Privacy and Your Data
Every app is a little like a store clerk who writes down everything you buy, and some of them sell the notebook.
AI runs on data, and much of that data is about people: what we write, search, buy, watch, and where we go. Lesson 2 called data the "food" of AI. This lesson asks the next question: whose food is it, and who gets to decide how it's used?
You'll learn what AI companies collect and why, what happens to your chatbot conversations, how to change the settings that matter, what never to type into any AI tool, how AI is used for surveillance, and what rights the law gives you.
Privacy can feel hopeless, as if everything is already out there. It isn't. A handful of choices, most taking just a few minutes, make a real difference. By the end of this page you'll know which ones.
How to use this page
- There are six sections. Each takes about 8 to 10 minutes.
- Start anywhere. If you only read one, read Section 4 (what never to share). If you have ten more minutes, do the checkup in Section 3.
- Press Focus at the top to show just one section at a time. Your checkmarks are saved on this device.
- This lesson pairs with the Making Sense of Numbers module, which covers anonymization and differential privacy in more depth.
Focus mode is on: one section at a time. Press Focus again to see everything.
AI companies collect data to train their models, to run and improve their products, to keep them safe, and in some cases to target advertising, and that data includes not just what you type but information about your device, your location, and how you use the service.
Analogy: the store clerk's notebook
Picture a clerk at the corner store who writes down everything you buy, when you came in, what you looked at but didn't buy, and who you came with. Some clerks use the notebook only to keep the shelves stocked. Some use it to suggest things you might like. And some sell copies of the notebook to other stores.
Every app keeps a notebook. The privacy question is always: what's written in it, who reads it, and who else gets a copy?
Four kinds of data collected
| Kind | Examples |
|---|---|
| What you give it | Your prompts, uploaded files and photos, voice recordings, account name, email, phone number, payment details |
| What it records about your use | When and how often you use it, which features, how long conversations last, thumbs-up or thumbs-down ratings |
| What your device reveals | Approximate location (from your internet address), device type, browser, operating system, sometimes precise location if you allow it |
| What it gets from elsewhere | Information from connected apps (email, calendar, drive) if you link them; data bought from partners or brokers by some companies; public web pages used for training |
Why they collect it
- To answer you. The AI needs your question and the conversation to respond at all.
- To provide features like chat history and memory (Lesson 3, Section 6).
- To train and improve models. Real conversations show companies what people actually ask and where the AI fails (Section 2).
- For safety and abuse prevention. Companies review some conversations to catch misuse, such as fraud or attempts to get dangerous information.
- For legal reasons. Records may be kept to comply with laws or respond to court orders.
- For advertising and business, in some cases. Companies whose main business is advertising (like Google and Meta) have long used information about users to target ads, and some AI products are beginning to show ads too. Each company's policy is different.
The hidden layer: data brokers
Beyond the apps you use directly, there's a whole industry you've probably never dealt with: data brokers. These companies gather information from public records, purchases, apps, websites, and loyalty programs, then build profiles and sell them. A single profile can include your address history, relatives, estimated income, health interests, and shopping habits. AI makes these profiles more powerful by predicting things you never shared: whether you're pregnant, likely to move, or struggling financially.
What AI adds to the old privacy problem
- Inference: AI can guess sensitive things from harmless-seeming data, such as health conditions from shopping patterns or personality from writing style.
- Intimacy: people tell chatbots things they'd never type into a search engine: worries, symptoms, relationship troubles, business plans.
- Scale: AI can sort through enormous amounts of data about millions of people quickly and cheaply.
Ask this: three questions for any AI app
- What does it collect? (Look for "Privacy Policy" or "Privacy Center," usually linked at the bottom of the website or in settings.)
- Does it use my conversations to train its AI, and can I turn that off?
- Does it share or sell my data, and to whom?
Tip: you can paste a long privacy policy into a chatbot and ask, "In plain language, what does this policy let the company do with my data?" Then spot-check the parts that matter.
Watch out for
- "Free" apps. When you don't pay with money, you often pay with data. That's not always bad, but it's worth knowing the trade.
- Connecting everything. Linking your email, calendar, and files to an AI assistant can be very useful, and it also gives that company access to a great deal. Connect only what you need.
On most free and personal chatbot plans, your conversations may be used to train future AI models unless you turn that setting off, and some conversations may be read by human reviewers, while business and enterprise plans generally don't train on customer data by default.
Analogy: the cooking class recording
Imagine taking a cooking class where the school records every session to train future instructors. Most of what you say is harmless ("How long do I knead this?"). But if you mentioned your divorce while chopping onions, that's on the recording too, and future instructors might learn from it.
Chatbot training is similar. The company's goal is to improve the AI, not to learn about you specifically. But whatever you said is part of what they're learning from.
How training on chats works
Conversations can be used to fine-tune future versions of a model and to test where it goes wrong (Lesson 2, Sections 2 and 5). Companies say they filter and de-identify training data, removing things like names and account numbers where they can. But filtering isn't perfect, and "de-identified" doesn't always mean "impossible to trace back," especially for unusual details.
Can the AI repeat what I told it to someone else?
It's unlikely for any particular conversation, because a single chat is a tiny drop in an ocean of training data, and companies work to prevent models from memorizing personal details. But it's not impossible: researchers have shown that models can sometimes reproduce rare pieces of their training data. The safest assumption: don't type anything into a chatbot that you'd be seriously harmed by if it were ever seen by someone else.
Human reviewers
Most major AI companies have trained staff or contractors review a small portion of conversations for safety and quality. Reviewed content may be kept longer than other data. Some companies' policies say that conversations flagged for safety reasons may be retained even if you've turned off training.
How plans differ
| Type of account | Training on your chats (typical) |
|---|---|
| Free and personal paid plans (the ones most individuals use) | Often on unless you turn it off. Setting names and defaults vary and change; check yours. |
| Temporary or incognito chats | Generally not used for training; deleted after a short period (days to weeks, depending on the company). |
| Business, team, education, and enterprise plans | Generally not used for training by default, under contract. |
| AI features inside other apps (email, social media, office software) | Follows that app's own policy, which may differ from the standalone chatbot's. |
How long data is kept also varies widely: from about 30 days for some temporary chats to several years for conversations used in training. Each company publishes its rules in its privacy policy.
Your chats aren't protected like a doctor visit
Conversations with a doctor are protected by health privacy law (HIPAA), and conversations with a lawyer are protected by legal privilege. Conversations with a consumer chatbot generally aren't, even when you're asking about your health or a legal problem. In some lawsuits, courts have ordered AI companies to preserve users' chat logs as evidence. The heads of major AI companies have publicly acknowledged this gap.
Watch out for
- "Delete" not meaning "gone." Deleting a chat removes it from your view, but companies may keep copies for a period for safety or legal reasons. Data already used in training can't be pulled back out of a finished model.
- Treating a chatbot like a confidential advisor. It may feel private. Legally and technically, it usually isn't.
Most AI tools give you real controls (turning off training, using temporary chats, managing memory, deleting history, and limiting connected apps), but they're often a few menus deep and change over time, so a short checkup a couple of times a year is worth the effort.
Analogy: the thermostat you never adjusted
Many people live for years with whatever temperature the builder set on the thermostat. It works, sort of. But a two-minute adjustment could make the house more comfortable and save money.
Privacy settings are the same. The defaults were chosen by the company, for the company. You're allowed to adjust them.
The settings that matter most
| Setting | What it controls | Words to look for |
|---|---|---|
| Model training | Whether your chats help train future AI | "Improve the model," "Help improve our AI," "Model training," "Activity" |
| Temporary chat | A conversation that isn't saved to history or used for training | "Temporary chat," "Incognito," a ghost or clock icon |
| Memory | What the app remembers about you between chats | "Memory," "Personalization," "Saved info" |
| Chat history | Past conversations stored in your account | "History," "Activity," "Delete all chats" |
| Connected apps | Access to your email, calendar, files, and other accounts | "Connectors," "Apps," "Integrations," "Extensions" |
| Data export | A copy of everything the company holds about you | "Export data," "Download your data" |
Where to find them in the big three
Menus change often, so treat these as starting points (as of late 2026):
- ChatGPT: Settings → Data Controls → "Improve the model for everyone." Temporary chats are available from the top of a new chat. Memory is under Personalization.
- Claude: Settings → Privacy → the setting about helping improve Anthropic's models. Memory and incognito options are also available.
- Gemini: Gemini Apps Activity (sometimes called "Keep Activity") controls whether chats are saved and used to improve Google's products.
If you can't find something, ask the chatbot itself: "Where do I turn off using my chats for training?" (Then confirm by finding the setting.)
Beyond chatbots: other AI privacy controls
- Social media: some platforms use public posts to train AI. Look for AI or data-use settings, and consider making accounts private.
- Phones: check which apps have access to your location, microphone, camera, and photos (Settings → Privacy on iPhone; Settings → Security & privacy on Android). Choose "While using the app" instead of "Always."
- Smart speakers and doorbells: review voice-recording history and sharing settings in their apps.
- Browsers: block third-party tracking cookies and consider a privacy-focused browser or extension.
- Global Privacy Control: a browser signal telling websites "don't sell or share my data." In several states, websites are legally required to honor it.
Getting off data broker lists
You can ask data brokers to delete your information. Doing it one company at a time is tedious, since there are hundreds. Options:
- California residents can use the state's free DROP (Delete Request and Opt-out Platform) to send one deletion request to every registered data broker. Since August 2026, brokers must check it regularly and process requests.
- Everyone else can use state privacy rights where available (Section 6), or paid services that send removal requests on your behalf. Compare them carefully; quality varies.
Do this: the 15-minute AI privacy checkup
- In each chatbot you use, decide on the training setting deliberately rather than leaving the default.
- Review saved memory. Delete anything wrong or anything you'd rather it forgot.
- Delete old chats containing personal details you no longer need.
- Check connected apps. Disconnect any you don't actively use.
- Learn where temporary chat is, for the next sensitive question.
- On your phone, review location, microphone, and camera permissions.
- Put a reminder on your calendar to do this again in six months.
Watch out for
- Settings that reset or move. App updates sometimes change where settings live or add new ones. The six-month checkup catches this.
- Opting out isn't retroactive. Turning off training stops future use; it generally doesn't remove data from models already trained.
Some information should never go into any AI tool, however helpful or trustworthy it seems: passwords, full ID numbers, financial account details, and other people's private information, because once typed, you lose control of where it goes.
Analogy: the postcard test
A letter in a sealed envelope feels private. A postcard can be read by every mail carrier who handles it.
Treat what you type into a chatbot like a postcard. It's probably fine that the mail carrier might read "Having a lovely time at the lake!" It's not fine if the postcard says your bank PIN.
The never list
| Never share | Why | What to do instead |
|---|---|---|
| Passwords, PINs, security answers, one-time codes | These unlock everything. No legitimate AI tool needs them. | Use a password manager. If an "AI assistant" asks for a code, it's a scam. |
| Social Security, passport, driver's license, Medicare numbers | Keys to identity theft. | Replace with "[SSN]" or "XXX-XX-1234" if the AI needs to know the document's format. |
| Bank, credit card, and investment account numbers | Direct path to your money. | Describe the situation without numbers: "My checking account was charged twice." |
| Full medical records with your name | Health data isn't protected by health privacy law in consumer chatbots. | Ask about terms and conditions generally, or remove names, dates of birth, and ID numbers first. |
| Other people's private information | They didn't consent. Their secrets aren't yours to share. | Use a first initial or "my friend," and leave out identifying details. |
| Confidential work material | Could violate your employer's rules, contracts, or the law. | Use only tools your employer has approved for that kind of information. |
| Photos of IDs, cards, or documents | Images carry every number on them, plus faces and addresses. | Crop or cover sensitive parts before uploading, or type only what's needed. |
The "redact first" habit
You can often get the same help without the sensitive parts. Replace them before you paste:
My name is Margaret Olsen, SSN 123-45-6789, and Medicare denied my claim #88213 at Parkland on 3/14. Account 4410-2297. What should I do?
Medicare denied a claim for a hospital outpatient procedure. The letter says "not medically necessary." What are my appeal options, what's the deadline usually, and what should I ask my doctor's office for?
Think twice about
Not forbidden, but worth a pause:
- Detailed health, mental health, or relationship information linked to your real name.
- Your exact home address and daily schedule.
- Photos of children, especially with names, schools, or locations.
- Legal matters you're actively involved in. Ask your lawyer first.
- Anything you'd be upset to see in a data breach headline.
Red flag: an AI asking you for sensitive information
Legitimate AI assistants don't ask for your password, PIN, or verification codes. If a chatbot, a "support agent," or an AI voice on the phone asks for these, stop. It may be a scam site imitating a real company, or a hijacked agent (Lesson 4, Section 6).
Do this
- Before pasting any document, scan it for numbers: account, ID, policy, phone. Remove or replace them.
- Use temporary chat for sensitive but necessary questions.
- If you've already shared something on the never list, delete the chat, and for passwords or account numbers, change them or call your bank.
Watch out for
- Fake AI apps and sites. Look-alike chatbot apps exist mainly to harvest what people type. Download only from the official company.
- Friendly familiarity. The more comfortable a chatbot feels, the easier it is to overshare. The postcard test still applies.
AI makes it possible to watch, identify, and track people at a scale no human workforce ever could, which can help find missing people and solve crimes but also enables mass surveillance, wrongful accusations, and pressure on free speech, and much of it happens without our knowledge.
Analogy: from the town gossip to a million watchers
In a small town, the gossip on the porch might notice who visited whom. Annoying, but limited: one person, one street, a fuzzy memory.
AI surveillance is like a million porches, each with a perfect memory, all comparing notes instantly, and never going to sleep. The difference isn't just more watching. It's a different kind of watching.
Facial recognition
(Lesson 4, Section 4 covered how it works.) Privacy concerns center on:
- Scraped face databases. Some companies built databases of billions of face photos copied from social media and websites without consent, then sold searches to police and others. Regulators in several countries have fined such companies and ordered data deleted.
- Use in public places to identify people at protests, stores, stadiums, and on streets.
- Errors and wrongful arrests, which have fallen hardest on Black Americans in documented U.S. cases.
- Biometric laws. A few U.S. states, notably Illinois and Texas, have laws restricting how companies collect face and fingerprint data. Under them, Meta and Google each agreed to pay Texas well over a billion dollars in separate settlements in 2024 and 2025.
Tracking
- Phone location data is collected by many apps and sold through data brokers. Investigations have shown it can reveal visits to homes, clinics, and places of worship, and that government agencies have bought such data without warrants.
- License plate readers: networks of cameras record where cars go, building histories of drivers' movements. Many police departments and neighborhoods use them; some cities have debated or dropped them over privacy concerns.
- Online tracking: cookies, tracking pixels, and device fingerprinting follow you from site to site, and AI turns those trails into predictions about you.
- Workplace monitoring: some employers use AI to track keystrokes, screen activity, or even facial expressions of employees.
Predictive policing
Predictive policing uses past crime data to forecast where crimes might happen, or who might be involved, so police can focus resources there.
The case supporters make
Police resources are limited, and data can help put them where they're most needed.
Patterns in crime are real, and ignoring them isn't neutral either.
Well-designed, audited tools could be more consistent than human hunches.
The case critics make
Past arrest data reflects past policing, not just past crime. Neighborhoods that were heavily policed produce more records, which sends more police, which produces more records: a feedback loop.
Lists of "likely offenders" can label people who haven't done anything.
Evidence of effectiveness has been mixed, and several programs have been dropped.
Real-world outcomes have reshaped the debate: Los Angeles and Chicago both ended prominent predictive policing programs after criticism and audits, and some cities, such as Santa Cruz, California, banned the practice. Other departments continue to use data-driven tools.
Why surveillance matters even if "I have nothing to hide"
- Mistakes happen, and they can happen to anyone: a wrong face match, a misread plate.
- Information outlives its purpose. Data collected for one reason can be used for another, by different people, years later.
- The chilling effect. When people know they're watched, many hesitate to attend a protest, visit a clinic, or speak freely, even when they've done nothing wrong.
- Power. Whoever holds detailed records about everyone holds power over them.
Do this
- Limit location sharing to apps that truly need it, and turn off precise location where it's optional.
- Reset your phone's advertising ID or limit ad tracking in phone settings.
- Find out what your city uses. Many city councils publicly discuss surveillance technology purchases. Attending or commenting is a way to have a say.
Watch out for
- "It's just for safety." Safety matters, and so do limits, oversight, and accountability. Good programs welcome all three.
- Hopelessness. Surveillance can feel unstoppable, but public pressure has repeatedly led cities and companies to change course.
You have more legal rights over your data than most people realize, including, in many places, the right to see what companies hold about you, correct it, delete it, and stop its sale, but in the U.S. those rights depend heavily on which state you live in.
Analogy: the patchwork quilt
Europe's privacy protection is like one big blanket covering the whole continent. The United States has a patchwork quilt: a square here for health records, a square there for children, and a growing number of state squares, with some gaps where no square covers you at all.
Knowing which squares cover you is the first step to using them.
The core rights, in plain terms
| Right | What it means |
|---|---|
| Know / access | Ask a company what personal data it has about you and get a copy. |
| Correct | Fix inaccurate information. |
| Delete | Ask a company to erase your data, with some exceptions. |
| Opt out | Stop the sale of your data, targeted advertising, and certain automated profiling. |
| Portability | Take your data with you in a usable format. |
| Sensitive data protection | Extra limits on health, biometric, precise location, and similar data; often requiring your consent. |
| No retaliation | A company can't punish you (for example, with higher prices) for using these rights, with limited exceptions. |
United States: federal laws (specific areas only)
There's no comprehensive federal privacy law in the U.S. Proposals have been introduced in Congress but haven't passed. Instead, federal laws cover specific areas:
- HIPAA: health information held by doctors, hospitals, and insurers. Not health apps or chatbots, in general.
- COPPA: online data of children under 13.
- FERPA: student education records.
- Fair Credit Reporting Act: credit reports, including the right to see and dispute them.
- The FTC can act against companies that break their own privacy promises or use unfair or deceptive practices, including with AI.
United States: state laws
As of 2026, about twenty states have comprehensive consumer privacy laws, including California, Virginia, Colorado, Connecticut, Texas, Oregon, and more recently Indiana, Kentucky, and Rhode Island (which took effect in January 2026). They generally give residents the rights in the table above, though details differ. California's law is the oldest and strongest, and it has a dedicated privacy agency.
If your state isn't on the list, many large companies still honor these requests for everyone, because it's simpler than tracking where each customer lives. It never hurts to ask.
Europe and beyond
- GDPR (European Union, since 2018): one of the world's strongest privacy laws, giving broad rights and allowing large fines. Many companies worldwide follow it.
- EU AI Act (being phased in 2024–2027): bans certain AI uses outright (like social scoring by governments and most real-time facial recognition by police in public), and sets rules for high-risk AI.
- Many other countries, including Canada, the UK, Brazil, Japan, and India, have their own privacy laws.
Do this: using your rights
- Find the request form. Look for "Your Privacy Choices," "Do Not Sell or Share My Personal Information," or "Privacy Request" links at the bottom of company websites, or in the privacy policy.
- Ask for a copy of your data from one AI company you use. Many offer a "Download your data" button. It's eye-opening.
- Turn on Global Privacy Control in your browser, if available, to signal opt-out automatically.
- Complain if ignored. Your state attorney general's office and the FTC (ReportFraud.ftc.gov) accept privacy complaints.
Your voice counts
Privacy law is still being written, at the state, national, and global levels. Public comments, calls to representatives, and local debates about surveillance technology have shaped real outcomes. Understanding the basics, as you now do, is what makes participation possible.
Watch out for
- Assuming "there must be a law against that." Often, in the U.S., there isn't, especially for data brokers, health apps, and chatbots.
- Laws that change. Privacy law is moving fast. Check your state's current rules before relying on a specific right.
Putting it together
Six things to remember about AI and your privacy:
- Every app keeps a notebook. AI companies collect what you type, how you use the service, and device details, for answers, features, training, safety, legal reasons, and sometimes ads.
- Your chats may train future models on personal plans unless you turn it off, and some are read by reviewers. Chatbot conversations aren't legally protected like a doctor or lawyer visit.
- You have real controls: training, temporary chats, memory, history, and connected apps. Do a 15-minute checkup twice a year.
- Apply the postcard test. Never share passwords, ID numbers, account numbers, or other people's secrets. Redact first.
- AI supercharges surveillance, through face recognition, location tracking, and predictive tools, with real benefits and real risks to fairness and freedom.
- You have rights, especially in the roughly twenty states with privacy laws and in Europe. Use them, and have a say in how new rules are written.
Check yourself
Try to answer each question in your head, then tap to see a sample answer.
1. Name three reasons AI companies collect data.
Any three of: to answer you, to provide features like memory, to train and improve models, for safety and abuse prevention, for legal reasons, and in some cases for advertising.
2. What is a data broker?
A company that gathers information about people from many sources, builds profiles, and sells them, usually without people knowing.
3. On a free or personal chatbot plan, are your chats used for training?
Often yes, unless you turn the setting off. Business and enterprise plans generally don't train on customer data by default. Check your own settings.
4. Are your health questions to a chatbot protected by HIPAA?
Generally no. HIPAA covers doctors, hospitals, and insurers, not consumer chatbots or most health apps.
5. What is the "postcard test"?
Treat anything you type into a chatbot like a postcard that others might read. Don't include anything you'd be harmed by if it were seen.
6. How can you get help from AI on a sensitive letter without oversharing?
Redact first: remove names, ID numbers, and account numbers, and describe the situation in general terms.
7. What is the feedback-loop criticism of predictive policing?
Heavily policed areas generate more arrest records, which the system reads as more crime, which sends more police there, generating still more records.
8. Name three data rights many privacy laws give you.
Any three of: know/access, correct, delete, opt out of sale and targeted ads, portability, sensitive data protection, no retaliation.
Glossary
- Biometric data
- Measurements of your body used for identification, such as face geometry, fingerprints, or voiceprints.
- Chilling effect
- When people hold back from lawful activities because they feel watched.
- Data broker
- A company that collects and sells profiles of people.
- De-identified data
- Data with obvious identifiers removed. It can sometimes still be traced back to individuals.
- DROP
- California's Delete Request and Opt-out Platform, which sends one deletion request to all registered data brokers.
- GDPR
- The European Union's General Data Protection Regulation, a broad privacy law in effect since 2018.
- Global Privacy Control (GPC)
- A browser signal telling websites not to sell or share your data.
- HIPAA
- U.S. law protecting health information held by healthcare providers, insurers, and their partners.
- Human review
- When company staff or contractors read some user conversations for safety or quality.
- Inference (privacy)
- AI guessing sensitive facts about you from data you did share.
- License plate reader
- A camera system that records license plates and locations of passing cars.
- Opt out
- Telling a company not to use your data in a particular way.
- Predictive policing
- Using data and algorithms to forecast where crime may occur or who may be involved.
- Privacy policy
- A company's official description of what data it collects and how it's used.
- Redact
- Remove or cover sensitive information before sharing.
- Temporary chat
- A chatbot conversation not saved to history or used for training.