First ask: does the task need this detail?
A request to make a message friendlier rarely needs the recipient’s home address. An outline for a school project does not need a child’s full name, school, or daily routine. Give an AI the shape of the problem with invented details whenever you can.
- Keep access secrets out: passwords, one-time codes, recovery codes, API keys, and private links that grant access.
- Keep people out of practice data: names, contact details, ID numbers, precise locations, faces, and voice recordings.
- Keep sensitive records out: customer lists, private conversations, financial or health records, employee information, and school records.
At work, check which tool and data use your organisation approves. The NCSC warns that sending company or customer information to unapproved AI services can reduce an organisation’s control over how that information is handled. Familiarity with a personal app is not workplace approval. [10]
Keep the useful task. Replace the private material.
Our teaching example is a reminder about a borrowed item. No real names, addresses, messages, or account details are needed.
An entire message thread
“Here is our club’s full chat history and member contact list. Write a reminder about returning a borrowed item.”
The task needs a tone and a few facts, not everyone’s conversation and contact details.
A fictional situation
Write a friendly two-sentence reminder for a fictional club member to return a borrowed board game by Friday. Leave names and contact details out. Do not invent a fee or a return location.
Read the draft yourself. Did it add a fee, change the day, or invent a location? Put any necessary real details into your final message after you finish using the AI.
Is changing the name enough?
Usually it is better to write a fresh fictional example. A real record may still identify someone through its dates, location, rare circumstances, attachments, or surrounding text. Replacing a name alone is not proof that the information is anonymous.
For a sensitive task, use an approved process or ask the responsible person for help. Do not upload the original record just to ask a chatbot whether it is safe to upload.
Check where your information goes.
Before using real work material, find the service’s privacy notice and relevant account settings. Check whether it retains conversations, uses them for model improvement, allows human review, or passes them to connected services. NCSC’s discussion of unapproved AI explains why these controls matter. [10]
- Account and plan: personal, workplace, and developer services can have different arrangements. Check the one you are actually using.
- Connected tools: examine what files, mail, microphone, or camera access you are granting. A request to explain a concept should not need your inbox.
- Sharing and deletion: check who can open a shared conversation and what deletion actually covers. Removing a visible chat is not evidence that every retained copy is gone.
The FTC’s general online-safety guidance also recommends examining app permissions and privacy policies, and warns that other people may retain copies of material you share. [12]
“Open weights” describes access to model parameters; it does not tell you where a particular app sends your prompt. Running a model on your own device is different from using one hosted on a website. Either setup needs its own data-handling check.
On Tinker & Tokens
Our live AI tools are archived while their service is offline. The lessons and browser-only activities remain available. Read the archive privacy notice for the current boundary and historical handling limits. An adult checking an example does not turn a general model into a child-safe service.