Lesson 9 of 11 in the free starter module · Progress is saved in this browser

All 11 lessons
  1. 1 What an AI chatbot actually is
  2. 2 Give it enough context to help you
  3. 3 Prompting Lab 1 — The Prompt Card
  4. 4 Prompting Lab 2 — Before and after
  5. 5 Improve the answer instead of starting over
  6. 6 Prompting Lab 3 — Fix-this-prompt drills
  7. 7 Prompting Lab 4 — Your personal prompt template
  8. 8 Check before you trust
  9. 9 Keep private things private
  10. 10 Run the full loop on a real task
  11. 11 Prompt Practice Lab and starter mastery check

Free starter · Lesson 9

Keep private things private

Time
about 30–35 minutes
Prompting skill
P7 Safe prompting (with P1–P4 review)
You'll need
Your free AI chatbot and a new chat. You'll practice with fictional examples. Don't paste real account numbers, passwords, or anyone's private information.

Listen Talk to your coach


1. Objective

By the end of this lesson, you will be able to:

  • Identify the details in a request that shouldn't go into a personal chatbot.
  • Rewrite a request so it still gets useful help, using a description of the situation and placeholders instead of private details.
  • Locate your chatbot's training or activity setting and say what your current choice is.

2. Hook

Listen: welcome 0:11 · AI voice: Bella
Transcript

A chatbot can only work with what you give it. That's why context helps, and why it pays to be choosy. Today you'll learn what to keep off the page, and how to still get the help you need.

You'd happily ask a stranger at a help desk, "Can you help me word a complaint about a double charge?"

You wouldn't hand that stranger your bank card, your password, and a printout of your medical history to do it.

A chatbot is more like that help desk than like a locked diary. This lesson shows you how to get the same help while handing over much less.

3. Warm-up (from lessons 1–8)

  1. (Lesson 2) Context is useful. Why might adding more detail sometimes be a bad idea? (Guess. This lesson answers it.)
  2. (Lesson 8) Name one way to make a reply easier to check.
  3. (Lesson 1) Is a chatbot a person who will keep your secrets?
  4. (Lesson 6) In the bank drill, what did the model fix replace the card number with?
Answers
  1. Some details are private, and anything you type may be stored or seen by others, depending on the product and your settings.
  2. Any of these: ask it to name the source; ask it to say if it's unsure; ask it to list the claims to verify.
  3. No. It's a program run by a company, and what you type is handled under that company's policies and your settings.
  4. A placeholder: [CARD LAST 4 DIGITS]. The draft didn't need the real number.

4. The idea

Listen: the key idea 0:44 · AI voice: Bella
Transcript

Anything you type or upload may be stored by the company that runs the tool, and depending on your settings, it may be used to improve their models or seen by a reviewer. You don't need to stop using these tools. You need a clear rule. Safe prompting means giving the chatbot the situation it needs, and leaving out the secrets it doesn't. Two moves do most of the work. Describe the kind of thing, not the thing itself. And use placeholders, like the word name or account number in square brackets, which you fill in later yourself, never in the chat. Skip passwords, ID numbers, account details, health records, and other people's private information. You can still get help on all of those topics.

A chatbot can only work with what you give it. That's why context helps (lesson 2). It's also why you need care. Anything you type or upload may be stored by the company that runs the tool, used to improve its models if your settings allow it, or in some cases seen by a human reviewer.

You don't need to stop using these tools. You need a clear personal rule:

Don't put anything in a prompt that you wouldn't be comfortable handing to a stranger at that company.

Your new prompting skill: safe prompting (P7)

Safe prompting means giving the chatbot the situation it needs, and leaving out the secrets it doesn't. Two moves do most of the work:

  1. Describe the kind of thing, not the thing itself. "A billing email that includes my account number" instead of pasting the email.
  2. Use placeholders. A placeholder is a clearly marked stand-in, like [ACCOUNT NUMBER] or [NAME], that you fill in later yourself through a trusted channel, never in the chat.

This is row 5 of your rubric. You met it briefly in lesson 3; here it gets its full treatment. It's the one row that must always score 2 before you send.

A simple "don't paste" list

Skip these (in a prompt, an upload, or a screenshot) unless you're using a workplace tool your employer has approved for that kind of data. Most personal chatbots aren't that tool.

  • Passwords, PINs, and one-time login codes
  • Government ID numbers (such as a Social Security number)
  • Bank or credit card numbers, and anything that would let someone into an account
  • Your full home address, or someone else's, when it isn't necessary
  • Health records, diagnoses, or detailed medical history
  • Other people's private information (their phone number, work situation, or family details)
  • Confidential work material, client data, or unpublished documents you're not allowed to share
  • Anything you wouldn't want stored, reused, or read by someone you've never met

You can still get help on all of those topics. Describe the kind of problem, not the secrets inside it.

Tool notes: what companies typically do with chats (checked mid-November 2026)

Privacy settings and policies change. The notes below were checked in mid-November 2026 against each company's own help pages. Before you rely on a setting, look at it in your account and read the company's current policy. If the screen labels differ from what you see here, follow the labels in front of you.

A common pattern across major consumer chatbots:

  • Your chats may be used to train (improve) future models, depending on your settings.
  • You can usually turn that off, or use a temporary / incognito / private chat mode that is not used for training.
  • Even after you opt out, feedback you choose to send (for example, a thumbs-up or thumbs-down) may still be used to improve models.
  • Companies may still keep or review chats for safety, security, or legal reasons, for a period of time.

Specific consumer settings, as of this writing:

  • ChatGPT (personal plans such as Free, Plus, or Pro): "Improve the model for everyone" is on by default. You can turn it off in Settings → Data controls. Regular chats still stay in your history. Temporary Chat is not used to improve OpenAI models while it remains temporary, and may be retained for up to 30 days for safety purposes. If you submit feedback on a reply, the related conversation may be used for training even if you opted out. OpenAI says it does not use content from ChatGPT Business, Enterprise, Edu, or Healthcare workspaces, or from its API, to train models by default. See Data controls in ChatGPT.
  • Claude (consumer Free, Pro, or Max): Anthropic uses your chats and coding sessions to improve Claude if you allow model improvement in your privacy settings (new users are asked to choose during signup). You can change this in Settings → Privacy, under "Help Improve our AI models." Turning it off means new chats are not used for future model training. Incognito chats are not used to improve Claude, even if model improvement is on. Safety-flagged conversations may still be used to enforce policies. See Is my data used for model training? and How do I change my model improvement privacy settings?.
  • Gemini Apps: If you are 18 or older, Keep Activity is on by default. When it is on, Google says it uses your Gemini Apps activity to provide, develop, and improve its services, including training generative AI models, and that a subset of chats may be reviewed by human reviewers. You can turn Keep Activity off in Settings & help → Activity. Temporary chats, and chats you have when Keep Activity is off (if you don't submit feedback), are not used to train Google's AI models; they are retained with your account for 72 hours for service and safety. See the Gemini Apps Privacy Hub and Manage & delete your activity in Gemini Apps.
  • Grok (grok.com or the Grok app): xAI says it may use your prompts and Grok's responses to train models, and that you control this. On the mobile app, go to Settings → Data Controls and turn off Improve the model. On grok.com, go to Settings → Data, then Improve the Model. After you opt out, new conversations are not used for training. Private Chat (the ghost icon) is not used for training and is deleted from xAI systems within 30 days, unless a legal, compliance, or safety reason requires keeping it longer. Voluntary feedback may still be used for training. If you use Grok on X, X's privacy policy and X's opt-out instructions apply instead. See xAI Consumer FAQs and the SpaceXSI Privacy Policy (updated 27 October 2026).

Google's help pages are explicit that some chats may be reviewed by people, and they warn you not to enter confidential information you would not want a reviewer to see. That is a good habit with any chatbot, even when a company is less explicit on the public help page.

If your work has official rules about AI tools, follow those. They come first.

5. Worked example

Before (don't do this; shown here only as an example):

Fix this complaint: "Hi, I'm Maya Chen, account 998877, 742 Maple Avenue. You charged my Visa 4111 1111 1111 1111 twice for October. Refund it now."

After:

Help me write a polite message to a utility company about a duplicate charge for October. I won't share my account details; use [NAME] and [ACCOUNT NUMBER] as placeholders. Ask for a refund of the second charge. Keep it under 120 words.

What changed:

Change Row
Name, account number, address, and card number removed; placeholders added 5. Safe to send (new)
"Duplicate charge for October… refund of the second charge" keeps the facts that matter 2. Context
"Polite… under 120 words" 3. Tone / 4. Shape and limits

The "after" prompt has everything the chatbot needs to help, and nothing a stranger could misuse. (The details in the "before" are well-known fictional examples. See Sources.)

6. Guided practice

There are three parts. Part A needs no chatbot. Part B uses a fictional prompt; please don't replace the fictional details with real ones.

Part A: Mark what should come out (paper or notes app)

Read this fictional draft. Don't paste it into a chatbot.

Hi, this is Maya Chen. My Social Security number is 078-05-1120, my Visa is 4111 1111 1111 1111, exp 09/28, CVC 123, and my password for the utility site is River$ky19. I live at 742 Maple Avenue, Springfield. I was charged twice for October. My account number is 998877. Please refund the second charge. You can reach me at 555-0142.

List every detail that should not go into a consumer chatbot. Then check the answer at the end of section 9.

Part B: Ask the chatbot to help you strip a draft

Start a new chat. This prompt describes the kinds of details in the Part A draft without pasting any of them. That's the habit this lesson teaches, and it also avoids tripping a chatbot's safety filters with credential-shaped data.

Prompt (copy and paste):

I need help wording a message about a billing problem. I will list the kinds of details that were in a draft, without pasting the details themselves.

The draft included: a person's name, a Social Security number, a full credit card number with expiration date and CVC, a website password, a home address, an account number, a phone number, and the fact that the person was charged twice for October and wants a refund of the second charge.

Tell me which of those details should not be pasted into a chatbot. Then write a short, polite message to the company that asks for a refund of the duplicate October charge. Use placeholders such as [NAME], [ACCOUNT NUMBER], and [PHONE] for anything I should fill in later. Do not invent numbers or passwords.

What to look for:

  1. A list of details to keep out of a chatbot: at least the Social Security number, the card number and codes, and the password. A careful reply will also flag the home address, account number, and phone number as things to add later, through a trusted channel with the company, not in the chat.
  2. A short message that still explains the duplicate October charge and asks for a refund, with placeholders instead of personal details.

If it doesn't go as described: read the rewrite once to make sure it didn't invent an ID number, card number, or password. If it did, don't use those invented details. Notice how much your own eye still matters.

Part C: Find your privacy setting (optional, recommended)

If you already have an account on ChatGPT, Claude, Gemini, or Grok, take two minutes to find the training or activity setting in the Tool notes above. Choose the option you actually want. You don't have to turn anything off to finish this lesson. You should know where the control is and what your current choice is.

7. Independent practice

Spot the unsafe prompt

Which prompt fails row 5 (Safe to send)?

  • A. My doctor's portal shows "LDL" and "HDL" on a cholesterol test. In general terms, what do those two terms mean? Give me three questions I could ask my doctor.
  • B. Here are my full lab results with my name, date of birth, and patient ID: [pastes everything]. What's wrong with me?
  • C. Help me write a short note asking my doctor's office to move my appointment from Tuesday to Thursday. Use [PATIENT NAME] as a placeholder.
Answer

B fails row 5: it pastes identifiers and health records, and it also asks a chatbot for a diagnosis. A describes the kind of result and asks for general meanings and questions for a professional. C uses a placeholder and shares only a scheduling fact.

Fix-it drills

Drill 1 (with a hint). Rewrite safely:

My coworker Dana Ruiz (dana.ruiz@example.com) keeps missing deadlines because of her divorce. Write an email to our manager about it.

Hint: Whose private information is this? What does the chatbot actually need to know to help you write a professional message?

Drill 2 (no hint). Rewrite safely:

My bank password is Sunflower#42 and I think someone used it. Write a message to the bank.
One strong version of each (not the only one)

Drill 1:

Help me write a short, professional email to my manager about a project that has missed two deadlines. I want to focus on the work and on a plan to get back on track, not on anyone's personal life. Use [COWORKER] as a placeholder. Under 150 words, calm and constructive.

(It also shows good judgment: a coworker's personal situation isn't yours to share, with a chatbot or in the email.)

Drill 2:

I think someone may have accessed my online bank account. Write a short, clear message to my bank asking them to secure the account and review recent activity. Use [ACCOUNT HOLDER NAME] as a placeholder. Also give me a three-item checklist of general steps people usually take when a password may be compromised.

(Never type a password into a chatbot. If you think a password has been exposed, change it directly on the official site or app, and contact the bank through its official phone number or app.)

8. Rubric self-check

You now have all six rows of the Prompt Check. Score your two drills:

Row Question 0 1 2
1. Task What do I want done and back? Neither One Both
2. Context Situation and facts? None Some Enough for a stranger
3. Reader and tone Who's it for, how should it sound? Neither One Both
4. Shape and limits Length, format, include, avoid? None Vague Specific
5. Safe to send Would I be fine with a stranger at that company reading every word? Private details included Mostly safe, one unneeded detail Only what's needed; placeholders for private parts
6. Check built in If it matters, did I ask for sources or uncertainty, and do I plan to check outside the chat? Stakes real, no check Partial Yes (or n/a for low stakes)

The one hard rule: row 5 must be 2 before you send. Everything else is "fix your lowest row next."

Where to next: If row 5 is below 2, find each detail a stranger could misuse and replace it with a description or a placeholder.

9. Retrieval check

  1. Why is "Would I hand this to a stranger at the company?" a useful test for a prompt?
  2. What's a placeholder, and when do you fill it in?
  3. You want help writing an email about a doctor's appointment. What belongs in the prompt, and what stays out?
  4. True or false: turning off model training means the company keeps nothing and no person will ever see your chat.
  5. You've opted out of training, then you tap thumbs-down on a reply. Why might that matter?
  6. (From lesson 8) True or false: asking the chatbot "Are you sure?" counts as checking.
Answers
  1. Chats may be stored, used to improve models (depending on settings), or reviewed by people for quality or safety. If you wouldn't hand the text to a stranger there, don't paste it.
  2. A clearly marked stand-in like [ACCOUNT NUMBER]. You fill it in yourself later, through a trusted channel, not in the chat.
  3. In: the kind of change you need, the tone, and non-sensitive scheduling facts. Out: medical record details, diagnoses, insurance ID numbers, and other people's health information.
  4. False. Opting out usually stops new chats being used to train future models. Companies may still keep chats for a time, use them to provide the service, or review them for safety. Temporary or private modes also have short retention periods. Read the current policy.
  5. Several companies say that when you send feedback, the related conversation may be used for training even if you opted out.
  6. False. Real checking uses a source outside the chat.

Part A, details that should come out: Social Security number; card number, expiration, and CVC; password; and, for a consumer chatbot, the full home address, account number, and phone number. The useful facts that can stay are: there was a duplicate charge for October, and the person wants a refund of the second charge. The name can go in the message you send to the company; the chatbot doesn't need a real name.

10. Recap

Listen: recap 0:30 · AI voice: Bella
Transcript

Recap. Don't paste passwords, ID numbers, financial account details, health records, other people's private information, or confidential work into a personal chatbot. Describe the kind of thing, and use placeholders for the private parts. Row five, safe to send, must score two before you send. Know your tool's privacy settings, and when in doubt, share less. Voice is data too, so the same rule applies when you talk to a coach or a chatbot.

  • Don't paste passwords, ID numbers, financial account details, health records, other people's private information, or confidential work into a personal chatbot.
  • Safe prompting: describe the kind of thing, and use placeholders for the private parts.
  • Row 5 (Safe to send) must be 2 before you send.
  • Know your tool's training and activity setting. As of mid-November 2026, ChatGPT, Claude, Gemini, and Grok each offered a way to limit training on your chats, plus a temporary or private mode.
  • Policies and screens change. When in doubt, open the company's own help page and share less.
  • If your workplace has AI rules, follow them.

You can now: get useful help on a sensitive topic without handing over sensitive details. Use it this week: before you paste anything into a chatbot (a document, an email, a screenshot), scan it for one thing you'd replace with a placeholder.

Next: lesson 10, putting the whole loop together on one everyday task.

11. Talk it through with your coach

  • Role-play: your coach plays a friend who's about to paste a whole bank statement into a chatbot to "make a budget." You talk them through what to leave out and how to ask instead.
  • Privacy rule: your coach reminds you at the start not to say any real account numbers, passwords, or health details out loud. Voice is data too.
  • Text-only alternative: Part A and the fix-it drills above.

Talk to your coach

River is an AI voice coach, not a person. River knows this lesson and can quiz you, listen to a prompt you're working on, and give hints. River won't do the exercise for you. You can talk out loud or type.

Sources

Settings and policy language checked in mid-November 2026:

The Social Security number and card number in this lesson are well-known fictional examples (078-05-1120 was a wallet-sample number; 4111 1111 1111 1111 is a common test card pattern). The names, addresses, phone numbers, emails, and passwords are invented. Don't use real personal data in their place.