions, predictions), an automator (actions taken automatically), or a decision-support tool (provides evidence, not the final decision).
• Interaction modes: Human-in-the-loop (human approves AI outputs), Human-on-the-loop (human supervises multiple AI actions and can intervene), and Human-in-command (human sets goals and rules; AI executes within them).
• Oversight mechanisms: continuous monitoring, explainability (AI shows reasons for outputs), confidence scores (AI shows how sure it is), thresholds for human review, regular audits, and feedback loops where human corrections improve the AI.
• Design principles: transparency, accountability, fairness, safety and usability. interfaces should show clear AI recommendations, supporting facts, and an easy way for users to correct or override AI.

Common risks and how oversight addresses them:

• Automation bias: people might over-trust AI. Oversight uses explanations and requires human approval for critical decisions.
• Errors and biases: AI trained on biased data can behave unfairly. Audits, diverse datasets and human review reduce this risk.
• Over-reliance: If AI is wrong, consequences can be large. Fallback procedures and human intervention options limit harm.

Practical workflow example: In a medical triage system, AI analyzes symptoms and gives a recommended priority and a confidence score. If confidence is below a threshold or the case is critical, a human doctor reviews and decides. The doctor’s corrections are logged and used to retrain the AI.

Why it matters for responsible AI: Human-AI interaction and oversight ensure AI systems help people without replacing essential human judgement, protect rights, and make systems trustworthy for society.

Overview: Artificial Intelligence (AI) changes how people live and work. Its social and economic impacts include improvements in productivity and services, shifts in the job market, changes in inequality, and new ethical and policy challenges. Responsible AI seeks to maximize benefits while reducing harms.

Social impacts:

• Positive: Better healthcare (faster diagnosis), improved accessibility (speech-to-text, assistive tech), personalized education, safer transport (advanced driver assistance).
• Negative: Privacy loss (surveillance, data misuse), biased decisions (discrimination in hiring, lending), algorithmic opacity (people don’t know how decisions are made), social isolation (over-reliance on automated systems).
• Distributional effects: Some groups gain more (tech-savvy, high-skill workers), while others can be left behind (low-skill workers, communities with less data access).

Economic impacts:

• Productivity & growth: AI can increase output per worker, lower costs, speed up R&D and create new goods and services.
• Employment effects: Automation may displace routine jobs but also creates new roles (AI engineers, data-labelers, maintenance). Net effect depends on pace of adoption and worker reskilling.
• Market structure: Large firms with more data can gain competitive advantage, potentially increasing market concentration.
• Wages & inequality: Demand rises for skilled workers, possibly increasing wage gaps unless policies/interventions close the skills gap.

Mitigation & Responsible Approaches:

• Design fair, transparent algorithms and perform bias audits.
• Invest in education and continuous training (reskilling/upskilling).
• Implement privacy protections, data governance, and explainability standards.
• Use social safety nets (unemployment support, job-transition programs) and promote inclusive access to AI benefits.

How students should think about it: When evaluating an AI application, ask: Who benefits? Who is harmed? Are decisions explainable and fair? What policies or training are needed to make the outcome equitable?

What are Ethical Design and Development Practices?

Ethical design and development practices are a set of principles and concrete steps that AI designers and developers follow to make sure AI systems are safe, fair, transparent, and respect people’s rights. These practices guide choices about data, models, testing, deployment, and monitoring so that AI benefits users and society while reducing harm.

Core ethical principles

• Fairness: Avoiding unfair bias against individuals or groups (for example, by gender, caste, religion, or socio‑economic status).
• Transparency and Explainability: Making decisions and inner workings understandable to users and auditors.
• Privacy: Protecting personal data and collecting only what is necessary.
• Accountability: Ensuring someone is responsible for decisions and that there are ways to correct mistakes.
• Safety and Robustness: Designing systems that are reliable under different real-world conditions.
• Inclusivity: Designing for diverse users, including people with disabilities.

Practical steps in the development process

1. Stakeholder analysis: Identify who will use the system and who might be affected. Include diverse voices early (users, domain experts, ethicists).
2. Careful data collection and labelling: Collect representative data and document its sources. Check for missing groups and label quality issues.
3. Preprocessing to reduce bias: Examine class imbalances, remove sensitive attributes when appropriate, or apply reweighting/sampling techniques to balance data.
4. Choose appropriate models: Prefer simpler or interpretable models when decisions affect people’s rights (e.g., loan approvals). Use complex models only when required and add explainability tools.
5. Privacy protection: Use data minimization, anonymization, and techniques such as differential privacy when handling personal data.
6. Testing and validation: Test models for accuracy, fairness (metrics by groups), robustness to noisy input, and safety edge cases.
7. Explainability tools: Provide understandable explanations of decisions (local explanations for individual decisions and global summaries of model behavior).
8. Documentation and audit trails: Keep documentation about data, model choices, testing results, and deployment decisions so audits are possible.
9. User consent and control: Inform users about data use and provide controls and appeals for automated decisions.
10. Monitoring and continuous improvement: Monitor performance after deployment, watch for data drift and fairness regressions, and update the system responsibly.

Why these practices matter: Ethical practices reduce harms such as discrimination, privacy breaches, and unsafe behavior. They increase trust in AI systems and often lead to more accurate and reliable systems because issues detected early are easier to fix.

Simple classroom checklist for ethical design

• Have we identified affected groups and obtained consent for data?
• Is the training data representative?
• Have we measured model performance separately for each group?
• Can the model’s decisions be explained in plain language?
• Is there a plan to monitor and correct errors after deployment?

Tip for students: When you build small AI projects, keep a short "Ethics log" that lists data sources, possible harms, fairness checks, and steps you took to reduce risks. This practice makes ethical thinking concrete and repeatable.

Overview
Mitigation measures and technical tools are actions, methods and software used to reduce harms, bias, privacy leaks, unfair outcomes and unsafe behaviour of AI systems. They combine organizational policies, careful data and model design, testing, and monitoring to make AI more responsible, fair, transparent and robust.

Typical mitigation workflow

• Identify risks: detect possible harms (bias, privacy, safety, misuse).
• Measure: choose metrics (accuracy, fairness metrics, privacy guarantees, robustness tests).
• Apply mitigations: use technical tools and process changes to reduce risk.
• Validate & test: evaluate on held-out and adversarial cases.
• Monitor & update: continuous logging, audits, human review and retraining.

Categories of mitigation measures

• Data-focused: better data collection, balancing under-represented groups, removing sensitive attributes when appropriate, labeling quality checks.
• Algorithmic: methods to reduce bias and increase robustness (pre-processing, in-processing, post-processing).
• Privacy: techniques to protect individuals (differential privacy, encryption, federated learning).
• Explainability & transparency: model explanations (LIME, SHAP), model cards and datasheets to communicate behaviour and limits.
• Human oversight: human-in-the-loop approvals for high-risk decisions and appeal processes.
• Operational: audits, impact assessments, access control, logging and incident response.

Key technical tools and short descriptions

• Pre-processing: change training data (reweighting, resampling) to reduce imbalance before training.
• In-processing: modify the learning algorithm to include fairness or robustness constraints or regularizers.
• Post-processing: alter model outputs (thresholds, calibration) to meet fairness criteria without changing model internals.
• Differential Privacy (DP): add controlled noise to outputs or gradients to limit what can be learned about any single individual. Provides a provable privacy bound (epsilon).
• Federated Learning: train models across many devices/clients without centralizing raw data; only model updates are shared.
• Explainable AI (XAI): local explanations (LIME, SHAP) show which features influenced a specific decision; global methods summarize overall behavior.
• Adversarial training & robustness: include adversarial examples at training time to reduce vulnerability to input manipulations.
• Model cards & datasheets: standardized documentation describing intended use, limitations, performance by subgroup and data provenance.
• Monitoring tools: dashboards tracking drift, fairness metrics and error rates in production; alerting for changes.

Trade-offs and best practices
No single tool solves every problem. For example, improving fairness may reduce a particular accuracy metric—choose the correct metric for the use case. Always combine technical tools with governance: policies, audits and human review are essential.

What this topic means
Legal, regulatory and policy aspects of Artificial Intelligence (AI) are the rules, laws and guidelines that govern how AI systems are designed, developed, deployed and used. They aim to protect people’s rights, ensure safety, promote fairness and provide ways to hold organizations accountable when AI causes harm.

Why they are needed
AI systems can affect privacy, fairness, safety and legal rights. Without rules, problems such as biased decisions, privacy violations, lack of accountability and unsafe systems can occur. Regulations and policies set minimum standards, create responsibilities and give people remedies when harms happen.

Key elements

• Data protection and privacy – rules about collecting, storing and using personal data (consent, purpose limitation, data minimization).
• Transparency and explainability – requirements to document how systems make decisions and to explain outcomes to affected people.
• Fairness and non‑discrimination – rules to detect and reduce biased outcomes across groups (e.g., race, gender).
• Safety and reliability – ensuring systems perform correctly and safely under expected conditions.
• Accountability and liability – who is responsible if an AI system causes harm (developer, operator, vendor).
• Standards, auditing and certification – processes for independent checks, audits and official approvals.
• Governance and oversight – institutional mechanisms (ethics boards, regulators) to monitor AI use.
• Cross‑border issues – handling data-transfer rules and differing laws between countries.

How regulation works in practice
Regulators create laws (binding rules), agencies publish regulations and guidance, and organizations adopt internal policies and technical controls to comply. Common practical steps include impact assessments (to spot risks before deployment), documentation (model cards, data sheets), testing for bias, logging for audits and complaint/redress mechanisms for affected users.

Stakeholder roles

• Developers – build systems following privacy-by-design, document decisions and enable explainability.
• Organizations that deploy AI – run risk assessments, monitor systems in production and provide user remedies.
• Regulators and policymakers – set rules, enforce laws and update policies as technology evolves.
• Users and civil society – raise concerns, demand transparency and push for fair practice.

Educational note for students
Knowing legal and policy aspects helps you design responsible AI systems. It means thinking beyond code: about people affected, how to explain choices, and how to prevent harm.

Case Studies and Real-world Examples

Case studies show how AI systems behave in real contexts and help students connect ethical principles (fairness, privacy, transparency, accountability, safety) to concrete outcomes. Studying specific incidents or deployments reveals where harms arose, which stakeholders were affected, and what technical or non‑technical fixes were used. Case studies also teach structured investigation: identify stakeholders, trace data and model choices, measure impacts, propose mitigations, and monitor results.

When analysing a case, consider the socio-technical lifecycle: data collection, data cleaning, model design, training, deployment, user interaction, and post-deployment monitoring. Ethical problems often come from biased data, opaque models, incentives that ignore harms, or missing governance. A good case study examines the root cause, not just the visible outcome.

Use a repeatable framework for each case:

1. Define the system and intended purpose.
2. Identify affected groups and potential harms (economic, social, safety, privacy).
3. Inspect the data and model decisions (features, labels, evaluation metrics).
4. Quantify harms with appropriate metrics (accuracy, false positive/negative rates, fairness measures).
5. Propose technical and policy mitigations (data rebalancing, explainability, human-in-the-loop, regulation).
6. Recommend monitoring, transparency, and accountability mechanisms.

Teaching note: Case studies are most useful when students practice measuring outcomes (confusion matrices, group‑wise errors), and discuss trade-offs — e.g., improving fairness for one group may change overall accuracy. Emphasize responsible deployment: informed consent, data minimisation, audit logs, clear owner for decisions, and remediation paths for affected people.

What it is: Assessment, Reporting and Redressal is the process used to evaluate AI systems for correctness, fairness, safety and privacy; to report problems or harms caused by AI; and to provide ways to fix mistakes and compensate affected people. These steps ensure AI is responsible, trustworthy and aligned with human values.

Why it matters: AI can make or support important decisions (loans, hiring, policing, medical advice). If an AI is wrong, biased, or harmful, people need a reliable way to detect the issue, inform the right parties, and obtain correction or remedy.

1. Assessment (Evaluation & Monitoring)

• Purpose: Measure if the AI meets requirements (accuracy, fairness, robustness, privacy).
• What to test: performance (accuracy, precision, recall), fairness across groups, safety under adversarial inputs, privacy leaks, and explainability.
• Methods: offline testing on labeled datasets, cross-validation, A/B tests in production, continuous monitoring of real-world outputs, third-party audits.
• Tools & artifacts: confusion matrices, ROC/PR curves, fairness metrics, model cards, test logs, audit reports.

2. Reporting (Transparency & Incident Reporting)

• Purpose: Make problems visible to developers, operators, users and regulators. Provide clear documentation about capabilities and limitations.
• Channels: in-app feedback button, incident reports to a support team, public reporting (transparency dashboards), regulator notifications when required.
• What to include in a report: description of issue, inputs that triggered it, user impact, timestamps, logs, and suggested priority.
• Documentation: model cards, data sheets for datasets, change logs, audit trails to track who changed what and when.

3. Redressal (Correction & Remedy)

• Purpose: Fix harms, correct decisions, compensate or make amends where required, and prevent recurrence.
• Steps: acknowledge receipt, investigate (reproduce the issue), remediate (bug fix, model retraining, rule override), communicate results to affected users, and update policies/processes.
• Mechanisms: human review/appeal processes, rollback of a model/version, targeted retraining on corrected data, fairness-aware reweighting, and procedural changes (improved testing, monitoring).

Stakeholders & Responsibilities

• Developers: build assessment tests, fix bugs, implement monitoring.
• Operators: monitor production, respond to incidents, maintain logs.
• Users: report problems, provide feedback.
• Regulators: set standards, require reporting for serious harms.

Best practices / Checklist

• Maintain clear documentation (model cards, dataset cards).
• Run regular audits for fairness and safety; include third-party audits when possible.
• Provide easy reporting channels and keep an audit trail of actions taken.
• Design an appeals process with human oversight for high-stakes decisions.
• Continuously monitor model drift and update models when performance degrades.

Example workflow (short): assess with test suite → detect bias via fairness metric → user files report → team investigates logs and reproduces issue → retrain or apply rules to fix issue → inform user and regulators if needed → add tests to prevent recurrence.

What it means

'AI for Good and Sustainable Use' refers to designing, deploying and governing artificial intelligence so that it benefits people and the planet while minimising harm, respecting rights and conserving resources. It connects Responsible AI principles (fairness, transparency, privacy, accountability) with sustainability (reduced energy use, low carbon footprint, long‑term societal benefit) and the United Nations Sustainable Development Goals (SDGs).

Key principles

• Human-centred design: AI should augment human capabilities, preserve dignity and require human oversight for critical decisions.
• Fairness and inclusion: Avoid biased training data, test systems across demographic groups and reduce disparate impacts.
• Transparency and explainability: Make how models reach decisions understandable to users and auditors.
• Privacy-preserving methods: Use techniques like differential privacy and federated learning to protect personal data.
• Energy efficiency and low-carbon operation: Choose smaller or pruned models, efficient hardware, and green data centres; measure and report emissions.
• Lifecycle thinking: Consider environmental and social impacts from data collection, model training, deployment, to disposal.

How to apply it (practical steps)

1. Start with clear, beneficial goals aligned with public good (e.g., healthcare access, disaster relief, energy savings).
2. Collect representative, consented data and document data sources (data sheets / model cards).
3. Choose model architectures that meet performance needs using minimal resources (model selection, pruning, quantization).
4. Measure performance and harms: accuracy, fairness metrics, energy consumption and CO2 emissions.
5. Deploy with human oversight, clear user consent, and regular audits to detect drift and unintended harms.
6. Use governance: stakeholder consultation, impact assessments and transparent reporting.

Benefits and risks

Benefits include faster diagnosis in healthcare, efficient energy use, better crop yields, improved accessibility, and faster disaster response. Risks include biased decisions, privacy breaches, increased inequality, and environmental cost from energy-intensive training. Balancing trade-offs (e.g., model accuracy versus energy use) is essential.