Based on the AIF-C01 exam guideReport an error

AIF-C01· AssociateWeek 5 · Day 4~37 min read

Day 4 - Data Governance and Compliance: Data Provenance and Quality, Model Governance, Auditing and Logging, Legal and Ethical Considerations of Generative AI

📌 Key Takeaways

Governance is "deciding who handles what under which rules, and proving it." The keywords are traceability and provability.
Data governance starts with provenance (lineage), quality, and classification. Data of unknown origin is a seed of copyright, bias, and regulatory violations.
Model governance manages a model's whole life through versioning, documentation, approval, monitoring, and access control (Model Registry, Model Cards).
CloudTrail = who did what (auditing), CloudWatch = how the system behaves (metrics and alarms), AWS Config = whether configuration complies with rules (configuration compliance).
Generative AI adds new legal and ethical issues: copyright, hallucination, bias, privacy, deepfakes, and transparency disclosure.

Data Governance: Starting with Provenance and Quality

Yesterday we looked at the technical defenses of AI security (IAM, encryption, PII, PrivateLink). Today is the organizational management layered on top of them. If security is "building walls," governance is setting rules and proving they were followed. Because AI is made from data, the starting point of governance is data too.

Data Lineage/Provenance: Tracking where the data came from and what transformations it went through. You should be able to answer "what data was this model trained on?" Data of unclear origin becomes a seed of copyright, bias, and regulatory violations.
Data Quality: Accuracy, completeness, consistency, and timeliness. "Garbage In, Garbage Out" is an iron rule of ML. Low-quality data produces biased and inaccurate models.
Data Classification: Classifying data by sensitivity (public/internal/confidential/PII) and applying different protection and access rules for each grade.

Data governance is directly tied to responsible AI. If you train on data of unknown origin, you will not notice bias (a fairness violation) or copyright infringement. If quality is low, model reliability drops. In other words, good governance creates "a state in which, when a problem arises later, you can trace the cause and be accountable."

The exam guide gives three examples of concepts for revealing provenance and recording the origin of data.

Concept	Question it answers	Analogy
Data lineage	Where did this data come from, and what transformations did it go through?	The distribution history of a food ingredient
Data cataloging	What data do we have, who owns it, and where is it?	A library's collection catalog
SageMaker Model Cards	What data was this model trained on, and what was it built to be used for?	A product manual

In generative AI, source citation means showing together which documents a response was based on. If you attach the source documents to a RAG response, users can verify its truthfulness themselves.

Six Data Lifecycle Strategies

The exam guide lists lifecycle, logging, residency, monitoring, observation, and retention as examples of data governance strategies.

Strategy	What it decides	Example
Data lifecycle	Rules for each stage: collect → store → use → archive → dispose	When and how to dispose of the raw data used for training
Logging	Records of who accessed data and models and when	Keep records of model invocations and data access
Data residency	The region where the data is physically located	Customer data is not taken outside the designated Region
Monitoring	Continuously checking for rule violations and anomalies	Detect a bucket that has become public
Observation	Watching how the system is actually used	Find out which questions come in most often
Retention	How long to keep it and when to delete it	Audit logs are kept for a set period

Residency and retention are often mixed up. Residency is where to put it; retention is how long to keep it.
As best practices for data engineering from the security angle, the exam guide lists assessing data quality, applying privacy-enhancing technologies, controlling data access, and data integrity.

Model Governance: Managing the Model's Whole Life

Model governance is putting control, recording, and accountability structures in place across the model's entire lifecycle (development → validation → deployment → operation → retirement).

Element	Meaning
Version control	Record which model version was deployed and when
Model documentation	Record training data, performance, limits, and intended use (model cards)
Approval procedure	A review and approval gate before deployment
Monitoring	Monitor performance and bias drift after deployment (Model Monitor)
Access control	Who can modify and deploy the model (IAM)

On AWS, you manage model versions and approval status with SageMaker Model Registry, and document model information with SageMaker Model Cards.
Model Cards are integrated with Model Registry, so you can attach audit information to a registered model version.
Registry answers "which version was approved and deployed," and Cards answer "what is this model and how was it evaluated."

🔍 Going deeper: Model governance connects to the tools you learned on Day 2. Measuring bias with Clarify, monitoring drift in operation with Model Monitor, and tracking versions with Model Registry are all governance in practice. "Governance" is not an abstract concept; it is weaving these tools into procedures.

Auditing and Logging: Compliance Means Being Able to Prove

The core of compliance is "can we prove that we followed the rules?" Audit logs are the foundation for that.

AWS CloudTrail: Records every API call that happens in an AWS account. It keeps "who did what, to which resource, and when." You can trace who deployed a model and who accessed the data.
Amazon CloudWatch: Operational monitoring centered on metrics, logs, and alarms. It detects and notifies you of performance anomalies and threshold breaches.
AWS Config: Tracks changes in resource configuration and evaluates whether they comply with compliance rules.

Tool	Main role	Core question
CloudTrail	API call audit log	"Who did what?"
CloudWatch	Metrics, logs, alarms	"Is it running well right now?"
AWS Config	Configuration change tracking and compliance evaluation	"Does the configuration match the rules?"

⚠️ Trap: It is easy to confuse CloudTrail and CloudWatch. CloudTrail = "who did what" (auditing and tracing), CloudWatch = "how the system behaves" (performance and metrics monitoring). If the exam says "API call history is needed for a compliance audit," CloudTrail is the answer.

CloudTrail records fall into two kinds. Management events, which create and change resources, are recorded by default, while the large volume of data events, which handle data inside resources, are recorded only if you turn them on separately. For an introductory exam, it is enough to be certain that "audit record of API activity = CloudTrail."
Compliance frameworks (GDPR, HIPAA, SOC, ISO, and others) differ by industry and region, and AWS provides certifications and tools that support them. At the AIF level, it is enough to understand that "auditability and traceability are the core of compliance."

"What Gets Recorded" in a Generative AI Call

When you use Bedrock, records are split into two layers. The names are similar, so they are easily mixed up.

Category	CloudTrail	Bedrock model invocation logging
What is kept	Who called which model, when, and from where	The full content of requests and responses (prompts and responses) and metadata
Default state	On once you create the account	Off by default. You must turn it on yourself
Storage location	CloudTrail event history; S3 if you create a trail	CloudWatch Logs, Amazon S3
Question it answers	"Who called it?"	"What was asked and what went out?"

A requirement such as "we must retain the prompt and response content for audit purposes" is not met by CloudTrail alone. You must turn on model invocation logging.
Invocation logs contain exactly what the user typed. Because personal information may be in there, access control and encryption must also be applied to the log storage.

Five AWS Services That Help with Compliance

As services that help with governance and compliance, the exam guide lists AWS Config, Amazon Inspector, AWS Artifact, AWS CloudTrail, and AWS Trusted Advisor.

Service	One-line role	Question it answers	Clue in the passage
AWS CloudTrail	Records API activity	Who did what?	"Audit," "who changed it"
AWS Config	Change history of resource configuration and compliance rule evaluation	Does the configuration match the rules, and when did it change?	"Track configuration changes," "resources violating rules"
Amazon Inspector	Automatically scans workloads for software vulnerabilities	Are there known vulnerabilities?	"Vulnerability scan"
AWS Artifact	Download AWS's security and compliance documents	Where do I get the certification reports AWS has received?	"SOC report," "ISO certificate," "submit to the auditor"
AWS Trusted Advisor	Checks the account and recommends best practices	Are there settings that deviate from best practices?	"Recommendations," "best practice check"

AWS Artifact is where you download AWS's compliance documents, such as ISO, PCI, and SOC reports, when needed. It also has a function for reviewing, accepting, and managing agreements with AWS. The documents and agreements are provided at no cost.
What Artifact proves is the compliance of AWS's infrastructure and services. A document proving our company's compliance must be created or obtained by the customer. The shared responsibility model applies to compliance as well.
Config and CloudTrail have different viewpoints. CloudTrail looks at actions ("someone changed the bucket settings"), while Config looks at state ("this bucket now has encryption turned off and violates a rule").

What Turns Governance into Procedure

Tools alone do not make governance. The exam guide gives the following as examples of procedures that follow governance protocols.

Element	Content
Policies	Document which data and models can be used for which purposes
Review cadence	Decide how often to look at models and policies again
Review strategies	Decide what is reviewed, by whom, and by what criteria
Governance framework	Use a framework that classifies risk systematically (for example, the Generative AI Security Scoping Matrix)
Transparency standards	Set criteria for what to disclose and notify
Team training requirements	Train members so they know the policies and risks

The Generative AI Security Scoping Matrix is a classification framework presented by AWS. It divides the ways of using generative AI into five scopes according to the degree of ownership over the model and data.

Scope	Name	What it is	Example
1	Consumer app	Use a public third-party generative AI service as is	An employee brainstorms ideas with a public chatbot
2	Enterprise app	A third-party enterprise application with built-in generative AI features	A scheduling app that writes meeting agendas
3	Pre-trained models	Call an existing FM through an API and build your own application	A support chatbot built with a Bedrock FM and RAG
4	Fine-tuned models	Refine an existing FM with our data	A summarization model tuned to medical terminology
5	Self-trained models	Train from scratch on our data	A purpose-built model made in-house

As the number goes up, ownership and control over the model and data grow, and so does the security responsibility the customer must bear.
For each scope, this framework prompts you to think about what to take care of across five aspects: governance and compliance, legal and privacy, risk management, controls, and resilience.

Legal and Ethical Considerations of Generative AI

Generative AI creates new legal and ethical issues that did not exist in traditional ML. These are what the exam covers.

Copyright and intellectual property: Copyrighted works may be included in the training data, and the copyright ownership of the generated output is also unclear.
Misinformation from hallucination: It can generate plausible falsehoods as fact, causing legal liability problems.
Bias and discrimination: Generated results can be unfair or harmful to particular groups.
Privacy: Personal information in the training data can leak into the output.
Deepfakes and misuse: Fake images, audio, and video can be abused for fraud and defamation.
Transparency obligations: Regulations requiring users to be told that content is "AI-generated" are increasing.

Pairing the legal risks of generative AI named in the exam guide with responses gives the following.

Legal risk	How it shows up	How to reduce it
Intellectual property infringement claims	The output is too similar to an existing work	Check the source and license of training and reference data
Biased model output	Results that disadvantage or insult particular groups	Bias measurement, per-group evaluation, human review
Loss of customer trust	Wrong or inappropriate answers become public	Guardrails, pre-release evaluation, disclosure of AI use
End-user risk	Users believe wrong advice and act on it	Set denied topics for high-risk areas, human review
Hallucination	Invents facts and sources that do not exist	Provide grounding with RAG, validate output

As ways to detect and reduce hallucination, the exam guide lists RAG grounding, output validation, and confidence scoring. You supply source documents (RAG), check that the response matches the source (validation), and handle low-confidence responses separately.

As a hypothetical case, a media company decided to write article drafts with generative AI. In terms of governance, it ① clarified the provenance of training and reference data to check copyright risk, ② to prevent hallucination, grounded facts in internal verified material with RAG and had a person do the final review (HITL), ③ kept transparency by stating "written with AI assistance," and ④ logged all generation and publishing work with CloudTrail to secure after-the-fact auditability. It is a case of building legal, ethical, and traceability structures together with the adoption of the technology.

Quick Review of Governance and Compliance

Data governance: Provenance (lineage), quality, classification. "Garbage in, garbage out."
Model governance: Versioning, documentation, approval, monitoring, access control. Model Registry/Cards.
Auditing and logging: CloudTrail (who did what), CloudWatch (how it behaves), Config (configuration compliance).
Legal and ethical issues of generative AI: Copyright, hallucination, bias, privacy, deepfakes, transparency disclosure.

Clue in the passage	Answer
"Trace what data this model was trained on"	Data lineage
"Find which datasets exist in the company and identify their owners"	Data catalog
"Customer data must not leave a particular country"	Data residency
"Keep logs for a set period and then delete them"	Data retention policy
"Check who deleted the endpoint last month"	AWS CloudTrail
"Continuously evaluate resources with encryption turned off against rules"	AWS Config
"The auditor requests AWS's SOC report"	AWS Artifact
"Keep even the prompt and response content for audit"	Bedrock model invocation logging
"Only approved model versions are put into production before deployment"	SageMaker Model Registry

Tomorrow is a comprehensive review that ties all of Week 5 (responsible AI principles, AWS tools, security, governance) into one big picture.

📖 Glossary

Governance: A management system that decides who handles what under which rules, and makes it possible to prove the rules were followed.
Compliance: Following laws, regulations, and internal rules, and showing that you followed them.
Data lineage: The history of where data came from and what transformations it went through.
Data catalog: A collection of the lists and descriptions of the data an organization holds.
Data residency: Requirements about the region where data is physically stored.
Retention: Rules about how long to keep data and when to dispose of it.
Audit log: Records kept so that who did what and when can be checked later.
Model invocation logging: A Bedrock feature that records request and response content and metadata in CloudWatch Logs or S3.
Deepfake: Fake images, audio, and video made with AI.
Generative AI Security Scoping Matrix: AWS's classification framework that divides the ways of using generative AI into five scopes so you can gauge the security level needed.

📝 Practice Questions

Click a choice to reveal the answer and explanation.

Question 1

"Which data was this model trained on, and what transformations did it go through?" You must be able to track this. What data governance concept does this describe?

Question 2

For compliance audit, you need "who, when, what resource, what action" API call history. What's the most appropriate AWS service?

Question 3

The best explanation of the CloudTrail and CloudWatch difference is?

Question 4

Which is least likely a legal·ethical consideration for generative AI adoption?

Question 5

Which is least clearly a model governance practice?

Official references