AI Privacy: What Tools Do With Your Data (and How to Protect Yourself)

You need to summarize a forty-page report, organize a financial spreadsheet, or draft a reply to a sensitive email. Under a tight deadline, you open the nearest AI assistant in your browser and paste the full text into the message box. That habit takes seconds and solves an immediate problem, yet it raises an uncomfortable question: where does that information go once you press Enter?
The direct answer requires clarity right away: pasting confidential company documents, identifiable client records, or trade secrets into free accounts or personal AI plans creates real privacy risks. Even after you turn off model training in your settings, your data does not vanish into thin air. It still travels to external servers and remains stored temporarily for security and abuse monitoring.
Understanding what happens behind the scenes in major assistants does not mean giving up productivity. With the right settings, clear rules on what can be shared, and simple anonymization habits, you can use AI tools with confidence and professional responsibility.
Article contents
What happens to the data you send to an AI tool
Whenever you send a message to an AI assistant, four distinct processes take place that are often confused. Telling them apart is the first step toward avoiding false assumptions about privacy.
Transient processing is the technical step in which the model reads your prompt and generates an immediate reply. Without this step, the service simply cannot run.
History and memory are features built for your convenience. The assistant saves conversations so you can revisit them later in the sidebar or let the system recall preferences across future chats.
Operational safety retention happens even when a conversation does not appear in your history or gets deleted from your interface. Timeframes depend on the interaction type and each provider’s policy: for temporary chats or after conversation deletion, technical retention usually lasts up to 30 days at OpenAI and Anthropic, or up to 72 hours at Google, subject to legal, safety, or feedback exceptions. That storage helps detect abuse, fraud attempts, and critical system failures.
Model training and improvement deserve the closest attention. By default, on many consumer-facing accounts, providers use user text, uploaded files, and ratings to train future versions of their algorithms. If confidential details enter that pipeline, they can be absorbed during training and, in rare edge cases, surface in fragments in answers generated for other users.
Another fast-growing issue inside organizations is Shadow AI. This term describes employees using AI tools at work without the knowledge, approval, or oversight of IT and security teams. The goal is usually just finishing a task faster, yet the practical outcome is exposing company assets through unvetted channels.
One frequent misconception involves paid personal subscriptions versus enterprise plans. Paying for an individual plan such as ChatGPT Plus or Claude Pro raises usage limits and speeds up access, but it does not turn your account into a business contract. Advanced legal commitments, formal Data Processing Agreements (DPAs), and contractual zero-training defaults belong exclusively to team and enterprise offerings.
How to configure ChatGPT, Claude, and Google Gemini to protect your data
Major AI platforms include privacy controls in their menus to reduce data exposure. The steps below follow official documentation checked in November 2026 and show where to find each setting (button labels may vary slightly by language, plan, and interface updates).
Privacy settings in ChatGPT
OpenAI’s assistant is the most widely used tool on the market, and you can explore its features in detail in our ChatGPT guide.
On personal Free, Plus, and Pro accounts, chats may be used to train new models by default. To stop that use:
- On desktop, click your profile name in the bottom-left corner and open “Settings”. On mobile, open the sidebar, tap your profile, and go to “Settings”.
- Select the “Data Controls” tab.
- Find the “Improve the model for everyone” toggle and turn it off.
- Close the window to save your preference.
Turning this control off prevents future conversations from entering OpenAI’s model training cycle. Your past chats remain visible in the sidebar, because storing history, maintaining memories, and training models are separate functions, as explained in OpenAI’s Data Controls FAQ.
To remove a conversation from your account, open the chat menu and choose delete (archiving only hides the thread from your main list). Permanent deletion from systems takes up to 30 days, subject to safety or legal exceptions. Pay close attention to a key detail if you upload files: deleting a conversation does not automatically delete files saved in your Library, which must be removed manually there, according to OpenAI’s chat and file retention rules.
When you need a one-off query without saving it to history, use Temporary Chat. Start a new chat and select the temporary option before sending your first message. Temporary chats stay out of your history, do not create new memories, and are not used for training, though a technical copy may be kept for up to 30 days for safety monitoring, per OpenAI’s Temporary Chat documentation.
Watch out for the feedback buttons: clicking thumbs up or thumbs down on an answer in a personal account allows OpenAI to review and use the associated conversation to improve services, even when general training is turned off, as stated in OpenAI’s model improvement documentation.
On managed commercial tiers such as ChatGPT Business, Enterprise, Edu, and direct API usage, inputs and outputs are not used to train models by default, as outlined on OpenAI’s Enterprise Privacy page.
Privacy settings in Claude
Built by Anthropic with a strong emphasis on safety and alignment, this assistant is covered in full in our Claude guide.
On personal plans (Free, Pro, and Max), users can choose whether to share conversations for model improvement:
- Open your account menu by clicking your name or initials.
- Go to “Settings” and select the “Privacy” section.
- Turn off “Help Improve our AI models”.
With that toggle off, Claude will not use your new or stored conversations in future model training runs covered by the setting, according to Anthropic’s model improvement privacy settings.
For isolated tasks, Claude offers Incognito chat. Start a new conversation outside project folders and click the ghost icon near the top of the screen. The interface highlights the border and displays an incognito notice. These threads do not enter your history, do not read conversation memories, and do not train company models, within the boundaries detailed in Claude’s Incognito chat guide.
When you delete a standard conversation in Claude, it disappears from your view immediately and is scheduled for permanent removal from servers within 30 days of deletion, subject to legal and safety exceptions. However, when model improvement is enabled or when you submit thumbs-up or thumbs-down feedback on personal accounts, de-identified data or the linked dialogue may be retained for up to five years for research and improvement, according to Anthropic’s data retention policy and personal training rules.
On commercial offerings such as Claude for Work, Team, Enterprise, and the API, customer data is not used for training by default unless explicitly opted in or submitted via feedback, as explained in Anthropic’s commercial policy. Organization administrators also have controls to disable feedback collection across the entire workspace.
Privacy settings in Google Gemini
Deeply integrated into the Google ecosystem, this model is examined in detail in our Google Gemini overview.
Data handling in Gemini includes specific details every professional should know. On personal Google accounts, chat history ties into the “Gemini Apps Activity” setting.
The most critical detail documented by Google involves human review. When activity is turned on, selected excerpts from conversations may be read and annotated by trained human reviewers to improve products and machine learning technologies. Google disconnects those excerpts from your Google Account, yet they can remain stored for up to three years, even if you delete your account activity later. For that reason, Google explicitly warns in the Gemini Apps Privacy Hub not to enter confidential information or anything you would not want a human reviewer to see.
To turn off this sharing:
- Open gemini.google.com in your browser or launch the mobile app.
- Open “Settings & help” and select “Activity”, or go directly to Gemini Apps Activity.
- Under “Gemini Apps Activity”, click the button showing “On”.
- Choose “Turn off” or select “Turn off and delete activity” to clear accumulated history, following the options in Google’s activity management guide.
Turning off activity stops future conversations from feeding model training, unless you submit feedback. However, this does not stop human review for safety and abuse prevention. Google maintains a residual retention window of up to 72 hours to deliver the service and protect its systems. In addition, notices updated in late 2026 in the Gemini Apps Privacy Hub note that these controls do not prevent interactions from being processed to create anonymized datasets that improve Google services.
Gemini also provides a Temporary Chat option next to the new chat button on signed-in personal accounts (this feature is not available on work or school accounts, per the Gemini usage guide). It stays out of recent chats and activity, does not train models, and remains subject only to the 72-hour operational safety window.
For corporate Google Workspace accounts, protection depends on the exact license your organization holds. If Gemini is included as a core service with enterprise data protections active, company prompts and files do not undergo human review and do not train external models. If access is enabled as an additional service or on plans like Workspace Individual, enterprise protections may not apply. Always verify your account’s exact license with IT using Google’s work and school account guidance.
What you can paste and what you should never send to personal chats
Not all data carries the same level of risk. Having a clear classification rule helps you make quick decisions at work without slowing down productivity.
The table below summarizes the main information categories and preventive guidance for AI tools on personal or unapproved accounts:
| Information Category | Workplace Example | Guidance for Personal Accounts | Recommended Alternative |
|---|---|---|---|
| Public and institutional data | Published press releases and public company announcements | Allowed with attention to added context | Use normally |
| Fictitious and simulated data | Spreadsheet with invented numbers to test formulas and charts | Safe for testing and learning, provided no real data is mixed in | Best practice for learning and building workflows |
| Aggregated and statistical data | Quarterly company totals without client-level breakdown | Evaluate carefully before sending | Confirm restricted figures cannot be deduced |
| Pseudonymized data | Report replacing names with codes such as Client A or Case 1 | Requires caution and judgment | Verify that exact dates, rare titles, or locations cannot re-identify people |
| Identifiable personal data | National IDs, phone numbers, email addresses, and named payrolls | Never paste into personal chats | Use an approved enterprise environment or mask all identifiers |
| Sensitive records and trade secrets | Medical charts, NDA-protected contracts, and sealed legal filings | Prohibited on personal consumer tools | Keep inside approved internal workflows or local offline models |
| Credentials and system access | Passwords, API keys, access tokens, and digital certificates | Never paste under any circumstances | Store keys in password managers and secure environment variables |
How to anonymize and prepare documents before prompting
One of the most common questions from professionals starting with AI is: “Can I paste my text into the chat and ask the assistant to anonymize the names first?”
The answer is clear: do not do that. The moment you paste real data and press Enter, that information has already been transmitted to the provider’s servers before the assistant begins generating a reply. If the account feeds training or retention pipelines, the exposure has already happened.
Data preparation and cleanup must always happen on your own computer before anything is sent.
Another vital distinction separates true anonymization from pseudonymization. In pseudonymization, you swap real names for generic labels while leaving surrounding context untouched. If a table states that the “Managing Director of the Austin branch hired in May 2024 received an 18% salary increase”, changing the name to “Employee X” hides nothing. Anyone inside the company immediately knows who that person is. True anonymization requires stripping away identifying links so the individual cannot be re-identified by reasonable means.
Before uploading spreadsheets or text documents, follow this quick cleanup checklist:
- In spreadsheets: check for hidden tabs, internal cell comments, or helper columns holding confidential figures.
- In documents and PDFs: inspect file properties to remove author names, company metadata, and tracked revision history.
- In prompt text: replace direct identifiers with neutral placeholders, generalize exact dates into broad timeframes, and group financial figures.
Compare a risky approach with a safely structured request:
Risky scenario (entirely fictitious example): “Review the consulting agreement between Alpha Construction LLC (Tax ID 12-3456789) and consultant John Doe (SSN 123-45-6789), covering a monthly fee of $6,500 for the Oakwood Condominium project in Austin.”
Safe and minimized scenario:
Act as a commercial contract drafting specialist. Review the termination clause below, identify potential legal ambiguities in the wording, and suggest improvements to balance the rights of both parties.
[Context: Consulting services agreement between two engineering firms. Financial figures and identifiers replaced with generic placeholders.]
Termination clause to review:
"Either party may terminate this agreement upon [Notice Period] written notice, whereupon [Client] shall pay all fees owed for services actually performed through the termination date."Notice how the second example gives the model the structural context needed to review wording clarity without exposing personal data, identification numbers, or business secrets. Even then, the generated output serves only as an initial drafting aid and requires qualified legal review before being used in a real contract.
AI at work and data protection laws: what professionals need to know
Across major jurisdictions, professional use of any technology processing personal data is governed by strict privacy frameworks—from Europe’s GDPR and US state privacy statutes to national frameworks like Brazil’s LGPD, whose web compliance principles we also explore in our guide on data protection and cookies.
When transferring workplace information to AI assistants, four core legal principles require close attention:
- Data minimization and purpose limitation: processing must stay strictly limited to what is necessary for a legitimate goal. There is no justification for pasting a client’s full name or tax ID when the prompt only aims to format a table or tighten paragraph clarity.
- Duty of security and governance: organizations must adopt technical and administrative safeguards to prevent unauthorized access and leaks. Employees using unapproved tools with regulated records can expose the company to compliance failures.
- Cross-border data transfers: because most major AI providers operate cloud infrastructure across international borders, sending personal records to external assistants triggers cross-border transfer rules. Organizations need to verify where processing occurs and confirm contractual safeguards match required standards.
- Special protection for sensitive data: health records, biometric identifiers, religious beliefs, and union membership receive heightened statutory protection. Pasting patient charts or medical leave notes into consumer chatbots creates severe privacy and regulatory exposure.
Regulators can impose substantial administrative penalties alongside reputational damage and lost client trust, which often cost far more than statutory fines. This discussion connects directly with ethical artificial intelligence, which advocates building and deploying technology grounded in transparency, accountability, and respect for fundamental rights.
When to turn to enterprise tools or local offline models
In many workplace scenarios, you need to analyze lengthy, complex files containing proprietary formulas or confidential clauses that cannot be easily scrubbed by hand. In those situations, the answer is not improvising in consumer chats, but choosing channels with proper technical and contractual protections.
The first option is requesting access to an approved enterprise edition of your chosen assistant. Business plans such as ChatGPT Enterprise, Claude for Work, or eligible corporate Google Workspace tiers typically operate under a formal Data Processing Agreement (DPA) with zero-training commitments. Specific controls such as detailed audit logs and centralized single sign-on vary by plan, so confirm exact contract terms before rolling out use. Even on developer APIs, zero training by default does not mean zero technical storage, as noted in OpenAI’s API data documentation.
The second option, suited for strict trade secrets or teams handling proprietary source code and patents, is running language models locally on your own hardware.
Tools such as Ollama and LM Studio let you download open-weight models and run them directly on your machine’s graphics card and memory:
- Offline execution: installing an app on your computer does not automatically make it air-gapped. You must verify that the model runs locally and that cloud connectors or web search tools are turned off so data stays inside your device, as detailed in the Ollama FAQ.
- Full control: you manage the model version, conversation logs, and data lifecycle directly.
- Hardware requirements: running capable local models requires modern hardware with dedicated GPU memory (VRAM) or high-bandwidth unified memory architectures. Physical device security, backups, and system updates also remain entirely your team’s responsibility.
A practical five-step routine before sending any prompt
To make safe habits automatic in your daily workflow, run through this quick checklist before pressing Enter:
- Check your environment: confirm whether you are signed into a personal consumer account or a company-approved enterprise workspace.
- Assess confidentiality: ask yourself whether the text in your prompt could appear in a public report without harming you, a client, or your organization.
- Minimize data upfront: send only the details required for the task and swap real names, IDs, specific figures, and brands for neutral placeholders whenever possible.
- Scrub attached files: remove hidden properties in PDFs and check spreadsheets for hidden tabs or internal notes.
- Know how to respond to accidental exposure: if you mistakenly paste confidential records into an unapproved tool, stop further messages, record which tool and account received the data, and immediately notify your security team or Data Protection Officer so containment steps can begin.
FAQ about AI privacy and workplace use
Are turning off chat history and turning off model training the same thing?
No. Saving conversations in your history and training models are separate controls in modern tools. In ChatGPT and Claude, you can disable training in settings while keeping your chat history visible in the sidebar. In Gemini, turning off “Gemini Apps Activity” stops both history saving and future model training. Neither setting removes information already incorporated into previously trained models, and providers still keep short-term safety logs (typically 72 hours at Google and up to 30 days at OpenAI and Anthropic, or longer for feedback and legal exceptions) to monitor abuse.
Do ChatGPT Plus or Claude Pro include enterprise privacy guarantees?
No. Paid individual plans offer faster replies, higher usage caps, and priority access to newer models, but they remain governed by consumer terms of service. They do not include enterprise Data Processing Agreements (DPAs), centralized audit logs, or administrative workspace controls. Organizations needing business-grade confidentiality must use dedicated team or enterprise plans.
Do human staff at AI companies read what I write?
On personal accounts with default settings enabled, that can happen. Google explicitly warns that when Gemini Apps Activity is turned on, excerpts from conversations may be reviewed and annotated by human teams and stored disconnected from your account for up to three years. Even with activity turned off, targeted human review for safety and abuse prevention can still occur within the 72-hour window. At OpenAI and Anthropic, human review happens primarily when automated safety systems flag a thread or when a user submits manual thumbs-up or thumbs-down feedback.
Can I ask the AI itself to anonymize a confidential document before summarizing it?
That practice is not recommended for confidential or regulated material. The moment you send the full text to an online assistant, the file has already been transmitted to the provider’s servers. Stripping names, replacing identifiers, and removing sensitive details must happen locally on your computer before anything is pasted online.
What happens when I delete an old conversation from my account?
When you delete a thread in ChatGPT or Claude, it disappears from your screen right away and permanent deletion from servers is scheduled within 30 days of deletion, subject to legal, safety, or feedback exceptions. In ChatGPT, remember to manually delete any uploaded files inside your Library, because deleting the chat does not remove them from there. In Google Gemini with activity enabled, deleting activity removes your personal history, but excerpts already selected and annotated by human reviewers can stay stored disconnected from your account for up to three years.
Are local AI models completely immune to data leaks?
Models running fully locally with cloud connectors turned off eliminate the risk of transmitting prompts over the internet to third-party AI vendors. However, information security then depends entirely on your computer. Outdated operating systems, malware infections, or unauthorized physical and network access to the machine still put your data at risk.



