8) AI Challenges and Responsibilities

Vocabulary flashcards covering Responsible AI, security risks, governance, compliance standards, AWS AI services, and monitoring concepts based on lecture notes.

Responsible AI

The practice of ensuring AI systems are transparent and trustworthy, and mitigating potential risks and negative outcomes throughout the AI lifecycle (design, development, deployment, monitoring, evaluation).

AI Security

The maintenance of confidentiality, integrity, and availability on organizational data, information assets, and infrastructure within AI implementations.

AI Governance

The framework of clear policies, guidelines, and oversight mechanisms established to manage risk, add business value, and align AI systems with legal and regulatory requirements.

AI Compliance

Ensuring adherence to legal regulations and guidelines, particularly in sensitive domains such as healthcare, finance, and legal applications.

Fairness (Eight Pillars of Responsible AI)

Promotes inclusion and prevents discriminatory practices or biased outcomes across demographic groups.

Explainability (Eight Pillars of Responsible AI)

Provides mechanisms to understand the nature, inputs, outputs, and behavioral mechanics of the underlying model.

Privacy and Security (Eight Pillars of Responsible AI)

Guarantees that individuals maintain complete control over when, how, and if their data is collected and processed, while maintaining system protection.

Transparency (Eight Pillars of Responsible AI)

Ensures openness regarding how AI models are built, trained, deployed, and evaluated.

Veracity and Robustness (Eight Pillars of Responsible AI)

Guarantees that AI solutions remain accurate, truthful, resilient, and reliable even when encountering unexpected inputs or adversarial conditions.

Governance (Eight Pillars of Responsible AI)

Defines, implements, and enforces formal operational workflows, accountability structures, and responsible AI practices.

Safety (Eight Pillars of Responsible AI)

Ensures that algorithms operate in a manner that is physically, socially, and psychologically safe and beneficial for individuals and society.

Controllability (Eight Pillars of Responsible AI)

Preserves human agency by ensuring the ability to align AI systems to human values, intent, and intervention mechanisms.

Amazon Bedrock Guardrails

An AWS security tool that filters harmful content, redacts PII, blocks undesirable topics, and enhances overall safety and privacy for foundation models.

SageMaker Clarify

An AWS tool used to evaluate foundation models on accuracy, robustness, and toxicity, as well as detect dataset and model bias.

SageMaker Data Wrangler

An AWS tool used to fix bias in machine learning workflows by balancing datasets, such as augmenting data for underrepresented groups.

Amazon Augmented AI (A2I)

An AWS service that provides human review of machine learning predictions to ensure quality and accuracy.

AWS AI Service Cards

A form of responsible AI documentation provided by AWS that helps users understand service features, intended use cases, limitations, design choices, and deployment best practices.

Interpretability

The degree to which a human can understand the cause of an AI model's decision and answer 'why and how', where more of this often results in lower model performance.

Decision Tree

A supervised learning algorithm used for classification and regression tasks that splits data into branches based on feature values; highly interpretable but prone to overfitting if branches are excessive.

Partial Dependence Plot (PDP)

A plot showing how a single feature influences the predicted outcome of an ML model while holding other features constant, aiding interpretability for 'black box' models.

Design for Amplified Decision-Making (HCD)

Enhances human decision efficiency while minimizing user cognitive load and operational error in high-pressure or stressful environment settings.

Design for Clarity, Simplicity, and Usability (HCD)

Ensures AI interfaces and output visual formats are easily interpretable without ambiguity.

Design for Reflexivity and Accountability (HCD)

Encourages human operators to critically reflect on AI recommendations rather than blindly trusting automated outputs.

Design for Unbiased Decision-Making (HCD)

Guarantees that the human-AI interaction is engineered to mitigate human cognitive biases and model bias. Decision-makers must be explicitly trained to recognize and counteract biases.

Cognitive Apprenticeship

A human-centered design concept where AI systems learn directly from human instructors and domain experts.

AI Toxicity

The generation of offensive, disturbing, or inappropriate output by an AI model, mitigated by curating training data and implementing guardrail models.

AI Hallucination

The generation of assertions or claims by Large Language Models (LLMs) that sound plausible and true due to next-word probability sampling, but are factually incorrect or non-existent.

Data Poisoning

The intentional introduction of malicious or biased data into an AI model's training dataset, leading to biased, harmful, or incorrect model outputs.

Prompt Injection

A vulnerability where an attacker embeds specific instructions into user prompts to hijack the model's behavior and produce unauthorized or malicious outputs.

Prompt Exposure

The risk of revealing sensitive or confidential information to an AI model during training or inference, potentially causing data leaks from the training corpus.

Prompt Leaking

The unintentional disclosure or leakage of internal system prompts, instructions, or operational context used within an AI model.

Jailbreaking

The technique of circumventing ethical and safety constraints implemented in a generative model to gain unauthorized access, functionality, or prohibited outputs.

Regulated Workloads

Applications in industries such as finance or healthcare that must satisfy strict legal requirements, audit procedures, archival rules, and custom security frameworks.

Model Cards

Standardized documentation detailing key aspects of an ML model, including dataset sources, licenses, biases, intended use, risk ratings, and performance metrics.

Data Lineage

The end-to-end tracing and documentation of data origins, including source citations, collection methods, pre-processing transformations, cleaning steps, and dataset cataloging.

AWS Shared Responsibility Model

A cloud security model defining AWS's responsibility for security 'of' the cloud (infrastructure, hardware, managed services) and the customer's responsibility for security 'in' the cloud (data, IAM, OS, configurations).

Generative AI Security Scoping Matrix

A framework that classifies generative AI deployments into 5 defined scopes (ranging from Consumer Apps at Scope 1 to Self-trained Models at Scope 5) to assess ownership and manage security risks.

Scope 1 - Consumer App

Using public, non-enterprise Generative AI applications directly (e.g. ChatGPT, Midjourney)

Scope 2 - Enterprise App

Leveraging enterprise SaaS platforms integrated with built-in AI features (e.g. Salesforce Einstein GPT, Amazon Q Developer)

Scope 3 - Pre-trained Models

Building custom applications on managed base foundation models (e.g. Amazon Bedrock base models)

Scope 4- Fine-tuned Models

Adapting pre-trained models using custom internal data (e.g. Amazon Bedrock customized models, SageMaker JumpStart)

Scope 5 - Self-trained Models

Designing, training, and hosting proprietary models from scratch (e.g. Custom training on Amazon SageMaker)

MLOps

An extension of DevOps practices to machine learning that automates model deployment, version control, testing, continuous integration, continuous retraining, and production monitoring.

Precision (ML Metric)

A performance evaluation metric defined as the ratio of true positive predictions relative to total positive predictions made by the model.

Recall (ML Metric)

A performance evaluation metric defined as the ratio of true positive predictions compared to actual positive instances present in the data.

F1-score

A model evaluation metric calculated as the average of precision and recall, serving as a balanced overall measure of model quality.

AWS Config

Tracks resource inventory and evaluates configuration compliance automatically.

AWS Artifact

Provides on-demand access to AWS compliance reports, SOC documentation, and security certifications.

AWS CloudTrail

Logs, monitors, and retains account activity and API calls across AWS infrastructure.

AWS Trusted Advisor

Evaluates environment settings against best practices for security, cost optimization, performance, fault tolerance, and service limits.

Amazon Inspector

Automated security assessment service that scans applications for vulnerabilities and exposure.

AWS Audit Manager

Continuously collects evidence to automate risk assessments and compliance audits.