A note from the workshop
The chatbot asked for context. You supplied your entire life.
A useful answer rarely needs every name, account number, and private conversation attached to the problem. Start by describing the job, then decide which information is actually necessary. Today’s exercise uses a fictional document so the privacy lesson does not create its own incident.
Three useful checks
- Make a fictional sample: Replace real names, addresses, identifiers, and amounts with made-up equivalents.
- Inspect the tool: Check the provider, retention settings, training controls, and connected services.
- Send less context: Try the smallest useful excerpt and review the answer before adding anything.
Put it into practice
Write a pretend message from a fictional landlord about a repair appointment. Ask the assistant to rewrite it more clearly. The structure of that task survives without your actual address, phone number, or lease. Compare the result with your original purpose before sharing additional context.
A visible document is not necessarily the complete file: comments, other sheets, attachments, and identifying details may matter. Work on a separate copy. If you cannot tell what will be sent, use a plain-text fictional example rather than experimenting with the private original.
Privacy controls deserve inspection, but a setting about model training is not automatically a statement about retention, access, or every connected tool. Read the relevant policy for the actual account type. Follow your workplace’s approved-tool rules when its documents are involved.
The document is only one part of the handoff
Start by separating the job from the material. Rewriting a repair request needs a description of the repair, the tone, and the desired response. It does not automatically need the lease, the tenant’s signature, and last year’s payment history. The useful boundary is the minimum information required to do this particular job.
There are also separate questions about the recipient. Which company runs the tool? Is this a personal or managed account? Will another connected service receive the content? A promise about one use of data does not answer every question about storage and access. Treat an answer you cannot find as an unresolved dependency, rather than filling the gap with the word private.
A worked example: rewrite the request without exporting your life
Imagine you want a calm, clear message about a leaking kitchen tap. This is an invented example: “A tenant needs a repair appointment for a tap that has dripped for three days. Write a polite request offering Tuesday afternoon or Thursday morning. Ask for confirmation. Do not invent rights, fees, or promises.” That instruction describes the task without identifying a real household.
Check whether the draft preserves the two appointment choices and asks for confirmation. If it invents a deadline or a legal entitlement, remove it and tighten the instruction. Only after the wording is settled would you add the real recipient details yourself, outside the chatbot. The useful product is a reusable writing pattern, not a complete private record sitting in another system.
- Write the purpose in one sentence before selecting any files.
- Create a fresh plain-text fictional sample; do not edit the only copy of a real document.
- List the facts the answer must preserve and the claims it must not invent.
- Read the draft against that list, then add private details locally if appropriate.
Where a careful-looking upload can still go wrong
Replacing a name with initials is a weak change if the rest of the story identifies the person. A distinctive workplace, a dated dispute, and a specific location can do the identifying together. When those details are irrelevant, change the scenario as well as the name. If the exact private event is essential to the task, a fictional version may not be suitable; use an approved workflow or do that part yourself.
Avoid turning redaction into a new file-handling mystery. If you are unsure whether comments, extra pages, or other content remain in a document, rebuild the small example in plain text. Do not upload a real file just to ask the assistant whether the file contains anything sensitive.
What a useful result looks like
Keep the fictional prompt, your acceptance checklist, and a note about which tool and account you reviewed. A good result is a draft that does the intended job while requiring less real information. It is not a finding that the service is universally confidential. Recheck the workflow when you change the tool, enable an integration, or start handling a different category of material. The next upload deserves its own decision; yesterday’s harmless recipe is not today’s personnel file.
Know the limits
Removing obvious names does not guarantee anonymity. Unique events and combinations of details can still identify someone. Keep sensitive material out until you understand the whole workflow.
Create one reusable fictional sample for a task you often give AI.
Print the private field sheet. Record actions and observations, not passwords, recovery codes, or private documents.
Keep exploring: the offline AI experiment, lost-phone recovery, or the Recovery Drill.
About this field note
Original CRYPT1K editorial guidance. Sources reviewed October 21, 2026. Procedures are reader exercises; no unperformed hardware, account, or AI test results are claimed. Character and artwork are fictional and AI-assisted.