Here is the short answer. Never paste passwords, government IDs, financial account details, patient records, confidential work files, or anything that links a real person to a sensitive situation into an AI chatbot. These tools are not secure vaults. They are internet services, and what you type into them can be stored, reviewed, trained on, or exposed in a breach.
But the longer answer is more important, because most people are not making obvious mistakes. They are making subtle ones. They paste a document to save time. They upload a photo to edit it. They describe a real situation using real names. None of those feel dangerous. Several of them are.
This guide covers everything you should keep out of an AI chatbot, why each one matters, what has already gone wrong for real people and companies, and what to do instead.
This is for anyone using ChatGPT, Claude, Gemini, Copilot, or any other AI chat tool on a personal or work account. It is especially relevant if you use AI at work, if you handle other people’s data, or if you have never once checked the privacy settings on these tools. If you have not changed your data settings yet, our guide on how to stop ChatGPT, Claude, and Gemini from using your chats for AI training covers the exact steps.
Why AI Chatbots Are Not As Private As They Feel
There is something about the chat interface that feels private. It is just you and a text box. Nobody is watching. It feels like a notepad.
That feeling is wrong.
When you type something into an AI chatbot and press enter, that text leaves your device, travels to a company’s servers, gets processed, and in most cases gets stored. Depending on your settings, it may be reviewed by human moderators, used to train future models, or exposed if that company suffers a breach.
According to information from Forbes’s technology coverage, there have been multiple instances of private AI conversations being leaked to third parties. A security researcher found over 143,000 conversations with tools like Claude, Copilot, and ChatGPT accessible on Archive.org. In July 2025, the BBC reported that hundreds of user conversations with Claude were exposed on the public internet, with links to the chats readily available on Google search results.
That is not a fringe risk. It happened. And the people whose conversations were exposed did not know it was coming.
Here is the practical rule to carry with you. If you would not be comfortable sending the information in an email to a stranger, do not paste it into a chatbot. Full stop.
Now let me walk you through each category of information you should keep out, with real examples of what goes wrong when people do not.
1. Passwords, API Keys, and Login Credentials
This is the most important one, so it comes first.
Never paste a password into any AI chatbot. Not to “test” it. Not to ask for help changing it. Not to show the chatbot what format it is in. None of those reasons justify it.
The same rule applies to API keys, two-factor authentication codes, backup codes, recovery keys, private keys, and security question answers. These are the keys to your accounts. Once they leave your device, you have lost control of where they go.
Here is why this matters more than people realize. A threat actor claimed to possess approximately 20 million OpenAI user credentials for sale on dark web forums in February 2025. In October 2024, a significant security breach resulted in over 225,000 sets of OpenAI credentials being discovered for sale on dark web markets, stolen through infostealer malware. If your account is ever compromised, every password you typed into that chatbot becomes accessible to whoever has your login.
In December 2025, security researchers uncovered a malware campaign involving malicious Google Chrome extensions designed to secretly collect user conversations from ChatGPT and DeepSeek and transmit them to attacker-controlled servers. The extensions looked like legitimate AI tools. They requested broad browser permissions. Once installed, they captured full conversation content including prompts and responses that may have contained sensitive credentials.
There is also a subtler risk. If you paste an API key into a chat asking for help using it, that key now sits in a log on a server you do not control. Anyone who later gains access to your chat history, through a breach, through account compromise, or through a shared link, gets that key too.
What to do instead: use a placeholder. If you need help with code that involves an API key, replace your real key with something like YOUR_API_KEY_HERE in the code. The chatbot can still help you without ever seeing the real value.
2. Government IDs, Passport Numbers, and Social Security Numbers
Uploading a photo of your passport or national ID card to an AI tool is one of the highest-risk things you can do with this technology. Government IDs are the foundation of identity theft. Once a bad actor has your full name, date of birth, and ID number together, they have nearly everything they need to impersonate you, open accounts in your name, or commit fraud.
Avoid uploading passports, national ID cards, driver’s licenses, residence permits, visas, tax documents, or insurance files to any general-purpose AI tool. This applies to uploading a photo of the physical document and to typing the number directly into a chat.
According to information from cybersecurity firm Bitdefender, any of these identifiers can be exploited to impersonate you, leading to identity theft, financial fraud, or other criminal misuse of your personal details.
Some AI tools now advertise identity verification features built into their products. Those are purpose-built, regulated services with specific legal obligations around data handling. A general consumer chatbot is not that. The difference matters enormously.
3. Financial Account Information and Payment Details
Your bank account number, credit card details, routing number, investment account information, paystubs, tax returns, and income details all belong off the chat window.
According to guidance from the University of Kentucky Information Technology Services, paystubs, banking and investment account information, and credit card details should not be shared with an AI chatbot. As a financial writer for Investopedia noted, if these documents are exposed, they can be used for blackmail, fraud, or tailored social engineering attacks against you or your family.
Here is a specific example of why this matters. According to information from Forbes, entering your credit card number or bank details into ChatGPT would mean that anyone who gains access to your account or chats could potentially use these details to commit fraud. Your chat history is not encrypted in the same way a bank’s systems are. If someone gets your ChatGPT login, they get your history too.
You can absolutely use AI tools for financial guidance. Ask for help understanding a concept, building a budget framework, or interpreting general tax rules. Just keep the actual numbers out of it. Replace real figures with generic examples if the chatbot needs numbers to help you.
4. Personal Medical and Health Information
This one catches people off guard because it feels so natural to describe your health situation to an AI. The chatbot listens patiently, does not judge, and often gives clear, helpful responses. That experience makes it easy to share more than you should.
Avoid sharing personal medical details, including healthcare records, diagnoses, medications, test results, or anything that links your name to a health condition. Even indirect details can be risky. As one security researcher noted in reporting shared across multiple outlets, asking for “heart-friendly dinner recipes” or sharing information about medications, sexual health, or specific conditions could allow the system or downstream users of the data to infer sensitive health profiles, which could potentially reach advertisers or insurers.
If you need AI help with a health question, keep the prompt general. Instead of “I have Type 2 diabetes and my A1C was 8.2 last week, what should I eat,” try “what diet changes are recommended for managing blood sugar.” You get the same useful information. You share nothing personal.
For anyone working in healthcare, the stakes are higher still. In the Netherlands, the Autoriteit Persoonsgegevens issued a warning in August 2024 after a GP practice employee entered patients’ medical data into an AI chatbot contrary to the agreements. The regulator treated it as a data breach. In Australia, a government agency experienced a breach in March 2025 when an employee uploaded a document containing flood program applicants’ personal and health information to ChatGPT. Both of these were ordinary employees trying to do their jobs faster. Neither intended to cause a breach.
5. Confidential Work Documents, Source Code, and Trade Secrets
This is where the most costly mistakes happen, because they happen at scale.
The Samsung case is the clearest example in the public record. In March 2023, Samsung Electronics allowed engineers at its semiconductor division to use ChatGPT to help with work tasks. Within 20 days, three separate incidents occurred. One engineer pasted the full source code of a semiconductor equipment measurement program to fix an error. Another pasted code from a yield and defect analysis program. A third fed in a transcript of a recorded confidential internal meeting to produce minutes. Samsung responded by capping prompts at 1,024 bytes as an emergency measure and later banned generative AI on company devices entirely.
The incidents were not the result of a cyberattack. No hacker was involved. Engineers were just trying to work faster. That is exactly what makes this pattern so common.
According to a Harmonic Security study from 2025, sensitive information appeared in over four percent of prompts and twenty percent of file uploads sent to AI tools. A LayerX report from the same year found that forty percent of files uploaded to generative AI tools contained PII or payment card data. IBM research from 2025 found that one in five organizations had already suffered a data breach tied to unsanctioned AI use, and that the average added cost of a breach involving shadow AI was approximately 670,000 dollars.
Following the Samsung incident, JPMorgan Chase restricted employee use of ChatGPT, citing concerns about sharing confidential financial data. Apple restricted use of ChatGPT and GitHub Copilot internally, concerned about source code leakage.
If you need AI help with a work project, redact it first. Replace client names with placeholders. Remove project codes, financial figures, and identifying details. Use the anonymized version to get the help you need, then apply it back to your real work. The chatbot does not need your actual client’s name to help you structure an email to them.
6. Photos with Hidden Location Data (EXIF Metadata)
Most people do not know that the photos on their phone contain hidden information called EXIF metadata. This embedded data can include the GPS coordinates of where the photo was taken, the date and time, and the device model.
When you upload a photo to an AI chatbot, that metadata travels with the file to the company’s servers. According to information from MetaClean’s published research, when you upload a photo to ChatGPT, the EXIF data including GPS coordinates travels with it. OpenAI retains that data and, unless you opt out, may use it to improve their models.
That means a photo you take at home and upload for an AI editing task could, without you knowing, be tagged with the GPS coordinates of your home address.
There is a second, less obvious risk that goes further. In April 2025, users discovered that ChatGPT running its o3 reasoning model could geolocate photos with striking accuracy using nothing but the image itself, no EXIF metadata required. It reads the visual content: buildings, signs, road markings, plants, the angle of the sun, and matches those elements against geotagged imagery. As reported by TechCrunch, the accuracy in many cases was described as surreal and, by privacy researchers, as genuinely alarming.
What this means practically is that stripping EXIF from a photo before uploading it reduces the risk but does not eliminate it. If you upload a photo taken in your back garden, near your child’s school, or at a regular location, an advanced AI model may be able to identify that location from the image alone. For most casual photo edits, this is a low risk. For anyone with safety concerns, including journalists, activists, or domestic abuse survivors, it is a serious one.
Before uploading any personal photo to an AI tool, strip the EXIF data using a tool like your phone’s built-in “remove metadata” option or a desktop app. And think carefully about what the image itself reveals in the background.
7. Real Names Paired with Sensitive Situations
This one is subtle, and it is the category where people make the most unintentional mistakes.
Entering your full name plus your home address is risky. But entering someone else’s name together with a sensitive situation is often more dangerous, and more common.
Here are some real patterns I see people fall into:
“My employee [full name] has been performing poorly and I need to write a termination letter.”
“My patient [full name] presented with the following symptoms and I need help writing notes.”
“My client [full name] is going through a divorce and I need to draft a letter about their financial situation.”
In each of these cases, the person is trying to do something legitimate and useful. But they are linking a real person’s identity to a sensitive, private situation inside a system that stores that conversation, and in many cases, uses it for training.
The fix is simple. Replace every real name with a placeholder before you start. “My employee John” becomes “my employee.” “My patient Sarah Jones” becomes “my patient.” The quality of the help you get does not change. The risk drops significantly.
According to information from cybersecurity researcher reporting cited by multiple sources, entering your name, address, birthdate, and even email into an AI tool may not seem risky until a malicious actor uses it to stitch that data together for phishing or impersonation attempts.
The same logic applies in reverse. Be careful about typing information that describes another real person in a way that could identify them without using their name. Descriptions of a specific job role, city, age, and situation can make someone identifiable even without a name attached.
8. Private Conversations, Relationship Details, and Emotional Confessions
AI chatbots can feel like a safe space to process difficult things. They are available at 2am, they do not get tired of the topic, and they never judge. I understand the appeal completely.
The risk is not that the AI will use your confession against you. The risk is that what you type is stored, potentially reviewed by a human moderator, potentially exposed in a breach, and potentially used to train future models.
According to information from cybersecurity coverage by Bitdefender, anything from political or religious views to relationship troubles or emotional struggles could be exposed if conversational logs are hacked or mishandled.
For processing genuinely personal or emotionally difficult topics, the safest approach is to use a platform’s Temporary Chat or Incognito Chat mode, remove your real name and identifying details, and keep the specifics vague enough that the conversation could not be traced back to you or the other person involved.
If you are struggling with something serious, mental health professionals operate under strict confidentiality rules that AI chatbots do not. That distinction matters.
9. Children’s Personal Information
This needs its own section because the risks are specific and serious.
Do not upload photos of children to AI tools. Do not describe situations that combine a child’s name with location, school, physical description, or daily routine. Do not ask an AI to help you write anything that would require entering a child’s personally identifying information to do it.
The EXIF metadata risk for children’s photos is the same as for any other photo, but the consequences of that information being exposed are more severe. Most countries have specific legal protections around children’s data for exactly this reason.
If you need AI help with something involving a child, use placeholder names, remove ages and school names, and keep the details general.
What Has Already Gone Wrong: A Short Record of Real Incidents
It helps to see what actually happened, not just what could happen. Here is a short factual summary from verified reports.
In March 2023, Samsung semiconductor engineers pasted confidential source code and a confidential meeting transcript into ChatGPT within their first three weeks of being allowed to use the tool. Samsung immediately restricted the service and later banned it on company devices.
In a reported incident from late 2024, an undisclosed AI provider was found to be retaining prompt data despite having deletion policies that said otherwise.
In January 2025, security firm Wiz discovered a DeepSeek database sitting completely open on the internet with no authentication required. The exposure included over a million lines of log data containing chat history, secret keys, and backend details, accessible to anyone who found the right URL.
In July 2025, Fast Company found that ChatGPT conversations shared using the “Make this chat discoverable” option were showing up in Google search results. Nearly 4,500 appeared in a single site search.
In August 2025, over 370,000 Grok conversations were indexed by Google, Bing, and other search engines, making them fully searchable by anyone.
In September 2025, Cybernews found that AI apps ImagineArt, Chatly, and Chatbotx had exposed 116GB of live user data including prompts, authentication tokens, and user agents on an unsecured server.
None of those incidents required users to do anything obviously wrong. In most cases, people were just using the tools.
The “What to Do Instead” Guide
Knowing what not to do is useful. Knowing what to do instead is more useful.
For passwords and credentials: Use a placeholder like [YOUR_API_KEY] or [PASSWORD_HERE] in any code or prompt. The chatbot can still help with the logic without seeing the real value.
For personal documents: Redact names, ID numbers, account numbers, and dates before pasting. Use find-and-replace in your document editor to swap real names for labels like “Client A” before uploading.
For medical questions: Frame questions in the third person and keep them general. “What are the treatment options for Type 2 diabetes” instead of describing your own records.
For work documents: Anonymize before you paste. Remove client names, project codes, financial figures, and any clause that is marked confidential. Most AI tools can help with structure, tone, and logic without needing the real details.
For photos: Strip EXIF data before uploading using your phone settings or a desktop tool. On iPhone, use the Markup tool to remove location before sharing. On Android, use Files by Google or a free EXIF editing app. For Windows, right-click any photo, select Properties, then Details, then “Remove Properties and Personal Information.”
For sensitive personal topics: Use Temporary Chat in ChatGPT or Incognito Chat in Claude. Remove your name and replace identifying details with generic labels. If the situation is genuinely serious, consider whether a human professional would serve you better.
When AI Chatbots Are Genuinely Safe to Use
This article is not a case against using AI tools. I use them regularly, and they are genuinely useful. The goal is to use them well, not to avoid them entirely.
AI chatbots are very safe for:
Research questions about general topics that contain no personal information. Drafting, editing, and improving text that has been anonymized or is fully fictional. Brainstorming ideas, names, concepts, and frameworks that contain nothing sensitive. Code help, once you replace real credentials and proprietary logic with placeholders. Learning new skills through Q and A style explanations. Planning, scheduling, and organizing tasks that do not involve private data from other people.
The guiding principle is simple. If the information in your prompt would be fine sitting in a stranger’s email inbox, it is probably fine in a chatbot. If it would not be, take sixty seconds to anonymize it first.
Limitations of This Guide
These risks apply most directly to consumer and free-tier accounts. Enterprise accounts on ChatGPT, Claude for Work, and Google Workspace have contractual data protections that consumer accounts do not. That said, even enterprise tools have had incidents, and assuming protection without verifying it is a risk in itself.
These risks also assume you are using official apps. Third-party AI tools, browser extensions that add AI features, and unofficial chatbot wrappers often have weaker or entirely absent privacy policies. According to reporting by Cybernews, multiple AI companion and chatbot apps have exposed user data at significant scale. The smaller the company, the less likely they have the security infrastructure of OpenAI or Google, and the more carefully you should check their terms before sharing anything sensitive.
Opting out of training, as covered in our guide on stopping ChatGPT, Claude, and Gemini from training on your chats, reduces the risk but does not eliminate it. Data can still be stored temporarily, reviewed for safety reasons, and exposed in breach events regardless of your training settings.
Frequently Asked Questions
What phrases should I avoid in ChatGPT?
There are no specific phrases that trigger privacy violations. The risk is not about phrases but about the type of information you include. Avoid any prompt that contains real names paired with sensitive situations, financial figures tied to your identity, passwords or keys, government ID numbers, or medical records. It is not about how you phrase it. It is about what information is in the prompt at all.
Can ChatGPT turn against you?
ChatGPT will not act against you intentionally. The risk is not that the tool becomes hostile. The risk is that information you type into it can be stored, accessed in a breach, reviewed by a human moderator, or exposed through a shared link or an indexing error. The tool itself is not the threat. The threat is what happens to your data after you press enter.
Are AI chatbots safe to use at work?
That depends on your employer’s policy and which tool you are using. Enterprise plans for ChatGPT, Claude, and Gemini have stronger data protections than personal accounts. According to IBM research from 2025, one in five organizations had already suffered a data breach tied to unsanctioned AI use. If you are using a personal account for work tasks, that is where the risk concentrates. Check your company’s AI policy before you paste anything work-related into a consumer chatbot.
What should I do if I already shared something sensitive?
Delete the conversation from your history immediately. This stops that specific text from being used in further training runs and removes it from your account. According to guidance from cybersecurity experts reported by HuffPost, deleting chat history stops exfiltration of data should anyone compromise your account. It cannot undo what was already used in a completed training run, but it limits future exposure. If you shared financial details or credentials, change them now regardless of whether you think anything happened.
Is it possible for AI to collect data without my consent?
Technically, all major platforms disclose data collection in their terms and provide opt-out controls, so consent is technically given when you accept those terms. In practice, many users do not read the terms and do not know the defaults are set in the company’s favor. In the EU, GDPR gives you the right to object to processing, request your data, and demand deletion. In the US, protections vary by state. If you believe your data has been misused, the FTC handles consumer data complaints at a federal level. In the UK, that is the ICO.
Before You Go
The most practical thing you can take from this article is one habit. Before you paste anything into an AI chatbot, pause for five seconds and ask yourself: would I be comfortable if this exact text showed up in a Google search result tomorrow?
If the answer is no, anonymize it first. Replace names with labels. Remove numbers. Cut the identifying context. Then paste what remains. You will still get the help you need. And you will have kept the information that matters where it belongs.
If you want to go further with your AI privacy setup, start by checking your data settings on the tools you use most. Our guide on how to stop ChatGPT, Claude, and Gemini from using your chats for AI training covers the exact steps, toggle by toggle. And if you use AI tools for writing, drafting, or research at work or school, our AI prompts examples guide shows you how to get strong results from fully anonymized, privacy-safe prompts.
All incidents and policy details in this article are based on verified, publicly available reporting and company documentation as of October 2026. AI platforms update their privacy policies regularly. Check each platform’s official documentation after major product updates.
