AI for Everyone study guide · Lesson 9 of 14

Verify and Protect

Verify Claims and Citations · Minimize Sensitive Data · Resist Injection and Impersonation · Know the High-Stakes Boundary

This is the lesson's study guide from the FlashFeed app, plus its key facts and a self-test.

What you'll be able to do

  • Verify material claims, minimize sensitive data, resist manipulation, and keep high-stakes decisions with qualified people.

Verify Claims and Citations

Verification begins by identifying what matters. Separate claims, calculations, quotations, citations, and recommendations; then choose the right check for each. Open primary or authoritative sources, compare exact language, recalculate numbers, and look for omitted qualifications.

Do not ask the same unsupported conversation to certify itself. A second prompt may catch an inconsistency, but independence comes from external evidence, tools, or qualified review. Record what was checked and what remains uncertain.

Three ideas to retain:

  • Verify the Material Claims: Focus checking effort on statements that drive a decision, create risk, or would be costly to repeat if wrong.
  • Existence and Support Are Separate: A citation can be real while failing to support the attached claim, or support only a narrower qualified version.
  • Independent Evidence Matters: Rephrasing the question to the same model can be useful, but authoritative sources, calculations, and external tests provide stronger confirmation.

Key term: Claim verification means “checking a statement against appropriate evidence.”

Minimize Sensitive Data

Before entering data, ask whether the task can be completed with less. Remove names, account numbers, health details, private messages, customer records, children's information, and confidential work unless use is authorized and protected by an appropriate service and agreement.

Review the provider's current retention, training, sharing, deletion, security, and administrator controls; these differ by product, plan, region, and settings. Use placeholders or synthetic examples for exploration. Data minimization is useful even when a system claims strong protection.

Three ideas to retain:

  • Data Minimization Reduces Exposure: Provide only the fields and detail required for the task, and remove direct or indirect identifiers when they are unnecessary.
  • Policies Depend on Product and Setting: Retention, training use, human review, sharing, deletion, and administrator controls can vary across services and plans.
  • Other People's Data Is Not Yours to Volunteer: Private messages, records, images, and voices can affect people who did not choose to use the system.

Key term: Personal data means “information linked or linkable to a person.”

Resist Injection and Impersonation

An AI system may read webpages, emails, or documents containing instructions aimed at the system rather than the user. Keep control instructions separate from retrieved content, restrict tools, and require confirmation before external actions. Do not let a document grant itself authority.

Synthetic text, images, and voices can make scams more convincing. Urgency, secrecy, payment demands, and channel switching are warning signs regardless of how familiar the voice sounds. Pause and verify through a known number, account, or person.

Three ideas to retain:

  • Retrieved Content Is Untrusted: Text found in a file or webpage can contain manipulative instructions and should not outrank the workflow's trusted controls.
  • Familiar Sound Is Not Authentication: Voice cloning can imitate a person well enough to support an emergency or payment scam.
  • Verify through a Trusted Channel: Use a known phone number, account, contact, or agreed code rather than contact information supplied in the suspicious message.

Key term: Prompt injection means “untrusted text that tries to redirect an ai workflow.”

Know the High-Stakes Boundary

Health, legal rights, safety, credit, employment, insurance, and major financial choices can cause serious harm. AI can help explain vocabulary, organize records, draft questions, or compare a document with a checklist. It should not impersonate a qualified professional or make the final decision.

Use authoritative, current, jurisdiction-specific sources and disclose uncertainty. Escalate emergencies immediately to appropriate local services. If an output affects another person's rights or access, require accountable human review and a way to challenge or correct the decision.

Three ideas to retain:

  • High Stakes Change the Standard: The greater the possible harm, the stronger the evidence, expertise, privacy, review, and accountability that are required.
  • Preparation Is a Safer Role: AI can help assemble questions, timelines, and source excerpts while a qualified person retains diagnosis, advice, or decision authority.
  • Affected People Need Recourse: When an AI-assisted decision changes access, opportunity, or rights, people need notice, human review, and a correction or appeal path.

Key term: Decision support means “information that helps a responsible person decide.”

Sources

Key terms

Corroboration
Independent evidence that supports or confirms a claim.
De-identification
Reducing the link between data and the person it describes.
Social engineering
Manipulating a person into revealing information or taking an unsafe action.
High-stakes use
An application where errors can materially affect safety, rights, health, livelihood, or access.

Key facts

Existence and Support Are Separate

A citation can be real while failing to support the attached claim, or support only a narrower qualified version.

Independent Evidence Matters
Independent Evidence MattersWikimedia Commons · Jay Walsh for the Wikimedia Foundation · CC BY-SA 4.0

Independent Evidence Matters

Rephrasing the question to the same model can be useful, but authoritative sources, calculations, and external tests provide stronger confirmation.

Policies Depend on Product and Setting

Retention, training use, human review, sharing, deletion, and administrator controls can vary across services and plans.

Other People's Data Is Not Yours to Volunteer

Private messages, records, images, and voices can affect people who did not choose to use the system.

Familiar Sound Is Not Authentication

Voice cloning can imitate a person well enough to support an emergency or payment scam.

Verify through a Trusted Channel

Use a known phone number, account, contact, or agreed code rather than contact information supplied in the suspicious message.

Preparation Is a Safer Role

AI can help assemble questions, timelines, and source excerpts while a qualified person retains diagnosis, advice, or decision authority.

Affected People Need Recourse

When an AI-assisted decision changes access, opportunity, or rights, people need notice, human review, and a correction or appeal path.

Matched pairs

Each item on the left goes with the one on the right.

Claim verification
Checking a statement against appropriate evidence
Citation alignment
Whether a source actually supports the attached claim
Personal data
Information linked or linkable to a person
Data minimization
Limiting collection and sharing to what is necessary
Prompt injection
Untrusted text that tries to redirect an AI workflow
Out-of-band check
Verification through a separate trusted channel
Urgency cue
Pressure designed to prevent careful verification
Decision support
Information that helps a responsible person decide
Decision authority
Power and accountability to make the final choice
Qualified reviewer
A person with relevant expertise and accountability

Step by step

Minimize data before using an AI tool

  1. Identify the task's essential fields
  2. Remove names and unnecessary identifiers
  3. Replace real values with placeholders when possible
  4. Confirm authorization, policy, and account settings
  5. Delete temporary inputs and outputs according to policy

De-identification can fail when remaining details still point to a person.

Self-test

Answer each question from memory, then check your work against the answer key.

  1. 1. What must be checked about a citation?

    1. AThat the URL looks long
    2. BThat the title sounds relevant
    3. CThat it supports the claim
    4. DThat the answer is confident
  1. 2. Which check is most independent?

    1. ARepeat the same prompt
    2. BAsk for more confidence
    3. CRequest a longer rationale
    4. DRecalculate from source data
  2. 3. Best way to test a sensitive workflow?

    1. AShare every identifier
    2. BUse synthetic sample data
    3. CPaste real customer records
    4. DDisable access controls
  3. 4. What should be checked before upload?

    1. ACurrent data-handling terms
    2. BOnly the interface color
    3. CThe model's tone
    4. DA marketing slogan
  4. 5. A familiar voice requests urgent money. First move?

    1. AShare an account password
    2. BTrust the caller ID
    3. CVerify through a known channel
    4. DPay before the call ends

Go deeper

Answer key

  1. 1. C That it supports the claimA source must be both genuine and aligned with the statement.
  2. 2. D Recalculate from source dataExternal inputs and arithmetic test the result beyond its wording.
  3. 3. B Use synthetic sample dataSynthetic examples reveal workflow needs without exposing real people.
  4. 4. A Current data-handling termsRetention and use depend on the actual product, plan, and settings.
  5. 5. C Verify through a known channelVoice and caller ID can be spoofed; independent contact reduces risk.

Image credits

Download all 14 lessons: PDF · EPUB