Answer: B. It limits unnecessary exposure. Limiting collection to its purpose reduces unnecessary exposure.
Which example is sensitive personal data?
A A medical diagnosis
B A public slogan
C A product color
D A generic greeting
Answer: A. A medical diagnosis. Health information can cause harm if improperly disclosed.
What should a privacy notice explain clearly?
A Model weights and code
B Every training token
C Other firms' policies
D Data uses and choices
Answer: D. Data uses and choices. A useful notice explains what data is used, why, and what choices people have.
How does purpose limitation guide AI data use?
A Reuse data without review
B Keep every input forever
C Use data for stated purposes
D Share all outputs publicly
Answer: C. Use data for stated purposes. Purpose limitation ties processing to specified purposes and requires review of new uses.
Which practice best reduces prompt data exposure?
A Paste complete customer files
B Remove identifiers before submission
C Include passwords for context
D Use public chats for records
Answer: B. Remove identifiers before submission. Removing identifiers and irrelevant details reduces personal information exposed to the model.
Where should access permissions be enforced in retrieval?
A At the data source
B Only in the system prompt
C Only after generation
D Inside the user interface
Answer: A. At the data source. Check authorization during retrieval so the model never receives restricted documents.
What does encryption in transit protect against?
A All insider misuse
B Incorrect model outputs
C Every endpoint compromise
D Network interception
Answer: D. Network interception. Transport encryption helps prevent interception while data moves between systems.
Which statement distinguishes anonymization from pseudonymization?
A Pseudonyms always hide identity
B Both guarantee zero privacy risk
C Anonymized data resists reidentification
D Anonymization only encrypts files
Answer: C. Anonymized data resists reidentification. Pseudonymized records can be relinked with extra information; assess residual reidentification risk.
Why can removing names fail to anonymize data?
A Names are never personal data
B Other attributes identify people
C Models ignore indirect identifiers
D Encryption reveals identities
Answer: B. Other attributes identify people. Combinations such as location, dates, or rare attributes can identify people.
What does differential privacy add to data analysis?
A A measurable privacy bound
B Guaranteed perfect accuracy
C Automatic consent collection
D A ban on model training
Answer: A. A measurable privacy bound. Calibrated randomness bounds how much one person's data can affect results.
How does federated learning limit central data collection?
A All updates reveal nothing
B It eliminates model attacks
C Raw data becomes public
D Training data stays on devices
Answer: D. Training data stays on devices. Federated learning can keep raw examples local, though updates may still need protection.
Which question matters before training on user data?
A Is the dataset largest?
B Can names be guessed?
C Was this use disclosed?
D Are records convenient?
Answer: C. Was this use disclosed?. Check whether training use is disclosed, authorized, and consistent with commitments.
What is a prompt injection privacy risk?
A It encrypts private files
B It can expose connected data
C It verifies user identity
D It deletes model logs
Answer: B. It can expose connected data. Malicious instructions may manipulate an app into revealing information it can access.
Why should system prompts contain no passwords?
A Prompts may be exposed
B Passwords reduce model accuracy
C Prompts encrypt themselves
D Models cannot read credentials
Answer: A. Prompts may be exposed. Prompts are not secure secret storage; keep credentials in protected stores.
How should a RAG system protect restricted documents?
A Trust the model to refuse
B Hide document titles only
C Filter answers after publication
D Check access before retrieval
Answer: D. Check access before retrieval. Verify authorization before retrieval, before restricted data reaches the model.
What is a safer default for storing chat histories?
A Keep every chat indefinitely
B Store chats in public logs
C Retain only what's needed
D Share history across accounts
Answer: C. Retain only what's needed. Keep data only as long as needed for a defined purpose.
Which log practice best protects personal information?
A Record full prompts forever
B Redact unnecessary sensitive fields
C Copy logs to public storage
D Disable every audit trail
Answer: B. Redact unnecessary sensitive fields. Logs help monitoring, but sensitive fields should be minimized and access restricted.
When is data residency especially relevant?
A Location affects legal obligations
B Models need larger prompts
C Users prefer faster answers
D Text includes common words
Answer: A. Location affects legal obligations. Data location may affect legal duties, contracts, and transfer requirements.
What extra care is appropriate for children's data?
A Collect more for personalization
B Assume guardian consent always
C Publish training examples
D Use strong protections and review
Answer: D. Use strong protections and review. Assess applicable rules, minimize collection, and apply age-appropriate safeguards.
Why is health information high risk in AI workflows?
A Health data has no identifiers
B Models cannot process medical text
C Disclosure may cause serious harm
D All health data is public
Answer: C. Disclosure may cause serious harm. Health data disclosure can affect safety, dignity, employment, or services.
Which control can reduce exposure in test environments?
A Copy live customer records
B Use synthetic test data
C Reuse production credentials
D Disable access logging
Answer: B. Use synthetic test data. Synthetic data can limit use of real records, though privacy still needs assessment.
What is a model inversion attack trying to infer?
A Training data details
B The user's screen size
C The model's color palette
D A website's page count
Answer: A. Training data details. Model inversion attempts to infer training examples from model behavior or outputs.
What does membership inference attempt to determine?
A Which device made a request
B Who wrote a system prompt
C Where a server is hosted
D Whether data trained the model
Answer: D. Whether data trained the model. It tests whether a particular record appeared in model training data.
Why can seemingly harmless outputs still threaten privacy?
A Outputs never contain personal data
B Only images expose private facts
C They may reveal sensitive attributes
D Short answers cannot be traced
Answer: C. They may reveal sensitive attributes. Repeated or combined responses can reveal attributes not obvious in one answer.
What is the first priority after discovering exposed personal data?
A Delete all audit evidence
B Contain the exposure
C Publish affected records
D Wait for model retraining
Answer: B. Contain the exposure. Contain exposure promptly, preserve evidence, and follow the incident process.
Which step helps assess an AI privacy incident?
A Identify affected data and people
B Guess the cause from memory
C Erase system logs immediately
D Notify unrelated users first
Answer: A. Identify affected data and people. Determine data involved, affected people, scope, and duration.
When is a privacy impact assessment useful?
A Only after public complaints
B After deleting all records
C For every spelling correction
D Before a high-risk deployment
Answer: D. Before a high-risk deployment. Early assessment identifies privacy risks while design changes remain practical.
What does privacy by design encourage teams to do?
A Add notices after launch
B Rely only on user caution
C Build safeguards into workflows
D Collect first and assess later
Answer: C. Build safeguards into workflows. Integrate privacy safeguards into the system and processes from early planning.
Which human review is useful for sensitive AI decisions?
A Approve every harmless suggestion
B Review meaningful adverse outcomes
C Replace all access controls
D Ignore affected people's concerns
Answer: B. Review meaningful adverse outcomes. Meaningful review can catch harm and provide a challenge path.
What should be checked before sending data to an AI provider?
A Terms, safeguards, and data use
B Provider's logo and colors
C Model's public popularity
D Social media post count
Answer: A. Terms, safeguards, and data use. Review retention, training use, safeguards, contract terms, and subprocessors.
Why should telemetry collection be limited?
A Telemetry always changes answers
B Metrics cannot help reliability
C Logs are never useful
D It may capture user information
Answer: D. It may capture user information. Telemetry may capture personal data; collect necessary signals and control retention.
What does least privilege mean?
A Give all users administrator rights
B Share one account widely
C Grant only necessary access
D Disable permissions permanently
Answer: C. Grant only necessary access. Each user or component receives only access needed for its task.
Where should API keys used by an AI app be stored?
A A public prompt template
B A protected secrets manager
C A shared spreadsheet
D A source code comment
Answer: B. A protected secrets manager. A secrets manager restricts access and keeps credentials out of prompts and code.
What reduces personal data in prompts?
A Use a relevant summary
B Paste the full case file
C Include unrelated identifiers
D Attach raw database exports
Answer: A. Use a relevant summary. A focused summary provides context without disclosing entire records.
Why check sensitive output before sharing?
A AI output is always public
B Models never repeat inputs
C Review weakens encryption
D Generated text can expose data
Answer: D. Generated text can expose data. Review may catch personal or confidential content reproduced in an answer.
What does a privacy threat model help identify?
A The most popular model
B The shortest prompt wording
C Assets, threats, and controls
D A guaranteed compliance result
Answer: C. Assets, threats, and controls. Threat modeling documents data assets, attack paths, threats, and mitigations.
Which governance measure supports accountable AI data handling?
A Let every team decide alone
B Assign clear data owners
C Avoid written procedures
D Remove privacy contacts
Answer: B. Assign clear data owners. Clear owners take responsibility for decisions and safeguards through the data lifecycle.
Why are audit records useful in privacy reviews?
A They show access and changes
B They prevent every breach
C They replace user notices
D They contain no sensitive data
Answer: A. They show access and changes. Audit records help investigate access and changes, and need protection themselves.
How can AI create unequal privacy harm?
A Every group has identical context
B Privacy risks affect only developers
C Bias removes personal information
D Some groups face greater exposure
Answer: D. Some groups face greater exposure. Different groups may face unequal surveillance or consequences, so assess impacts across populations.
What does transparency about AI data use provide?
A Automatic permission for reuse
B Proof that systems are safe
C Clear information for decisions
D A substitute for security
Answer: C. Clear information for decisions. Clear information supports informed choices but does not itself guarantee safety.
How do privacy and security relate?
A They mean exactly the same
B They overlap but differ
C Privacy requires no security
D Security guarantees privacy
Answer: B. They overlap but differ. Security protects against threats; privacy governs appropriate collection, use, access, and disclosure.
When should a cross-border transfer be reviewed?
A Before transferring personal data
B After deleting the source system
C Only when data is public
D Whenever a prompt is short
Answer: A. Before transferring personal data. Review applicable law, safeguards, contracts, and destination risks before transfer.
What should an opt-out mechanism make clear?
A Erase models instantly
B Stop all data use
C A hidden settings location
D Its scope and effect
Answer: D. Its scope and effect. People need to know what opt-out changes, how to use it, and when.
Why is an access request process important?
A It guarantees perfect predictions
B It removes all model errors
C People may have data rights
D It prevents every cyberattack
Answer: C. People may have data rights. A defined process supports applicable access, correction, or deletion requests.
What follows when deletion affects model training data?
A Promise instant unlearning always
B Assess feasible deletion obligations
C Ignore the source records
D Delete all model files
Answer: B. Assess feasible deletion obligations. Identify applicable duties and feasible methods; training influence may need specialized assessment.
Which measure helps prevent accidental disclosure in generated answers?
A Apply output checks
B Trust fluent wording
C Expose hidden system prompts
D Disable user authentication
Answer: A. Apply output checks. Output checks can detect sensitive content before release and complement access controls.
What is a privacy-preserving default for transcript sharing?
A Post transcripts publicly
B Share across all workspaces
C Include private context automatically
D Keep sharing off by default
Answer: D. Keep sharing off by default. A restricted default prevents unnecessary disclosure unless sharing is authorized.
Why test privacy controls after deployment?
A Testing removes legal duties
B Users stop submitting data
C Data flows can change
D Models never receive updates
Answer: C. Data flows can change. New features and integrations can change data flows and risks.
Which outcome best indicates responsible privacy handling?
A Maximum collection achieved
B Risks assessed and mitigated
C All prompts stored forever
D Every output published
Answer: B. Risks assessed and mitigated. Responsible handling assesses risks, applies safeguards, and checks effectiveness over time.
What privacy documentation helps explain a safeguard decision?
A Rationale, owner, and review date
B The model's marketing slogan
C An unverified accuracy claim
D A list of user passwords
Answer: A. Rationale, owner, and review date. Documenting reasons, ownership, and review timing makes decisions traceable.