Is it safe to give AI your photos and personal details?

The prompt: check a file with me before I upload it

Paste it into a new chat before you attach anything, and describe the photo or document in your own words. It asks who and what is in it, sorts each detail into stop, fix first or fine, and ends with a verdict: upload it as it is, upload a safer version, or keep it out.

PromptCheck a file with me before I upload it

You are my privacy check before I upload a photo or document to an AI app. Once a file is uploaded, a copy can stay on the company's servers for weeks, and in some cases years, and depending on my settings it can be read by reviewers or used to train future models, so anything you miss here is something I cannot take back. Your job is to help me decide whether to upload this file at all, and if I do, what safer version to upload. I have not uploaded the file and will not until we finish, so you work only from what I tell you about it.

ABOUT THE FILE (I fill this in once)
- What it is: [e.g. "a selfie I took in my kitchen", "a photo of a letter from my bank", "a screenshot of a group chat", "a scan of my child's school report", "a photo of my passport page"]
- What I want the AI to do with it: [e.g. "turn it into a cartoon avatar", "explain what this letter is asking me to do", "fix the lighting", "tell me what this rash might be", "fill in a form for me"]
- Who is in it or named in it: [e.g. "just me", "me and two friends", "my 6-year-old", "a neighbour's child in the background", "no people, but names are printed on it", or "not sure"]
- The app I will upload it to, and whether I have turned off model training there: [e.g. "ChatGPT, training is off", "Gemini, not sure", "Claude, I have not checked"]
- The device the file is on: [e.g. "iPhone", "Android phone", "Windows laptop", "Mac"]
- The country I live in: [country]

BEFORE YOU START
Ask me up to eight short questions, in two rounds, only where my answer changes what you tell me. Do not ask me to type out any name, number or address; ask only what kind of detail it is ("is there an account number on it?", not "what is the account number?").
Round one, what is in the file: who is visible, including anyone in the background or in a photo on a wall or screen; any text I can read on it (letters, labels, screens, name badges, school logos, whiteboards, calendars, receipts, tickets, barcodes or QR codes); any number that identifies a person or an account (passport, driving licence, national ID or tax number, bank or card number, patient or policy number, phone number, home address, car number plate).
Round two, how and why: whether the photo was taken with a phone camera (it may store where it was taken inside the file); whether the job needs the whole file or only one part of it; and whether anyone else in it has agreed to this.
Wait for my answers. If I say "not sure" to anything, treat that detail as present.

STEP 1: SORT EVERY DETAIL
Go through each person and each detail I told you about and put it in one of three groups:
- STOP: should not go into an AI app at all for this job.
- FIX FIRST: fine once it is cropped, covered, blurred or removed.
- FINE: the job needs it, or it identifies nobody.
Then apply these rules:
- If a child who is not mine is visible or named, put them under STOP unless I tell you their parent has agreed. Offer a way to get the job done without their face, such as cropping them out or describing the scene in words.
- If my own child is in it, put them under FIX FIRST at least: crop out school names, uniforms and anything that shows where they live or go to school, and tell me to check that training is off before I upload.
- If other adults are in it and have not agreed, put their faces under FIX FIRST and suggest a crop or a blur.
- If a full identity document is in it (passport, driving licence, national ID card, bank card), put it under STOP unless the job truly needs the document itself. If the job is to understand or fill in something, suggest I cover the numbers and photo first, or type out just the sentence I need explained.
- If a card number, security code, password, PIN or a login code is visible, it is always STOP. It must be covered or cut out, whatever the job.
- If health, legal or money details are in it, keep what the job needs under FINE and put names, dates of birth, addresses and reference numbers under FIX FIRST.
- If it is a phone camera photo, put the stored location under FIX FIRST and tell me to share or save a copy without location before I upload.
- If the app itself, not a person, has asked me for an ID photo or a selfie to check my age or identity, say that this is a different case: it should only ever happen inside the app's own settings or sign-in screens, never in reply to a chat message, email, text or link asking for it. Tell me to read that app's own help page on who handles the image and how long it is kept before I go ahead. Do not guess those details.
- If I have not turned off training, or I am not sure, add it as the first item under FIX FIRST.

STEP 2: THE SAFER VERSION
For every FIX FIRST item, give one plain instruction I can do on my device: crop, cover the detail with a sticky note or a piece of paper before I take the photo, blur or draw over it with the phone's built-in markup tool, take a new photo of only the part I need, or type the text out instead of uploading it. Pick the quickest fix for each item, not the most thorough one. If you are not sure of the exact menu or button name on my device, say "look for" and describe it, and mark it [CHECK ON YOUR DEVICE]. Never present a menu path you are unsure of as fact.

STEP 3: THE VERDICT
Give one verdict, with one line of why:
- UPLOAD AS IS: nothing under STOP or FIX FIRST.
- UPLOAD THE SAFER VERSION: fixes listed, nothing left under STOP.
- DO NOT UPLOAD: something under STOP cannot be fixed. Then give me the best way to get the job done without the file, such as describing it in words or asking a person.

WRITE IT IN THIS SHAPE
WHAT I LOOKED AT: [one line summing up the file and the job]
STOP: [each item and why, or "nothing"]
FIX FIRST: [each item, why, and the fix]
FINE: [each item]
VERDICT: [one of the three, with one line of why]
IF YOU UPLOAD: [what to do after, such as deleting the chat and the file when the job is done, or using a temporary or incognito chat so there is less to clean up, and not rating the reply with thumbs up or down, in plain words]

RULES FOR YOU
- Work only from what I told you. If I did not mention something, do not assume it is there, except where I said "not sure".
- Never ask me for the actual number, name or address, and never repeat one back if I type it by mistake; tell me to delete it from my message.
- Do not state the law where I live as fact. If consent or privacy law might matter, say "this may be covered by privacy law where you live" and leave it there.
- Keep it short. A normal photo should need no more than about fifteen lines.

CHECK YOUR ANSWER BEFORE YOU SHOW ME
1. Every person and every detail I mentioned appears in exactly one group.
2. Nothing appears that I did not tell you about.
3. The verdict matches the groups: no UPLOAD AS IS with anything under STOP or FIX FIRST.
4. No menu path is stated as fact unless you are sure of it.
5. Name the one thing you are least sure about, and why.

What it catches in an ordinary photo (an illustration)

The file: a kitchen selfie you want turned into a cartoon avatar, taken on your phone, going into an app where you have not checked the training setting.

Stop: nothing.

Fix first: the training setting; the background, where post, screens, a house number or a school logo tend to hide, such as the newsletter on the fridge with your child's name on it, so crop to head and shoulders; the location saved in the photo, so remove it before you attach.

Fine: your face, since the avatar needs it.

Verdict: upload the safer version.

What to upload, fix or keep out

Other people, especially children

Ask first: Google's own Gemini privacy page tells users to get permission before using other people's faces or voices in Gemini.
For a child: that decision belongs to the parent, so a friend's child, a niece or a pupil stays out unless the parent has said yes.
Fix first: crop them out, cover their face with your phone's markup tool, or describe the scene in words instead.
Verdict: no photo of someone else's child without the parent's OK.

IDs, letters and anything with a number on it

What the AI needs: to explain a bank letter or a form, it needs the wording, and your account number, address and reference number add nothing.
Fix first: cover the numbers with a sticky note before you take the photo, or type out only the part you want explained.
Always keep out: card numbers and security codes, passwords, PINs and login codes.
Verdict: keep full ID documents out of ordinary chats.

When ChatGPT asks for a selfie or your ID

Why it asks: if ChatGPT puts an adult's account into its teen version by mistake, you can prove your age on the web under Settings, Account, Verify age.
Who sees the image: a company called Persona runs the check and, depending on your country, asks for a live selfie, a government ID or both. OpenAI's page says Persona deletes them within 7 days of the check, and OpenAI only receives your date of birth or age result.
Cover what it does not need: the same page says your ID numbers do not have to show, so cover them before you upload.
Verdict: only inside ChatGPT itself, from Settings, Account, Verify age or a notice in the app, never in reply to a chat message, email, text or link asking for it.

Turn off training and delete what you uploaded

As of 25 September 2026, these are the switches that decide what happens to an upload in each app.

ChatGPT

Training: your chats are used unless you opt out, and that covers images and files as well as what you type. Go to Settings, Data controls, and turn off Improve the model for everyone; on a phone, open the sidebar and tap your profile icon to reach Settings.
Private option: a Temporary chat is not used for training and is kept for up to 30 days for safety, and the files you add to it stay out of your Library, as long as you do not save the chat.
Delete: open the chat's ••• menu and tap Delete; OpenAI removes it from its systems within 30 days, unless it must keep it for security or legal reasons.
The catch: deleting a chat leaves any copy of the file saved in your Library, so open Library from the sidebar, select the file and delete it too.

Claude

Training: when the model improvement setting is on, or when a chat is flagged for a safety review, and then it covers the whole chat, including anything you upload. Go to Settings, Privacy, and switch off Help improve our AI models; the steps are the same on a phone.
Private option: an incognito chat is not used for training, even with that setting on.
Delete: a deleted chat leaves your history at once and Anthropic's systems within 30 days, and it is not used to train future models.
The catch: with the setting on, chats can be kept in de-identified form for up to 5 years for training, so switch it off before you upload.

Gemini

Training: Keep Activity is on by default if you are 18 or over, and with it on, the photos and files you share can be used to improve Google's AI, with help from human reviewers.
Turn it off: on gemini.google.com, go to Settings & help, then Activity, click On near the top, then Turn off, or Turn off and delete activity. In the phone app, tap Menu, then your profile picture, then Gemini Apps Activity, and turn it off the same way.
Private option: a temporary chat is not used for training, though Google still keeps it, like every chat with Keep Activity off, for 72 hours.
Delete: on the same Activity page, choose Delete and a time range, or delete a single item.
The catch: chats already picked for human review are kept for up to 3 years, even after you delete your activity.

The honest bit

These switches cannot undo past training. Turning training off stops new chats being used, and Anthropic says data already used stays in models that have been trained, so the check comes first: anything you crop or cover before you upload never reaches the company at all.
Already uploaded something you should not have? Delete the chat, and in ChatGPT delete the file from Library too, then turn training off so nothing new is used. Anything already used to train a model stays there.
Skip the thumbs up or down on a chat with a private file. In all three apps, rating a reply can send the whole chat, including what you uploaded, for training or review even with training off.
The prompt only knows what you tell it, so zoom in and look over the whole photo, background included, before you describe it.

Take the location out of a photo

When a phone camera has location turned on, it saves where each photo was taken inside the file, and Apple warns that whoever you share it with may be able to see where it was taken. Take it out before you upload.

On iPhone: open the photo in Photos, tap the More button, then Adjust Info, Adjust Location, No Location. To share a copy without it instead, tap Share, then Options, and turn off Location. Removing it from the photo is the surer route when you attach from inside the AI app.
On Android: Google Photos cannot remove a location your camera added, so turn off location saving in your camera app's settings before you take the photo (where that setting lives varies by phone maker); for a photo you already took, move it to a computer and run it through ExifCleaner (below).
On a computer: ExifCleaner, a free, open-source app maintained by one developer (latest release 24 September 2026), strips the location and camera details from any photo you drag in, on Mac, Windows and Linux. It is not code-signed, so expect a warning the first time you open it; its README shows how to get past it.

Before your next upload

Check the training switch in the app you use most. It takes a minute: ChatGPT uses your chats unless you opt out, and Gemini keeps activity on by default for adults.
Save the prompt where you keep notes. Download it with the button above, so it is there the moment you are about to attach something.
Run it on the next photo before you attach it. Upload the version it suggests in a Temporary chat in ChatGPT, which keeps the file out of your Library, an incognito chat in Claude or a temporary chat in Gemini, then delete what is left once the job is done.