Learning path · Level 1 · Tutorial 4

Applying Basic AI Safety Checks

Apply a basic safety checklist before sharing information with AI and before using its output, without needing a paid tool or sharing real sensitive data.

  • Level: Beginner
  • Duration: 12 min
  • 6 steps
Start tutorial Back to the path
Applying Basic AI Safety Checks

What you will go through

  1. Spot Sensitive Information Identify personal, confidential, proprietary, and combined details that require protection before they are shared with an AI tool. 2 min
  2. Share Less—or Withhold Reduce a fictional task input to only what a low-risk request needs, use clearly invented details for generic examples, and withhold sensitive material when safe sharing is uncertain. 2 min
  3. Check Permission for the Intended Use Decide whether fictional material is authorized for an intended AI task, and choose a safe alternative when permission is unclear. 2 min
  4. Verify Claims Before Acting Check consequential AI claims against suitable evidence before using them, and pause when support is missing. 2 min
  5. Keep a Human Responsible Use `Before input` and `Before action` to keep a named human responsible for an AI-assisted task. 2 min
  6. Apply the Full Safety Checklist Apply the two safety checkpoints to a complete fictional community-event task and explain what may be shared and acted on. 3 min

What you will learn

  • Identify sensitive information in a fictional input and decide what to remove, replace with invented details, or withhold entirely.
  • Determine whether proposed input material is authorized for the intended use and choose an alternative when permission is unclear.
  • Identify consequential claims that require checking against a trusted original source or with a qualified professional before acting.
  • Apply a before-input and before-action checklist, identifying who remains responsible for the final decision.

Before you start

  • Deciding When AI Should Assist

Step 1 of 6

Spot Sensitive Information

Identify personal, confidential, proprietary, and combined details that require protection before they are shared with an AI tool.

Start with the two checkpoints

Two safety checkpoints

Before using an AI tool, pause at two checkpoints:

  • Before input: look at what you are about to share. Identify details that need protection.
  • Before action: look at the AI output before anyone relies on it or takes a real-world action.

This step focuses on Before input. AI can help with a draft, but it does not take responsibility for information you share or decisions made from its output.

A two-stage flow shows a document checked before entering an AI tool, then an AI draft checked by a human before a decision.

The first gate represents Before input; the second represents Before action.

Recognize direct and combined details

What can need protection?

Sensitive information is not limited to one category.

  • Personal information can include contact details and health information.
  • Confidential information can include nonpublic budgets, customer records, and internal plans.
  • Proprietary information can include business plans or other organization-owned material that is not meant for public sharing.

These concerns can overlap. For example, a customer record may be both personal and confidential. Also, ordinary details can become identifying when combined: a neighborhood, a volunteer role, and a precise schedule may point to one person even without a name.

Example

Read the notes closely

Fictional input

“Create a public announcement for Cedar Community Center’s Community Workshop. The event takes place on 14 September 2026, starts at 10:00 a.m., and costs $10 USD per person.

Background notes also mention Alex’s contact details, a participant’s allergy note, a nonpublic draft budget, and an early partnership plan. Invented example: Sample Participant is described by a small neighborhood, a volunteer role, and an exact arrival pattern.”

The event date, time, and fee are public event facts in this fictional task. The background notes contain details that need protection before input. The last combination is important: each detail may seem ordinary alone, but together it can identify Sample Participant.

Practice: separate public facts from protected details

Check your recognition

All notes below are fictional. Which assessment most accurately distinguishes details that need protection from explicitly public event facts?

Step 2 of 6

Share Less—or Withhold

Reduce a fictional task input to only what a low-risk request needs, use clearly invented details for generic examples, and withhold sensitive material when safe sharing is uncertain.

Choose the smallest useful input

Start with the minimum needed

At the Before input checkpoint, begin with the task—not with the document you happen to have. Ask: What is the smallest set of facts needed to complete this request?

For a public event announcement, the needed facts may be the event name, date, time, and public fee. A full registration record, contact detail, or private participant note is not needed just because it appears in the same document.

Use these choices:

  • Remove unnecessary details.
  • Replace with invented details when a generic practice example needs a personal detail.
  • Withhold information when the task would require sensitive material and you are uncertain about sharing it safely.

Reduce event notes before sharing

A crowded event note is filtered into a short announcement containing only event facts.

Keep the facts required for the requested public announcement; do not copy every detail from the source notes.

Remove, replace, or withhold

Example

Worked reduction

Fictional input

Cedar Community Center is planning a Community Workshop. It takes place on 14 September 2026, starts at 10:00 a.m., and costs $10 per person (USD). Alex is the event coordinator. Contact: [contact removed]. Notes also contain participant attendance and dietary information.

Reduced input for an AI draft

“Draft a short public announcement for Cedar Community Center’s Community Workshop. State that it is on 14 September 2026, begins at 10:00 a.m., and costs $10 per person.”

The reduced input supports the drafting task without including a contact detail or participant records.

Tip

Replace only when a generic detail helps

Sometimes a generic example needs a person-like detail, but no real person is needed. Make that explicit:

Invented example: “Write a friendly sample reminder addressed to Sample Participant.”

The label matters: it tells you that the detail was created for practice. Simply changing a real person’s name is not enough. Other details in a record can still identify them or expose private information.

If a request requires a private record and you are uncertain whether the tool adequately protects that sharing, choose Withhold. Do not try to solve that problem by changing only the name.

Practice: select only what is needed

Choose the reduced input

Which input is sufficient for asking an AI tool to draft a public announcement?

Practice: revise and decide

Apply the three input choices

Write a safe response to this situation:

You need an AI tool to draft a public announcement for Cedar Community Center’s Community Workshop. Use the event facts provided earlier. The draft also needs a generic attendee example, but no real participant information is needed.

Separately, someone asks you to have the tool summarize a participant’s private health note. You are uncertain whether sharing that note is adequately protected.

In your response:

  1. Write the reduced prompt for the announcement.
  2. Include a clearly labeled invented attendee detail and explain why it is invented.
  3. State the input-handling label you would use for the health note and why.

Write at least 180 characters (0/180).

Step 3 of 6

Check Permission for the Intended Use

Decide whether fictional material is authorized for an intended AI task, and choose a safe alternative when permission is unclear.

Check the intended use

Access is not authorization

At the Before input checkpoint, ask two permission questions:

  1. Is this material authorized for the task I want to do?
  2. Is it authorized to share with this selected AI tool?

Being able to read, download, or receive material does not automatically answer yes. Access may allow you to view a record, while uploading it to an AI service or reusing it for a new purpose may require separate approval.

This is separate from privacy. Removing a person’s name can reduce personal-data exposure, but it does not create permission to use the material.

A diagram contrasts being able to view a document with needing a separate approval before sharing it with an AI tool.

Access to material and authorization to share or reuse it are different checks.

Compare permission conditions

When permission is unclear

Use the stated condition—not an assumption—to decide what to do. Public availability is not unrestricted permission, and a record can remain restricted even after names are removed.

If permission is unclear, use Withhold for the material. Then choose a safe path: ask the appropriate owner, use material explicitly authorized for this task and tool, or create a fictional example. This is a practical safety decision, not legal advice.

Three fictional materials

Compare the material, its stated condition, and the appropriate handling at Before input.

Permission conditions guide the decision.
MaterialStated conditionAppropriate handling
Cedar Community Center event briefExplicitly approved for an AI-assisted public announcementUse only for that stated task and tool
Restricted participant recordStaff may view it; no approval to upload or reuse itWithhold and ask the appropriate owner
Public third-party flyerVisible online; no AI-use permission is statedWithhold; ask or use a fictional example

Practice: permission decisions

Choose the authorized path

Match each fictional material to the safest handling decision at Before input.

Tap an item and then its matching pair.

Step 4 of 6

Verify Claims Before Acting

Check consequential AI claims against suitable evidence before using them, and pause when support is missing.

Recognize when an AI claim has consequences

Before action: check consequential claims

Claims that need a check

At Before action, pause before using an AI draft to make a decision, publish a promise, or give advice. A claim is consequential when an error could affect health, safety, money, rights, or an important commitment.

AI can produce confident, useful-looking wording, but confidence is not evidence. An AI-provided citation—or another AI giving the same answer—is also not verification.

For a consequential claim, compare the claim with a relevant, current trusted original source. When the claim depends on professional judgment, seek confirmation from a qualified professional before acting.

Verification is a comparison

A diagram showing an AI draft checked against an original source or qualified professional, with a pause when support is missing.

A consequential claim needs support from the right kind of evidence before it is used.

Match the claim to its check

Choose the appropriate check

Match each situation with the best next move.

Tap an item and then its matching pair.

Compare the draft with the evidence

Example

Source excerpt (fictional)

Use the practice materials below for the next activity. The Source excerpt (fictional) is invented for this tutorial. It helps you practice comparison, but it is not a real-world authority. In a real task, find the relevant, current trusted original source.

Source excerpt (fictional)

Cedar Community Center — Community Workshop

The Community Workshop takes place on 14 September 2026, starts at 10:00 a.m., and costs $10 USD per person.

This excerpt contains no food-safety or medical assurance.

Example

AI draft

AI draft

“Join Cedar Community Center's Community Workshop on 14 September 2026 at 10:00 a.m. Admission is $15 per person. All snacks are safe for every dietary need.”

Compare each claim with what the source excerpt actually says. Notice both direct disagreement and claims the source does not address.

Apply the comparison

Make the before-action decision

Using the AI draft and Source excerpt (fictional) from the previous screen, identify the fee discrepancy, name the claim that needs separate confirmation, and explain why this excerpt cannot serve as real-world authority.

Write at least 120 characters (0/120).

Step 5 of 6

Keep a Human Responsible

Use Before input and Before action to keep a named human responsible for an AI-assisted task.

See the two-checkpoint flow

Two checkpoints, one human decision

AI can help produce a draft, but it does not take responsibility for sharing information or acting on claims. Use two checkpoints in every AI-assisted task:

  • Before input: check whether the material is necessary, sensitive, authorized for this use, and adequately protected for the proposed sharing.
  • Before action: check consequential claims and make sure a human review owner is ready to approve, revise, or pause.

The flow is not “ask AI, then act.” It is “check, draft, check again, then decide.”

The safety flow

A four-stage flow from selected protected notes, to an AI draft, to human review against a source, to either approval or pause.

Use Before input before sharing material. Use Before action after the AI draft and before any decision or publication.

Tip

Keep the decision human

A completed checklist is not a guarantee. If a concern remains unresolved, revise the input, seek permission or evidence, or pause the related action.

Give review a real owner

Name a review owner

A review owner is a named person who can check the relevant content, seek qualified help when needed, and approve or withhold action. It is not merely someone who receives the AI draft.

For the Community Workshop, Alex is the event coordinator and can be the review owner for the announcement. Alex can check the date, time, and fee against a trusted original source. But Alex must not treat an AI statement as established just because it sounds confident.

For example, the authorized event brief supports 14 September 2026, 10:00 a.m., and a fee of $10 USD per person. It does not establish a medical or food-safety assurance. If an AI draft adds such an assurance, Alex must seek appropriate qualified confirmation, remove the claim, or pause approval.

What each checkpoint decides

Use this compact decision guide whenever an AI-assisted task moves from input to action.

A concern at either checkpoint can stop or change the task.
CheckpointCheckIf unresolved
Before inputNecessary, sensitive, authorized, adequately protectedRevise the input, seek permission, or Withhold material
Before actionConsequential claims verified; review owner identifiedSeek evidence or qualified help, revise the draft, or pause action

Check before approving

Practice: handle an unsupported claim

AI draft: “The Community Workshop is safe for every dietary need.”

Source excerpt (fictional): “Community Workshop — 14 September 2026, 10:00 a.m.; $10 USD per person.”

What should happen before the announcement is approved?

Apply the full responsibility check

Apply both checkpoints

Choose the safe sequence for Alex’s AI-assisted event announcement. The AI draft includes an unchecked claim that snacks are safe for all allergies, and the Source excerpt (fictional) does not support that claim.

Step 6 of 6

Apply the Full Safety Checklist

Apply the two safety checkpoints to a complete fictional community-event task and explain what may be shared and acted on.

Connect the checkpoints

Run the complete check

Use the two checkpoints as one connected decision.

At Before input, reduce the material to what the task needs, then check that it is authorized to share with the selected AI tool. Choose Remove, Replace with invented details, or Withhold where needed.

At Before action, check consequential claims against a trusted original source or with a qualified professional. If a claim is unsupported, pause that related action. A named review owner—not the AI—checks and approves the final decision.

Two checkpoints, one human decision

A flow diagram showing event notes filtered before input, an AI draft checked against an original brief before action, and paths to approve or pause.

First decide what may enter the tool. Then decide what evidence supports acting on its output.

Prepare the input

Case file

Review this compact case. Every detail is fictional. Your task is to draft a public announcement, not to upload records or make health assurances.

Example

Fictional input

Fictional input

Alex, the event coordinator at Cedar Community Center, wants an AI tool to draft a public announcement for the Community Workshop.

Approved event facts: the Community Workshop takes place on 14 September 2026, starts at 10:00 a.m., and costs $10 USD per person.

Alex’s direct contact is [contact removed]; it is not needed in the announcement.

A training preview may use an attendee placeholder. If one is needed, use: Invented example — Sample Participant.

A participant roster contains names, contact details, and dietary and access notes. It is not needed for the announcement.

Permission conditions

Sharing conditions for the case

Access to material does not by itself authorize sharing it with an AI tool.
MaterialConditionUse for this task
Approved event briefExplicitly approved for an AI-assisted public announcementMay be used, after reducing it to needed facts
Participant rosterRestricted to registration staffNot authorized to upload or reuse
Training placeholderAllowed only as a clearly invented exampleMay use an invented name, not a real participant

Make the safety decision

Example

Source excerpt (fictional)

Source excerpt (fictional)

“Cedar Community Center Community Workshop: 14 September 2026, beginning at 10:00 a.m. Admission is $10 USD per person.”

This brief gives event logistics and the fee. It makes no statement about food safety, allergies, or individual health needs.

Example

AI draft

AI draft

“Join Cedar Community Center’s Community Workshop on 14 September 2026 at 10:00 a.m. for $10 per person. All food is allergen-safe and approved for every health need.”

Apply the full checklist

Using the case file, write a short end-to-end safety decision.

Include all of the following:

  1. What to Remove, Replace with invented details, and Withhold.
  2. Which material is authorized as input for the announcement.
  3. Which claim can be checked against the trusted original source, and what to do about the unsupported health-related claim.
  4. The named review owner and who remains responsible for the final decision.

Write at least 220 characters (0/220).

Take the checklist forward

Summary

Reusable closing checklist

  • Before input: share only what is necessary, handle sensitive details deliberately, and confirm authorization for the intended use and tool.
  • Use Remove for unnecessary details, Replace with invented details for clearly labeled generic practice content, and Withhold material that must not be submitted.
  • Before action: check consequential claims against a relevant trusted original source or seek qualified confirmation.
  • If evidence or permission is unresolved, revise the input, seek confirmation, or pause the related action.
  • Name a review owner. The person who reviews or approves the AI-assisted result remains responsible for the final decision.

Safety checks complete

Congratulations! You completed: Applying Basic AI Safety Checks

You completed an end-to-end AI safety review: protect the input, confirm permission, verify consequential claims, and keep a human responsible.

Download the App now to have access to + 5000 free courses, exercises, certificates and lots of content without paying anything!

  • 100% free online courses from start to finish

    Thousands of online courses in video, ebooks and audiobooks.

  • More than 60 thousand free exercises

    To test your knowledge during online courses

  • Valid free Digital Certificate with QR Code

    Generated directly from your cell phone's photo gallery and sent to your email

Cursa app on the ebook screen, the video course screen and the course exercises screen, plus the course completion certificate