AWS Learning Material Certification Cheatsheets
AIF-C01 Domain 4 14% domain

Task 4.1: Explain the development of AI systems that are responsible

Guidelines for Responsible AI · 4,791 words · source: Certified-AI-Practitioner-AIF-C01/domain-4/task-4-1-cheat-sheet.md

Domain 4: Guidelines for Responsible AI

Task Statement 4.1: Explain the development of AI systems that are responsible

Domain 4 is "Guidelines for Responsible AI" and represents 14% of the scored AIF-C01 exam content. Task 4.1 focuses on building AI systems responsibly: recognizing responsible AI characteristics, choosing models and datasets carefully, identifying legal and trust risks, understanding bias and variance, and knowing which AWS tools help detect, monitor, and reduce risk.

Use this sheet for questions such as "which responsible AI risk is present?", "which AWS service helps with bias detection or human review?", "how should a team make a dataset more inclusive?", "what does Guardrails do?", and "how should environmental impact affect model selection?"

Official Study Path

Use these in order:

  1. AWS Skill Builder: AWS Artificial Intelligence Practitioner Learning Plan
  2. AWS Skill Builder: Exam Prep Plan: AWS Certified AI Practitioner (AIF-C01)
  3. AWS Skill Builder: Domain 4 Review: AWS Certified AI Practitioner
  4. AWS Skill Builder: Domain 4 Practice: AWS Certified AI Practitioner
  5. AWS Exam Guide: Content Domain 4, Task Statement 4.1
  6. AWS Docs: Responsible AI, Amazon Bedrock Guardrails, SageMaker Clarify, SageMaker Model Monitor, Amazon A2I, SageMaker Ground Truth, and the AWS Well-Architected Generative AI Lens

Exam Objective Map

You should be able to answer questions about:

Objective What to know for the exam
Responsible AI features Bias, fairness, inclusivity, robustness, safety, veracity, explainability, privacy and security, controllability, governance, transparency
Amazon Bedrock Guardrails Content filters, prompt attack detection, denied topics, word filters, sensitive information filters, contextual grounding checks, automated reasoning checks
Responsible model selection Fit the model to the use case, avoid unnecessary GenAI, prefer smaller or optimized models when quality is sufficient, consider cost, latency, energy use, and sustainability
GenAI legal and trust risks Intellectual property claims, biased outputs, hallucinations, loss of customer trust, end-user harm, privacy exposure, unsafe content
Dataset characteristics Inclusive, diverse, representative, balanced, curated, high-quality, documented, labeled consistently, appropriate for the target population
Bias and variance Bias can create systematic unfairness or underfitting; variance can create instability or overfitting; both affect reliability and subgroup outcomes
Detection and monitoring tools SageMaker Clarify, SageMaker Model Monitor, Amazon A2I, human audits, subgroup analysis, label quality review, Ground Truth

One-Minute Mental Model

Responsible AI is not a single feature. It is a lifecycle discipline:

  1. Define the use case: Is AI appropriate, and what harm could happen if it is wrong?
  2. Select data: Is the dataset representative, inclusive, balanced, curated, and legally usable?
  3. Select the model: Is the model accurate enough, cost-effective, efficient, and appropriate for the risk level?
  4. Add controls: Use guardrails, policies, validation, human review, and monitoring.
  5. Evaluate subgroups: Check whether quality, error rates, and outcomes differ across user groups.
  6. Monitor in production: Detect drift, bias, hallucinations, unsafe content, and loss of trust.
  7. Improve continuously: Use feedback, audits, retraining, better data, and governance processes.

Exam shortcut:

Responsible AI = maximize benefit while minimizing risk across data, model, application, users, and operations.

1. Responsible AI Features

AWS describes responsible AI through dimensions such as fairness, explainability, privacy and security, safety, controllability, veracity and robustness, governance, and transparency. The exam objective for Task 4.1 emphasizes bias, fairness, inclusivity, robustness, safety, and veracity.

Feature What it means Exam clue
Bias Systematic skew in data, labels, model behavior, or outcomes "Certain demographic groups receive worse results"
Fairness The system considers impacts across groups and avoids unjust outcomes "Evaluate whether decisions are equitable"
Inclusivity The system works for diverse users, languages, contexts, and abilities "Dataset should represent the target population"
Robustness The system remains reliable with noisy, unexpected, or adversarial inputs "Model should handle typos, edge cases, or unusual inputs"
Safety The system reduces harmful outputs, misuse, and unsafe actions "Prevent harmful content or dangerous recommendations"
Veracity Outputs should be truthful, grounded, and correct "Detect hallucinations or unsupported claims"
Controllability Humans and systems can monitor, steer, restrict, or disable behavior "Need mechanisms to enforce scope and policy"
Governance Policies, ownership, documentation, review, and accountability exist "Who approves, audits, and manages the AI lifecycle?"
Transparency Stakeholders understand when AI is used and what its limitations are "Users should know AI is involved and where answers come from"

Exam Pattern

If the question mentions unfair outcomes across groups, think bias, fairness, subgroup analysis, and SageMaker Clarify.

If the question mentions harmful content, prompt injection, denied topics, PII, or hallucination filtering, think Amazon Bedrock Guardrails.

If the question mentions low-confidence predictions or sensitive decisions needing review, think Amazon A2I or human-in-the-loop review.

If the question mentions production drift, changing live data, or alerts, think SageMaker Model Monitor and SageMaker Clarify monitoring.

2. Bias, Fairness, and Inclusivity

Bias can enter an AI system at many points:

Source of bias What happens Example
Historical bias Historical inequities are reflected in training data Past hiring data underrepresents some groups
Sampling bias Data does not represent the target population Voice assistant trained mostly on one accent
Selection bias Some users, cases, or records are more likely to be included Survey data excludes people without internet access
Measurement bias Features or labels measure the wrong thing or measure groups differently Device data is less accurate for some populations
Label bias Human labels are inconsistent, subjective, or prejudiced Reviewers score similar content differently by group
Aggregation bias One model is used for groups with different patterns Medical model ignores demographic-specific risk factors
Deployment bias The system is used in a different context than intended A model trained for recommendations is used for eligibility decisions
Feedback-loop bias Model decisions affect future training data A fraud model investigates one group more often, creating more labels for that group

Fairness Review Questions

Ask these before trusting the model:

  • Is the training data representative of the population that will use the system?
  • Are important groups underrepresented or missing?
  • Are labels accurate and consistent across groups?
  • Are protected or sensitive attributes used directly or indirectly through proxies?
  • Are error rates different across demographic groups or user segments?
  • Was the model evaluated on subgroups, not only on average performance?
  • Can affected users challenge, correct, or appeal an AI-supported decision?

Exam Trap

High overall accuracy can hide unfair subgroup performance. A model can look good on average while failing badly for a smaller group.

3. Dataset Characteristics

Responsible AI starts with responsible data. For the exam, know what a good dataset looks like and what weak data signals look like.

Dataset characteristic Why it matters
Representative The data reflects the users, cases, languages, locations, and conditions expected in production
Inclusive Diverse groups and accessibility needs are included
Diverse The model sees enough variation to generalize beyond narrow examples
Balanced No important group or class is so underrepresented that the model ignores it
Curated Data sources are selected, cleaned, reviewed, and appropriate for the task
High quality Records are accurate, complete, deduplicated, and relevant
Label quality Labels are consistent, accurate, reviewed, and produced with clear instructions
Documented Source, permissions, collection method, limitations, and intended use are known
Current enough Data reflects the period and context where the model will be used
Legally usable The organization has rights to use the data for the intended AI purpose

Dataset Red Flags

  • Data comes from unknown or untrusted sources.
  • Sensitive groups are missing or underrepresented.
  • Labels were created without clear guidelines.
  • The dataset reflects outdated behavior or policy.
  • The model is deployed to a population different from the training data.
  • The data contains copyrighted, private, or regulated content without proper rights or controls.
  • The class distribution is extremely imbalanced without mitigation.

Label Quality and Human Audits

Label quality matters because supervised ML systems learn from labels. Poor labels create poor models, and biased labels can create biased outcomes.

Use these methods:

  • Clear labeling instructions and examples.
  • Multiple reviewers for subjective labels.
  • Annotation consolidation or agreement checks.
  • Human audits of random samples and edge cases.
  • Review labels by subgroup to detect inconsistent treatment.
  • Use Amazon SageMaker Ground Truth when managed human data labeling is needed.

4. Amazon Bedrock Guardrails

Amazon Bedrock Guardrails provides configurable safeguards for generative AI applications. It can be used to help detect and filter undesirable content, protect sensitive information, restrict topics, and check whether responses are grounded in source information.

Guardrail Components to Know

Guardrail component What it does Exam clue
Content filters Detect and filter harmful text or image content Hate, insults, sexual content, violence, misconduct
Prompt attack detection Detect jailbreaks, prompt injection, and prompt leakage attempts "User tries to override system instructions"
Denied topics Block topics that the application should avoid "Banking assistant must not discuss illegal investment advice"
Word filters Block configured words or phrases Profanity, competitor names, custom blocked terms
Sensitive information filters Detect, block, or mask PII and custom regex entities "Redact customer personal information"
Contextual grounding checks Detect responses that are not grounded in source content or not relevant to the query "RAG answer invents facts not in retrieved passages"
Automated reasoning checks Validate that responses follow logical rules and policies "Response must comply with defined business rules"

What Guardrails Is Good For

Use Bedrock Guardrails when the question says:

  • Block harmful or unsafe content.
  • Prevent answers on restricted topics.
  • Detect prompt injection or jailbreak attempts.
  • Mask or block PII in prompts and responses.
  • Reduce hallucination risk in RAG by checking grounding.
  • Enforce application-specific policy boundaries.

What Guardrails Is Not

Guardrails does not replace:

  • Dataset quality work.
  • Bias analysis across demographic groups.
  • Human review for high-risk decisions.
  • Security controls such as IAM, encryption, and network controls.
  • Legal review of data rights and output use.
  • Full evaluation of business outcomes.

Exam Shortcut

Bedrock Guardrails = runtime safeguards for GenAI inputs and outputs.
SageMaker Clarify = bias and explainability analysis for data, models, and predictions.

5. Responsible Model Selection and Sustainability

Responsible model selection means choosing a model that meets the business need without unnecessary risk, cost, latency, or environmental impact.

Model Selection Questions

Question Why it matters
Is GenAI necessary? A simpler search, rules, analytics, or traditional ML approach may be safer and more efficient
What is the task risk? Higher-risk decisions need stronger evaluation, controls, transparency, and human oversight
What quality is required? Do not choose the largest model if a smaller model meets the target
What latency is acceptable? Real-time applications may need smaller or optimized models
What cost is acceptable? Larger models and longer outputs usually cost more
What data will be sent to the model? Sensitive data requires privacy, security, and governance controls
Does the model support needed controls? Guardrails, evaluation, customization, region availability, and logging can matter
What is the environmental impact? Right-sized models and optimized inference reduce resource consumption

Sustainability Patterns

AWS Well-Architected guidance for generative AI emphasizes:

  • Ask whether generative AI is the right solution.
  • Use the simplest architecture that meets the requirement.
  • Right-size the model for the use case.
  • Prefer smaller models when they meet quality goals.
  • Use optimized inference techniques when appropriate.
  • Consider model distillation, quantization, pruning, and efficient architectures.
  • Evaluate optimized models to confirm quality, latency, throughput, and cost.
  • Avoid wasting compute with over-provisioned or idle resources.

Exam Pattern

If the question says "reduce environmental impact" or "improve sustainability" while maintaining acceptable quality, favor smaller models, right-sized models, optimized inference, or model distillation.

If the question says "simple task with strict latency and cost constraints," do not pick the largest available model.

If the question says "GenAI is being proposed for a simple lookup," consider whether a simpler approach is more appropriate.

6. Legal, Trust, and End-User Risks of GenAI

Generative AI introduces risks that must be identified and managed before deployment.

Risk What it means Mitigation pattern
Intellectual property infringement Training data, prompts, or generated outputs may involve copyrighted or licensed material Use approved data sources, review licenses, add legal review, cite sources when needed
Biased model outputs Outputs treat groups unfairly or reinforce stereotypes Diverse data, subgroup testing, SageMaker Clarify, human audits
Hallucinations The model generates plausible but unsupported or false claims RAG, citations, contextual grounding checks, evaluation, human review
Loss of customer trust Users lose confidence because the system is wrong, opaque, unsafe, or misleading Transparency, clear limits, quality monitoring, feedback mechanisms
End-user risk Users may act on incorrect, unsafe, or inappropriate recommendations Guardrails, disclaimers, human escalation, restricted scope
Privacy exposure Sensitive data is exposed in prompts, outputs, logs, or training data PII filtering, data minimization, access controls, encryption, logging controls
Unsafe content The model produces harmful, abusive, or policy-violating output Bedrock Guardrails content filters and denied topics
Overreliance Users treat AI output as authoritative when it is uncertain Human review, confidence thresholds, source citations, clear limitations

Exam Trap

Hallucination is not the same as bias. Hallucination is unsupported or false output. Bias is systematic skew or unfair treatment.

7. Bias and Variance

Bias and variance are classic ML concepts, but the exam connects them to responsible AI outcomes.

Concept Meaning Failure mode Responsible AI impact
High bias Model is too simple or makes strong wrong assumptions Underfitting Inaccurate results across many users or systematic errors for groups
High variance Model is too sensitive to training data noise Overfitting Unstable predictions and poor generalization to new users or edge cases
Bias in fairness sense Data or model produces systematic unfair outcomes Disparate treatment or impact Some demographic groups receive worse results

Underfitting vs Overfitting

Problem Training performance Test/production performance Exam clue
Underfitting Poor Poor Model too simple, misses real patterns, high bias
Overfitting Very good Poor Memorizes training data, fails on new data, high variance
Good fit Good Good Generalizes to unseen data

Effects on Demographic Groups

Bias and variance can affect groups differently:

  • Underrepresented groups may have higher error rates.
  • Overfit models may perform well on majority groups and poorly on rare cases.
  • Labeling inconsistencies can create unfair treatment.
  • Imbalanced data can cause a model to ignore minority classes.
  • A model can be accurate overall but inaccurate for specific subgroups.

Exam Shortcut

High bias = too simple or systematically wrong.
High variance = too sensitive, overfit, unstable.
Fairness bias = unequal or unjust outcomes across groups.

8. Tools to Detect and Monitor Bias, Trustworthiness, and Truthfulness

Tool or method Use it for Exam clue
Amazon Bedrock Guardrails Runtime GenAI safeguards, PII filtering, denied topics, prompt attacks, contextual grounding "Block unsafe output, redact PII, prevent hallucinated RAG answers"
Amazon SageMaker Clarify Detect pre-training bias, post-training bias, explain predictions, monitor bias and feature attribution drift "Find whether data or model predictions are biased"
SageMaker Model Monitor Monitor deployed models for data quality, model quality, bias drift, and explainability drift "Production model drift or threshold alerts"
Amazon A2I Add human review for low-confidence predictions or random audits "Human-in-the-loop review"
SageMaker Ground Truth Build high-quality labeled datasets with human labeling workflows "Need managed data labeling and label quality"
Human audits Expert review of samples, outputs, labels, and edge cases "SMEs must review outcomes"
Subgroup analysis Compare metrics across demographic or user segments "Model works overall but may fail for one group"
Label quality analysis Check consistency and correctness of labels "Labels may be wrong, subjective, or inconsistent"

SageMaker Clarify

SageMaker Clarify helps with fairness, explainability, and bias detection.

Know this:

  • It can identify bias in pre-training data.
  • It can identify post-training bias in data and model predictions.
  • It can explain model predictions using feature attributions.
  • It can generate reports for data science, business, risk, and compliance stakeholders.
  • It can monitor production models for bias drift and feature attribution drift.
  • Fairness should be considered throughout the lifecycle: problem formation, dataset construction, algorithm selection, training, testing, deployment, monitoring, and feedback.

Use Clarify when the question asks:

  • Is the dataset biased before training?
  • Did training introduce or amplify bias?
  • Why did the model make a prediction?
  • Are prediction outcomes different across groups?
  • Has production bias drifted beyond a threshold?

SageMaker Model Monitor

Model Monitor helps detect production issues after deployment.

Know this:

  • Production data can differ from training data.
  • Drift can be temporary or permanent.
  • Monitoring only at training time is not enough.
  • Model Monitor can work with Clarify to monitor bias drift and feature attribution drift.
  • Alerts can be configured through Amazon CloudWatch when thresholds are exceeded.

Use Model Monitor when the question asks:

  • Did live data distribution change?
  • Is model quality degrading in production?
  • Are monitored metrics outside acceptable ranges?
  • Do teams need scheduled checks and alerts?

Amazon A2I

Amazon Augmented AI adds human review to ML workflows.

Know this:

  • It is used for low-confidence predictions or random prediction samples.
  • It helps avoid building custom human review systems.
  • It provides built-in workflows for common use cases and custom workflows for other ML models.
  • It can support content moderation, document extraction, real-time inferences, sentiment analysis, transcription, translation, and tabular data review.

Use A2I when the question asks:

  • Humans should review uncertain predictions.
  • Sensitive decisions need human oversight.
  • A random sample of predictions should be audited.
  • Human review results should help improve future model behavior.

9. Responsible AI Lifecycle Checklist

Lifecycle stage Responsible actions
Problem definition Confirm AI is appropriate; define stakeholders, harms, success metrics, and escalation paths
Data sourcing Use curated, legal, representative, diverse, inclusive data sources
Labeling Use clear instructions, quality checks, reviewer agreement, and audits
Training/customization Avoid unnecessary complexity; evaluate bias, variance, and subgroup performance
Model selection Right-size the model; consider safety, latency, cost, sustainability, and controls
Evaluation Test quality, robustness, safety, veracity, subgroup outcomes, and edge cases
Deployment Add guardrails, logging choices, access controls, human review, and user transparency
Monitoring Track drift, bias, harmful outputs, hallucinations, user feedback, and business outcomes
Improvement Update data, adjust guardrails, retrain if needed, and document decisions

10. Decision Shortcuts for Exam Questions

Scenario Best answer pattern
Chatbot must avoid harmful content Amazon Bedrock Guardrails content filters
User tries to override system instructions Bedrock Guardrails prompt attack detection
App must not discuss restricted topics Bedrock Guardrails denied topics
App must redact PII in prompts or responses Bedrock Guardrails sensitive information filters
RAG answer invents facts not in retrieved context Bedrock Guardrails contextual grounding checks
Need to check if training data is biased SageMaker Clarify pre-training bias analysis
Need to check if model predictions are biased after training SageMaker Clarify post-training bias analysis
Need production alerts for bias drift SageMaker Clarify with Model Monitor and CloudWatch alerts
Need human review for low-confidence predictions Amazon A2I
Need high-quality labeled training data SageMaker Ground Truth
Need to see whether one demographic group has worse results Subgroup analysis and fairness metrics
Need to reduce carbon footprint while preserving quality Smaller model, optimized inference, model distillation, or right-sizing
Need to reduce hallucinations Ground responses with RAG, citations, evaluations, and contextual grounding checks
Need to address IP risk Use approved/licensed data and legal review
Overall accuracy is high but one group has poor outcomes Do not trust average-only metrics; evaluate subgroup performance

11. Common Wrong Answers

  • Use only overall accuracy to prove the model is fair.
  • Assume a larger model is always more responsible.
  • Use GenAI for a task that a simpler system can solve.
  • Treat Bedrock Guardrails as a replacement for legal review or dataset quality.
  • Treat SageMaker Clarify as a runtime harmful-content filter.
  • Ignore subgroup performance because aggregate metrics look good.
  • Deploy a model without monitoring because it passed offline evaluation.
  • Assume removing a sensitive attribute automatically removes bias; proxies can still exist.
  • Fine-tune a model on unlicensed or poorly curated data.
  • Let low-confidence predictions make high-impact decisions without human review.
  • Assume RAG eliminates hallucinations completely.
  • Optimize for sustainability without re-evaluating quality and safety.

12. Mini Flashcards

Q: What is responsible AI? A: Designing, developing, and using AI to maximize benefits while minimizing risks.

Q: Name six responsible AI features from Task 4.1. A: Bias, fairness, inclusivity, robustness, safety, and veracity.

Q: What AWS service provides runtime safeguards for GenAI applications? A: Amazon Bedrock Guardrails.

Q: What Guardrails feature helps detect hallucinations in RAG responses? A: Contextual grounding checks.

Q: What Guardrails feature helps protect PII? A: Sensitive information filters.

Q: What AWS service helps detect bias in datasets and model predictions? A: Amazon SageMaker Clarify.

Q: What service helps monitor production models for drift? A: SageMaker Model Monitor.

Q: What service adds human review for low-confidence predictions? A: Amazon Augmented AI (Amazon A2I).

Q: What service helps create high-quality labeled training datasets? A: Amazon SageMaker Ground Truth.

Q: What does high bias usually cause? A: Underfitting and systematic inaccuracy.

Q: What does high variance usually cause? A: Overfitting and poor generalization to new data.

Q: Why are subgroup metrics important? A: Overall metrics can hide poor performance for specific demographic or user groups.

Q: What should you consider for sustainable model selection? A: Whether GenAI is needed, model size, resource use, optimized inference, cost, latency, and quality.

Q: Name common legal or trust risks of GenAI. A: IP infringement, biased outputs, hallucinations, privacy exposure, loss of customer trust, and end-user harm.

13. Practice Questions

  1. A loan screening model has high overall accuracy, but approval errors are much higher for one demographic group. What should the team investigate?

    • Answer: Bias and subgroup performance. Use subgroup analysis and SageMaker Clarify to detect and measure bias.

  2. A customer support chatbot must block violent, hateful, and insulting responses. Which AWS capability fits best?

    • Answer: Amazon Bedrock Guardrails content filters.

  3. A RAG assistant answers from company policies but sometimes adds claims that are not in the retrieved passages. Which Guardrails capability helps?

    • Answer: Contextual grounding checks.

  4. A company wants humans to review model predictions when confidence is low. Which AWS service should it use?

    • Answer: Amazon Augmented AI (Amazon A2I).

  5. A deployed model was fair during testing, but live traffic has changed and bias may now exceed a threshold. What should the team use?

    • Answer: SageMaker Clarify bias monitoring with SageMaker Model Monitor and CloudWatch alerts.

  6. A team is choosing between a large FM and a smaller FM. The smaller model meets quality targets with lower latency and lower resource use. Which is more responsible?

    • Answer: The smaller/right-sized model, because it meets requirements with lower cost, latency, and environmental impact.

  7. A model performs extremely well on training data but poorly on new examples. What is the likely issue?

    • Answer: High variance and overfitting.

  8. A model performs poorly on both training and test data because it is too simple. What is the likely issue?

    • Answer: High bias and underfitting.

  9. A team trained on data from one region and wants to deploy globally. What dataset issue should it evaluate?

    • Answer: Representativeness and inclusivity across regions, languages, cultures, and user groups.

  10. A generative AI app may produce copyrighted-looking output or use unapproved training data. What risk is this?

    • Answer: Intellectual property infringement risk.

  11. A developer wants to prevent users from asking a banking assistant for illegal investment advice. Which Guardrails component applies?

    • Answer: Denied topics.

  12. A dataset has inconsistent human labels for the same class. What responsible AI practice should be used?

    • Answer: Label quality analysis, clear labeling guidance, annotation consolidation or reviewer agreement checks, and human audits.

  13. A product team says the model is safe because protected attributes were removed from the training data. What is the problem with this reasoning?

    • Answer: Proxy variables can still encode sensitive information, so subgroup analysis and bias testing are still needed.

  14. A team wants to reduce compute consumption for a GenAI workload without hurting task quality. Which approaches fit?

    • Answer: Right-size the model, use a smaller model, optimize inference, use quantization or pruning where appropriate, or use model distillation.

  15. A model gives fluent but false answers. Which responsible AI feature is most directly involved?

    • Answer: Veracity, or truthfulness.

14. Last-Day Review Checklist

  • I know Domain 4 is 14% of the scored AIF-C01 exam.
  • I can define responsible AI as maximizing benefit while minimizing risk.
  • I can explain bias, fairness, inclusivity, robustness, safety, and veracity.
  • I can distinguish hallucination risk from bias risk.
  • I know Bedrock Guardrails is for runtime GenAI safeguards.
  • I can list Guardrails components: content filters, prompt attacks, denied topics, word filters, sensitive information filters, contextual grounding checks, and automated reasoning checks.
  • I know SageMaker Clarify detects bias in data and model predictions and supports explainability.
  • I know SageMaker Model Monitor helps monitor deployed models for drift and alerts.
  • I know Amazon A2I provides human review for low-confidence predictions or audits.
  • I know SageMaker Ground Truth supports managed human data labeling.
  • I can explain why representative, inclusive, balanced, curated datasets matter.
  • I can explain how poor label quality can create biased or unreliable models.
  • I can distinguish high bias/underfitting from high variance/overfitting.
  • I can explain why subgroup analysis is necessary even when overall metrics are good.
  • I can identify GenAI legal and trust risks: IP claims, biased outputs, hallucinations, privacy exposure, end-user harm, and loss of trust.
  • I can choose smaller, right-sized, or optimized models when they meet requirements and improve sustainability.

Official Sources