Are AI Chatbots Safe? A Privacy and Safety Guide (2026)

By John White · Senior Tools & Tech Writer · Updated October 19, 2026

Part of our AI Chatbots & Assistants guide

For most people, yes. Popular AI chatbots like ChatGPT, Claude, Gemini and Copilot are mostly safe for everyday tasks such as writing, planning and studying. You just need to follow a few simple rules.

There are four main risks:

Privacy, meaning what happens to what you type
Wrong answers that sound sure of themselves
AI-powered scams
Kids or teens leaning on a chatbot too much

Most of these risks are easy to lower. You just need to know which settings to change and what to keep to yourself.

Regulators are watching closely. In December 2024, Italy's data protection authority fined OpenAI 15 million euros over how ChatGPT handled personal data.

In September 2025, the U.S. Federal Trade Commission ordered seven companies, including OpenAI, Meta and Alphabet, to explain how their companion-style chatbots affect children and teens. For a wider look at the tools themselves, see our AI chatbots and assistants guide.

The Main Risks of Using AI Chatbots

Privacy and data training

Most chatbots save your chats, and many use them to train future AI models unless you say no. On most personal plans, training is on by default. Companies say they remove or mask personal details first, but no filter catches everything.

The good news is that most major chatbots let you opt out, as shown below.

Data leaks and sharing

Your chats can end up in places you did not expect. According to Bloomberg, Samsung restricted staff use of AI tools in 2023. An engineer had uploaded sensitive internal code to ChatGPT.

In mid-2025, some shared ChatGPT conversations showed up in Google search. Users had ticked a "discoverable" box, and OpenAI removed that feature. In January 2025, security firm Wiz found an exposed DeepSeek database holding chat history.

The lesson: a share button can make a chat public, and any company can have a security slip.

Wrong answers and hallucinations

Chatbots sometimes make things up and state them with total confidence. This is called a hallucination. They predict likely words and do not look up facts like a librarian.

In 2023, two New York lawyers and their firm were fined $5,000. They had filed a court brief that cited fake cases ChatGPT invented.

In 2024, a Canadian tribunal ordered Air Canada to pay a customer. The customer had relied on wrong fare advice from the airline's own website chatbot.

Scams, deepfakes and phishing

Criminals use AI too. The FBI warned in December 2024 that scammers use generative AI in three ways:

To write convincing messages
To clone the voices of family members
To create fake photos and videos

It suggested that families agree on a secret word or phrase to confirm who is really calling.

An AI agent that can browse websites or read your email can be tricked by hidden instructions on a web page. Give these tools only the access they truly need.

Emotional over-reliance

Chatbots are always available, patient and agreeable. That can make them feel like a friend.

A 2025 study by OpenAI and the MIT Media Lab looked at heavy users. People who used ChatGPT the most each day tended to report more loneliness and emotional dependence. The study did not prove the chatbot caused those feelings.

In October 2025, OpenAI estimated that about 0.15% of its weekly users have conversations with explicit signs of possible suicidal planning. A chatbot is not a therapist.

In the U.S., anyone who is struggling can call or text 988 to reach the 988 Suicide and Crisis Lifeline.

Kids and teens

Young people often use chatbots as companions. A July 2025 Common Sense Media survey found that 72% of U.S. teens had used an AI companion at least once. And 52% used one regularly.

Kids may overshare, believe wrong answers or get attached to a chatbot character. Character.AI stopped allowing open-ended chats for users under 18 in November 2025.

What Happens to Your Chats?

Your message goes to the company's servers, gets processed and is usually saved to your account.

Storage. Your history stays in your account until you delete it. OpenAI and Anthropic say deleted chats are removed from their systems within about 30 days. Google auto-deletes Gemini activity after 18 months by default for adults, and you can change that to 3 or 36 months.

Training. If training is on, parts of your chats may be used to teach future models. Anthropic says it may keep de-identified chats for up to five years when you allow training, versus 30 days when you do not. Turning the setting off later does not undo past training.

Human review. Real people sometimes read chats. Google says trained reviewers, including outside service providers, read some Gemini conversations. Those copies are separated from your account but kept for up to three years, even if you delete your activity.

Safety and legal holds. Even "temporary" chats are not erased instantly. ChatGPT Temporary Chats and Claude incognito chats are kept for up to 30 days. Gemini keeps temporary chats for up to 72 hours.

Anthropic says chats flagged by its safety systems can be kept for up to two years. In 2025, a court order in The New York Times' copyright lawsuit required OpenAI to keep some deleted ChatGPT chats for several months.

No legal privacy shield. A chatbot is not a doctor or lawyer. OpenAI CEO Sam Altman said in July 2025 that ChatGPT conversations do not have that kind of legal confidentiality. So they could be requested in a lawsuit.

How the Major Chatbots Handle Your Data

Work, school and enterprise plans usually have stronger protections and are not used for training by default. This table covers personal accounts. It summarizes public policies and reporting as of early October 2026, so check your settings, because policies change often.

Chatbot	Trains on chats by default?	Can you opt out?	Notes
ChatGPT	Yes, on personal plans (Free, Go, Plus and Pro)	Yes. Settings, then Data Controls, then turn off "Improve the model for everyone"	Temporary Chats are not used for training and are deleted within 30 days. Ads are being tested on the Free and Go plans.
Claude	Only if "Help improve Claude" is on. Anthropic asked users to choose in 2025, with the switch preset to on	Yes. Settings, then Privacy	With training on, chats may be kept up to five years; with it off, 30 days. Incognito chats are never used for training. Anthropic says Claude will stay ad-free.
Google Gemini	Yes, for adults with Keep Activity on, which is the default	Yes. Turn off Keep Activity (formerly Gemini Apps Activity)	Even with it off, chats are kept up to 72 hours. Chats read by human reviewers are kept up to three years.
Microsoft Copilot	Yes, for many signed-in personal users	Yes, where available. Profile, then Privacy, then turn off model training on text and voice	Microsoft says opting out does not stop use for ads, safety and general product improvement. Copilot can show ads.
Meta AI	Yes. Meta says AI chats help improve its AI	No simple off switch for most U.S. users	Since December 16, 2025, Meta AI chats can shape the ads and posts you see on Facebook and Instagram, except in the EU, UK and South Korea. Meta says it skips sensitive topics like health and religion for ads.
DeepSeek	Yes. Its privacy policy says inputs may be used to train models	The policy describes a right to opt out; look in the app's data settings	Data is stored on servers in China. Several countries limited or blocked the app in 2025.
Perplexity	Yes, for personal accounts	Yes. Look for the AI data retention setting in your account settings	Enterprise accounts are not used for training. Its Comet AI browser can see the pages you visit, so review its permissions.

OpenAI began testing ads in ChatGPT's Free and Go plans on February 9, 2026. It has since expanded the test well beyond the U.S.

OpenAI makes three promises about these ads:

Advertisers do not see your chats.
Ads stay away from health, mental health and political topics.
Users under 18 do not see ads.

Anthropic pledged in February 2026 to keep Claude ad-free. Google said in April 2026 that it is open to ads in the Gemini app but is not rushing. To see what you get without paying, read is ChatGPT free.

A simple rule: do not type anything you would not want a stranger to read. Keep these out of your chats:

Passwords, PINs and one-time login codes
Credit card, bank account and Social Security numbers
Answers to security questions, like your first pet's name
Medical records or test results that show your name
Private work files, client data or company secrets
Your home address and daily routine
Your child's full name, school or photos
Details of legal problems you would not want read in court
Other people's private messages or personal information
Intimate photos or anything you would not want made public

Need help with a sensitive document? Remove names, numbers and addresses first, or swap in placeholders like "Patient A."

How to Make Your Chatbot More Private

Use the steps below. Menu names change, so look for similar wording if your app looks different.

Turn off model training. In ChatGPT, go to Settings, then Data Controls, and turn off "Improve the model for everyone." In Claude, go to Settings, then Privacy, and turn off "Help improve Claude." In Gemini, turn off Keep Activity. In Copilot, turn off model training under Privacy. In Perplexity, turn off AI data retention.
Use temporary chats for sensitive questions. ChatGPT's Temporary Chat, Claude's incognito chat and Gemini's temporary chat stay out of your history and out of training. Switch them on for each new chat.
Check what the chatbot remembers. Review saved memories, delete anything personal or turn memory off. In ChatGPT, also check the "reference chat history" setting.
Delete old chats and set auto-delete. In Gemini, you can set activity to auto-delete after 3 months instead of 18.
Review connected apps. Disconnect any email, calendar, cloud drive or browser links you do not use.
Manage shared links. Anyone with a share link can usually read that chat. Delete old links in your settings.
Turn off ad personalization. OpenAI says ChatGPT users can turn off ad personalization and clear the data used for ads.
Lock down your account with a strong, unique password and two-step verification.
Download only official apps from the company's website or its verified app store listing.

Are AI Chatbots Safe for Kids?

AI chatbots can help kids learn, but most were built for adults, so kids need supervision. Companion and character apps carry extra risks. Our guide to Character AI alternatives lists their age rules app by app.

Start with the age rules for the main assistants:

ChatGPT: Users must be at least 13, and teens under 18 need a parent's permission. OpenAI also uses age prediction to apply teen settings.
Claude: Anthropic requires Claude.ai users to be 18 or older.
Gemini: Teens can use their own Google account where local rules allow. Since May 2025, children under 13 can use Gemini through parent-managed Family Link accounts.
Meta AI: Teens can use the Meta AI assistant with age-appropriate protections. In January 2026, Meta paused teen access to its AI characters while it builds a version with parental controls.

ChatGPT offers some of the most detailed parental controls. Since September 2025, parents can link accounts with their teen and then:

Set quiet hours
Turn off memory, voice mode and image creation
Opt the teen out of model training
Get an alert if the system detects signs of serious distress

California's SB 243 has been in effect since January 2026. It requires companion chatbot apps to remind known minors at least every three hours that they are talking to an AI and should take a break.

Practical tips for parents:

Keep young kids' chatbot use in shared family spaces.
Explain that a chatbot is a computer program, not a friend or therapist, even when it sounds caring.
Set homework rules: AI can explain and quiz, but it should not write the assignment.
Check in often. Watch for secrecy, pulling away from friends or distress when the app is off-limits.

Can You Trust What a Chatbot Says?

Trust it the way you would trust a smart friend who sometimes guesses.

In October 2025, a study led by the European Broadcasting Union and the BBC covered 22 public broadcasters in 18 countries. It found that 45% of AI assistant answers about news had at least one significant problem.

In March 2025, Columbia University's Tow Center for Digital Journalism tested AI search tools. They gave incorrect answers to more than 60% of requests to identify the source of a news article.

Here is how to check an answer before you rely on it:

Ask for sources, then open them. Make sure each link exists and says what the chatbot claims.
Cross-check important facts with an official source or a second tool.
Be extra careful with health, legal and money questions. Confirm with a professional.
Watch for red flags like exact numbers with no source or confident claims about recent events.
Ask clearer questions. Our guide on how to write better AI prompts shows how specific prompts lead to better answers.

The Bottom Line

AI chatbots are safe enough for everyday use when you treat them like a helpful public tool, not a private diary. The biggest risks come from what you share, settings you never changed and answers you never checked.

Turning off training, using temporary chats and keeping personal details out of your prompts removes most of the privacy risk.

The chatbots are not all the same. ChatGPT, Claude, Gemini and Copilot give you clear privacy controls. Meta AI offers fewer choices, and DeepSeek stores data in China.

If privacy matters to you, compare your options in our list of the best AI chatbots.

Next steps: Open your chatbot's settings today and do four things:

Decide whether to turn off model training.
Review saved memories.
Review connected apps.
Set up parental controls for any teen in your home.

Recommended

Explore the complete AI chatbots guide

Compare every major assistant, see free vs paid plans and find the right AI for each job.

Read the guide →

Frequently Asked Questions

Is ChatGPT safe to use?

Yes, ChatGPT is generally safe for everyday tasks, and OpenAI offers solid privacy controls. By default, chats on personal plans can be used for training. Turn off "Improve the model for everyone" if you prefer.

Use Temporary Chat for sensitive topics and never share passwords or financial details. Double-check important answers, since ChatGPT still makes mistakes.

Does ChatGPT sell your data?

OpenAI says it does not sell your data. It says advertisers cannot see your chats, chat history or memories. Your chats can still be used to train its models unless you opt out.

On the Free and Go plans, ChatGPT may use the topic of your current chat to choose an ad. Advertisers get only combined, anonymous stats. You can turn off ad personalization in your settings.

Is DeepSeek safe?

DeepSeek's app carries more privacy risk than most major chatbots. Its privacy policy says data is stored on servers in China. Laws there can require companies to share data with authorities.

Italy blocked the app and South Korea paused new downloads in 2025. Several governments have banned it on official devices. If you use it, share nothing personal, or run the open-weight model on your own computer so your data never reaches DeepSeek.

Are deleted chatbot conversations really deleted?

Mostly, but not instantly. OpenAI and Anthropic say deleted chats are removed from their systems within about 30 days.

Google keeps Gemini chats that human reviewers have read for up to three years, even after you delete your activity. Companies may also keep some data longer for safety or legal reasons.

Is it safe to use a chatbot for health or mental health advice?

Chatbots can help you understand medical terms or prepare questions for a doctor. But they should not replace professional care. They can give wrong or outdated answers, and your chats do not have the legal privacy of a doctor's visit.

If you are in crisis in the U.S., call or text 988 to reach the 988 Suicide and Crisis Lifeline.