Easy ReviewAIF-C01
Fundamentals of AI and ML27

Responsible AI

The dimensions of responsible AI, how bias enters data and models, and how to make AI decisions transparent and explainable.

Este conteúdo de estudo ainda não foi traduzido, então aparece em inglês. Ajude a traduzir

Responsible AI means designing, building and using AI to maximize benefits and minimize harm. These sections follow the dimensions AWS defines, from fairness and explainability to safety and governance.

Dimensions of responsible AI

Responsible AI weighs the benefits of an AI system against its risks, across eight dimensions, from the first decision about whether AI is needed to the system’s retirement.

How it works

Dimension Question it asks
Fairness Does it treat stakeholder groups equitably?
Explainability Can we understand and evaluate its outputs?
Privacy and security Are data and models obtained, used and protected properly?
Safety Does it avoid harmful output and misuse?
Controllability Can we monitor and steer it?
Veracity and robustness Are outputs correct, even with unexpected input?
Governance Are best practices built into the AI supply chain?
Transparency Do people know when and how AI is used?

Exam cues

  • If the question says equal treatment of groups, think fairness.
  • If the question says correct output despite adversarial or unexpected input, think veracity and robustness.
  • If the question says disclose AI use and its limits, think transparency.
  • If the question says monitor and steer or switch off, think controllability.

Remember

  • Eight dimensions; veracity weighs more for generative AI.
  • Start by asking whether AI is required at all.

Trap: Explainability and transparency differ: explainability is understanding why the system produced an output; transparency is telling people that AI is used and how.

Da documentação da AWS (9)
  • Responsible AI is the practice of designing, developing and using AI technology with the goal of maximizing benefits and minimizing risks.
  • AWS defines responsible AI through eight dimensions: fairness, explainability, privacy and security, safety, controllability, veracity and robustness, governance, and transparency.
  • Fairness considers the impact on different groups of stakeholders, for example verifying that a mortgage system's decisions aren't influenced by discriminatory factors.
  • Veracity and robustness mean achieving correct system outputs, even with unexpected or adversarial inputs.
  • Controllability means having mechanisms to monitor and steer AI system behavior, including procedures to adjust or disable the system while keeping human oversight effective.
  • Safety means reducing harmful system output and misuse, keeping the system within a clearly defined use case scope.
  • Transparency lets stakeholders make informed choices about their engagement with an AI system, by communicating when and how AI is used, what data informs decisions and what controls exist.
  • Some responsible AI dimensions, such as veracity (truthfulness), weigh more heavily for generative AI systems than for traditional machine learning.
  • The Responsible AI Lens starts by clarifying the business problem and verifying that AI is actually required to solve it.

Bias, fairness and datasets

Bias usually enters through data: groups that are missing or underrepresented, labels that carry annotators’ assumptions, or examples that don’t match real users. It shows up as a model that works well for some groups and badly for others.

How it works

  1. Plan representative data: compare datasets with the real users and use cases; fill gaps.
  2. Check balance: measure each group’s share (class imbalance) and add data or use stratified sampling.
  3. Control labels: train annotators, measure agreement and look for annotator bias.
  4. Measure outcomes: compare accuracy, precision and recall across groups (subgroup analysis) and test paired inputs.
  5. Keep auditing: human audits and monitoring in production catch bias that appears later.

Exam cues

  • If the question says the model performs worse for one group, think subgroup analysis and unrepresentative data.
  • If the question says one group has far fewer examples, think class imbalance; collect or augment data.
  • If the question says inconsistent labels, think label quality controls.
  • If the question says detect bias and explain predictions on SageMaker AI, think SageMaker Clarify.

Remember

  • Fair datasets are representative, diverse, balanced and well labeled.
  • Measure fairness by comparing metrics across groups.

Trap: Overall accuracy can hide unfairness. A model can be 95% accurate overall and fail most often for a small group.

Da documentação da AWS (8)
  • Datasets should reflect the real-world conditions and users of the system; look for missing user groups or interaction styles and fill the gaps, especially in fine-tuning and evaluation sets.
  • To minimize unwanted bias, analyze the demographic makeup of datasets for over- and under-represented groups and fill gaps with more data from underrepresented groups or stratified sampling.
  • Group fairness is measured by comparing performance metrics such as accuracy, precision and recall across demographic groups and checking the differences against acceptable thresholds.
  • Testing pairs of inputs that differ only in a demographic attribute reveals whether demographic factors inappropriately change the output.
  • Label quality needs controls on human annotators, such as training, inter-rater agreement and checks for annotator bias, and synthetic labels should be validated against human judgment.
  • Fairness harms include harms to individuals, such as wrongful denials, and to groups, such as performance that varies across demographic groups.
  • Class imbalance is a pre-training bias metric that measures imbalance in the number of members between groups (facets), such as too little data for one age group.
  • SageMaker Clarify detects bias in pre-training data and in model predictions and explains predictions with feature attributions; it is no longer open to new customers.

Transparency and explainability

A transparent system tells people that AI is involved and how it was built; an explainable model lets you see why it produced an output. Both build trust, support audits and satisfy regulators.

How it works

Tool What it shows
Shapley values (SHAP) Each feature’s contribution to a prediction
Partial dependence plots How the prediction moves with a feature
SageMaker Model Cards Intended use, risk, training data, evaluation
AWS AI Service Cards Intended use and responsible-use notes for AWS AI services
Human oversight People review hard or consequential decisions
  1. Agree on the level of explainability the use case needs before choosing a model.
  2. Prefer interpretable models when decisions must be explained; add explanation techniques to complex ones.
  3. Document models (model cards), disclose AI use to users, and keep feedback paths and human review.

Exam cues

  • If the question says explain an individual prediction, think Shapley values (SHAP).
  • If the question says document a model’s intended use and risks, think SageMaker Model Cards.
  • If the question says intended use and limitations of an AWS AI service, think AWS AI Service Cards.
  • If the question says regulated decisions must be explainable, think interpretable models, accepting lower accuracy.

Remember

  • Interpretability: understand the model’s internal reasoning. Explainability: explain why it produced an output.
  • Simpler models are more transparent; complex ones often more accurate.

Trap: Explaining a model and documenting it are different jobs. Model Cards document; SHAP and PDPs explain.

Da documentação da AWS (8)
  • Explainability provides an understanding of how and why a model makes predictions, which builds trust, enables auditing and supports regulatory requirements.
  • Highly complex models such as deep neural networks may predict better but act as an opaque system, while simpler models such as decision trees and linear regression are more transparent; explainability needs must be balanced against accuracy.
  • Shapley values measure how much each feature contributed to a model's prediction, for a single prediction or for the model as a whole.
  • Partial dependence plots (PDPs) show how the predicted outcome depends on one or more input features.
  • SageMaker Model Cards document a model's intended use, risk rating, training details, metrics and evaluation results in one place.
  • AWS AI Service Cards explain the intended use cases of an AWS AI service, how it uses machine learning and key considerations for its responsible design and use.
  • Human review belongs at points where inputs or outputs are hard to judge or where human expertise helps with consequential decisions.
  • Model interpretability is the ability to understand and explain the reasoning behind a model's outputs.

Legal and business risks of generative AI

Generative AI creates new legal and business exposure: output that copies protected work, biased or toxic content, confident falsehoods, leaked data and the loss of trust that follows.

How it works

Risk Example Mitigation
Intellectual property Output resembles copyrighted work Content review, IP indemnity on covered services
Biased output Unfair treatment of a group Data curation, bias testing, guardrails
Hallucination Wrong policy given to a customer RAG grounding, contextual grounding checks, human review
Privacy Model reveals private data Data minimization, PII filters, access control
Loss of trust Users stop relying on the product Transparency, disclosure, feedback loops

Exam cues

  • If the question says output resembles a protected character, brand or text, think IP infringement claims.
  • If the question says users acted on false AI answers, think hallucination and end-user risk.
  • If the question says disclose AI use and limits to keep trust, think transparency.

Remember

  • AWS’s IP indemnity covers copyright claims on output of Amazon models and listed services.
  • Bias and toxicity come from training data; filter and test.

Trap: An indemnity covers certain copyright claims; it doesn’t make the output accurate, fair or safe to publish.

Da documentação da AWS (6)
  • AWS offers an uncapped intellectual property indemnity for copyright claims arising from the generative output of Amazon models and other generative AI services listed in the AWS Service Terms.
  • Foundation models can pick up hate speech and inappropriate undertones from training data, so developers should filter training data and encode specific norms into their models.
  • Foundation model answers on some subjects can be unreliable, inappropriate, toxic or incorrect.
  • A hallucination is output that is inconsistent, factually incorrect or unrelated to the input prompt.
  • Openness about an AI system's capabilities and limitations, including when and how AI is used, builds user trust and encourages adoption.
  • Using proprietary data to customize generative AI models raises data privacy and security concerns, so responses must not give unauthorized access to that data.

Flashcards

toque para virar
Resposta

·

Todos os 27 flashcards em lista

Name the eight dimensions AWS uses to define responsible AI.

Fairness, explainability, privacy and security, safety, controllability, veracity and robustness, governance, transparency.

A chatbot must keep giving correct answers even when users send typos, odd phrasing or adversarial input. Which responsible AI dimension is this?

Veracity and robustness.

A company tells customers when they are chatting with an AI and what data it uses. Which responsible AI dimension does this serve?

Transparency.

A loan model must not treat applicants differently because of gender or race. Which responsible AI dimension is this?

Fairness.

Operators need a way to monitor an AI system and switch it off if it misbehaves. Which responsible AI dimension is this?

Controllability.

Which responsible AI dimension matters more for generative AI than for traditional ML because models can invent facts?

Veracity (truthfulness).

Is responsible AI a final check before launch?

No. It applies across the whole lifecycle, starting with whether AI is needed at all.

A face verification model fails much more often for one region's users. Which check finds this?

Subgroup analysis, comparing error rates (accuracy, precision, recall) across groups against acceptable thresholds.

A training dataset has 90% of its examples from one age group. What is the problem and how do you fix it?

Class imbalance (under-representation). Collect more data for the underrepresented groups or use stratified sampling.

Name three characteristics of a dataset that supports fair models.

Any three of: representative of real users, diverse and inclusive across groups, balanced, accurately labeled, curated from relevant sources.

How can a team test whether a hiring model is biased by gender?

Send pairs of identical applications that differ only in gender and check whether the outputs change.

Different annotators label the same images inconsistently. Which responsible AI practice addresses this?

Label quality control - annotator training, clear guidelines, inter-rater agreement and expert review of disagreements.

Which SageMaker AI capability was built to detect bias in data and predictions and explain predictions?

SageMaker Clarify. It's no longer open to new customers.

Does removing the gender column from training data guarantee a fair model?

No. Other features can stand in for it, so measure outcomes across groups instead of assuming fairness.

A regulator requires a bank to explain which factors drove each individual loan decision. Which technique gives per-prediction explanations?

Feature attributions with Shapley values (SHAP), which show each feature's contribution to a specific prediction.

Which document tells customers the intended use, limitations and responsible-use considerations of an AWS AI service?

An AWS AI Service Card.

A highly regulated team must be able to read exactly how its model decides. Which kind of model fits better than a deep neural network?

An inherently interpretable model, such as a decision tree or linear regression, accepting some loss of accuracy.

What is the trade-off between model performance and interpretability?

Complex models (deep neural networks) often predict better but are opaque; simpler models are easier to explain but may be less accurate.

Which SageMaker AI feature documents a model's intended use and risk rating for governance?

SageMaker Model Cards.

An AI system recommends insurance plans. Name two human-centered design practices that build user trust.

Disclose that the recommendation is AI-generated and explain its main factors, and give users a way to give feedback or reach a human for consequential decisions.

Which visualization shows how a model's predicted outcome changes as one input feature changes?

A partial dependence plot (PDP).

Is an open-weight model automatically transparent and explainable?

Not automatically. Open weights help inspection, but you still need documentation of training data, licensing and explanation methods.

A marketing team's image generator produces a picture very similar to a famous cartoon character. Which legal risk is this?

An intellectual property (copyright) infringement claim from the character's owner.

Name four legal or business risks of using generative AI.

Any four of: intellectual property infringement claims, biased or toxic outputs, hallucinations that mislead users, loss of customer trust, exposure of private data.

A support bot confidently gives a customer wrong refund rules, and the customer acts on them. Which risks does this create?

End-user harm and loss of customer trust caused by a hallucination, plus possible legal liability.

What does AWS offer to customers facing copyright claims over output from Amazon's own generative models?

An uncapped intellectual property indemnity for copyright claims on output of Amazon models and the services listed in the Service Terms.

Does using a model from a large provider remove the risk of biased output?

No. Models pick up bias from their training data; the deployer still has to test, filter and monitor outputs.

Você concluiu todos os tópicos da AIF-C01!

Seu badge ficou dourado. Próximo passo: agendar a prova na AWS. Depois de passar, adicione-se ao Quem passou.

Compartilhar

Conte que você concluiu esta certificação

Concluí todos os tópicos da AWS Certified AI Practitioner (AIF-C01) no Easy Review. Venha estudar também!