AIGP · Domain 1
AI Governance
About 25% of the exam
AI foundations vocabulary
How models learn
- Supervised
- labeled examples, spam or not
- Unsupervised
- no labels, clustering, segmentation
- Reinforcement
- reward and penalty after actions
- Deep learning
- many-layer neural networks
- Generative AI
- creates text, images, code, audio
- Transformer
- self-attention over the whole input
Model concepts
- Foundation model
- broad pre-training, many downstream tasks
- LLM
- predicts likely next tokens
- Fine-tuning
- continue training on domain data
- Instruction tuning
- trained to follow directions
- RAG
- retrieve external context at query time
- Agent
- plans, calls tools, acts
Evaluation words
- Features
- input variables such as income, age
- Held-out test set
- estimates performance on unseen data
- Cross-validation
- rotate folds for stable estimates
- Recall
- catch every true case
- Precision
- flagged cases really are cases
- Calibration
- confidence matches real probability
Model versus system
- Model: trained weights that produce outputs
- System: model plus data, interfaces, people
- Governance attaches to the system in context
- One base model, many downstream uses
- RAG keeps knowledge outside and updatable
- Fine-tuning bakes knowledge into weights
- Temperature sampling makes outputs nondeterministic
- High variance: fits training, fails unseen
Why AI needs its own governance
- Autonomy: acts without a human step
- Adaptivity: behavior shifts after deployment
- Opacity: outputs resist explanation
- Scale: one flaw touches millions
- Probabilistic outputs, not deterministic code
- Agent errors compound across steps
- Aggressive pruning hurts rare cases most
- Token costs differ by language, equity issue
Traditional software governance assumes fixed behavior; AI governance has to keep re-checking
Governance principles
- Accountability
- named humans answer, never the tool
- Transparency
- disclose use, data, limits
- Explainability
- decisions can be understood
- Fairness
- no unjustified group disadvantage
- Human agency and oversight
- intervene, override, halt
- Contestability
- a path to challenge and correct
- Safety and robustness
- reliable under stress and attack
- Precautionary principle
- hold back until safety shown
Frameworks at a glance
NIST AI RMF
- Govern: culture, policies, accountability
- Map: context, purpose, risks identified
- Measure: analyze, benchmark, monitor
- Manage: prioritize, treat, accept
- Function, category, subcategory hierarchy
- Valid and reliable is the foundation
- Voluntary, not certifiable
ISO/IEC
- 42001: certifiable AI management system
- 23894: AI risk guidance, advisory
- 38507: governing body oversight of AI
- TR 24368: ethical and societal concerns
- 42001 Annex A: reference controls, justified
- 42001 requires an AI system impact assessment
OECD and others
- OECD: five values, five policy recommendations
- EU HLEG: seven requirements, ALTAI checklist
- UNESCO: four values, non-binding
- IEEE 7000 series: values into requirements
- EDM: evaluate, direct, monitor
- Principles-to-practice gap is the critique
Roles and structures
- Board
- sets appetite, receives reports
- Executive sponsor
- authority, funding, escalation
- Governance committee
- cross-functional decision rights
- Ethics board
- value-laden questions, may block
- Program owner
- runs the AI governance program
- Model owner
- accountable for one system
- Second line
- oversight and challenge
- Internal audit
- third line, independent assurance
A quorum of three engineers is a quorum without legal, privacy or the business owner
Building the program
- Charter
- Inventory
- Policies
- Intake and tiering
- Stage gates
- Monitoring
- Reporting
- Charter fixes decision rights and authority boundaries
- Registry catches shadow AI before it spreads
- Re-tier when scope, data or context changes
- Gate holds deployment until criteria are met
- Exception: documented waiver, risk acceptance, expiry
- Dated decision records with rationale and criteria
- Regular board cadence plus event escalation
- Embed in existing GRC and ERM channels
Reuse the DPIA, security assessment and risk register; parallel structures duplicate work and split accountability
Policies and literacy
- No confidential data in unapproved public tools
- Enterprise or no-training vendor tiers
- Role-based literacy, executives to engineers
- Completion rate measures attendance, not behavior
- Scoring rubrics with calibration reviewers
- Published weighted criteria resist gaming
- Principle-based guidance adapts to novel cases
Third-party AI
- Due diligence before procurement
- Contract: audit rights, change notice, bias evidence
- Model documentation and evaluation results
- No provenance: attestations and independent tests
- Availability SLA says nothing about accuracy
- Pin model versions, revalidate on change
- Map the supply chain beyond one tier
- Shared responsibility written down
Data governance for AI
- Data owner
- authorizes use of a domain
- Data steward
- quality and definitions day to day
- Catalog
- inventory with lineage and quality
- Business glossary
- one meaning per data element
- Provenance
- origin, ownership, handling history
- Datasheet
- composition, collection, intended uses
- Feature store
- shared definitions, point-in-time retrieval
- Training-serving skew
- features computed differently in production
Training data rules
- Assess quality, bias, consent, provenance, representativeness
- Strictly separate train, validation, test
- Best-fold reporting is optimistic bias
- Undocumented vendor data is a material risk
- Track lawful basis when sources merge
- Copyleft licenses conflict with proprietary release
- Feature relevance analysis supports minimization
- Special category only to detect bias
- Label drift needs versioned labeling guides
- Immutable dataset versions bound to documentation
Documentation artifacts
- Model card
- intended use, disaggregated performance, limits
- Datasheet
- how the data was gathered
- System card
- whole system, safeguards, evaluations
- AI inventory
- every system, owner, tier
- Decision log
- who approved what, why
- Impact assessment
- harms, affected groups, mitigations
- Instructions for use
- what deployers need to comply
Rapid recall: which framework
- Certifiable management system
- ISO/IEC 42001
- AI risk guidance, advisory
- ISO/IEC 23894
- Board-level oversight
- ISO/IEC 38507
- Four functions, voluntary
- NIST AI RMF
- Generative AI companion
- NIST AI 600-1
- Self-assessment checklist
- ALTAI from the EU HLEG
- Binding treaty
- Council of Europe convention
- Adversary tactics catalog
- MITRE ATLAS
Reference strip: foundations, frameworks, program, data, documents
Foundations
- Supervised, unsupervised, reinforcement
- Foundation model, fine-tune, RAG
- Model is not the system
- Autonomy, adaptivity, opacity, scale
Frameworks
- NIST: govern, map, measure, manage
- 42001 certifiable, 23894 advisory
- OECD five principles
- HLEG seven requirements, ALTAI
Program
- Charter with decision rights
- Inventory, intake, tiering, re-tiering
- Stage gates and exceptions
- Cadence reporting to the board
Data
- Owner authorizes, steward maintains
- Provenance and lineage
- Separate train, validation, test
- Datasheets tied to versions
Documents
- Model card, system card
- Decision log with rationale
- Impact assessment on file
- Instructions for deployers
Quick exam traps
- Trap: ISO/IEC 42001 and ISO/IEC 23894 are both certifiable
- Trap: An AI model and an AI system are the same thing for governance
- Trap: A high training completion rate proves AI literacy improved
- Trap: A 99.9% availability SLA covers model quality
- Trap: Govern is one of four equal steps done in sequence
- Trap: A numeric quorum guarantees a sound committee decision
- Trap: Accountability can be assigned to the AI system itself
- Trap: A system keeps the tier it received at intake
cybercertprep.com · original revision sheet written from the public body of knowledge