Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsDo not enter non-public information into an AI tool when disclosure could cause privacy, security, legal, contractual, safety or reputational harm—unless your organization has approved that exact service for that data classification and verified its controls. In practice, that means keeping credentials, personal and regulated records, confidential business material, and official or classified information out of public AI applications. When a task can be completed with public, synthetic, aggregated or genuinely de-identified information, use that instead.
The core rule: a prompt is a data disclosure
Typing, pasting or uploading information to an AI service is a disclosure to the service and potentially to its provider, contractors, logging systems or connected tools. The U.S. Department of Energy warns that confidential information entered into generative-AI tools may be stored or processed and could be revealed to unauthorized personnel. You therefore need to evaluate the destination before evaluating the convenience of the answer.
As an Amazon Associate I earn from qualifying purchases.
A familiar chatbot name does not establish that your information is protected. Public consumer services and approved enterprise deployments can differ in retention, model-training use, support access, data residency, tenant isolation, deletion and contractual remedies. Your employer’s policy, a confidentiality agreement, privacy law, professional-secrecy duty or sector regulation still applies even when a provider advertises security features.
Data that should stay out of public AI tools
Credentials and secrets
Never paste anything that can authenticate to a system or unlock an account. Microsoft security guidance states: “Never paste API keys, passwords, or connection strings into a prompt.” The same rule covers access tokens, private cryptographic keys, recovery codes, signing secrets, database credentials, cloud credentials and credentials embedded in configuration files or source code.
#1 Best Overall
Examples include a password you want rewritten into a stronger one, an API token copied from a dashboard, a database connection string used to debug an error, or a private key attached to a coding question. Even if the secret is short-lived, treat it as exposed once submitted: revoke or rotate it through the system that issued it rather than relying on a promise that a chat will be deleted.
Personal and sensitive personal information
Do not submit information that identifies a person or reveals sensitive facts about them. UK government guidance says: “You should never put sensitive information or personal data into these tools.” This includes identity details, contact information linked to an identifiable individual, employee or customer records, and another person’s private correspondence.
For a personal-finance reader, practical examples include a complete tax return, a bank or card statement with identifying details, a loan application, a credit report, a payroll file, or a customer list. A request such as “summarize my spending” does not make the underlying records safe to upload; the transaction descriptions, dates and amounts can expose a person’s identity, location, relationships and financial circumstances.
Health and other regulated records
Health information deserves the same exclusion. Do not put medical records, diagnoses, treatment notes, insurance claims or appointment histories into a public service. CMS guidance also identifies protected health information, personally identifiable information and financial information as categories that must not be disclosed to publicly accessible AI platforms.
The same caution applies to any record governed by a specific rule in your jurisdiction or industry, including employment, education, legal-services, insurance or child-related records. If a law or professional duty limits who may receive the information, a general-purpose public chatbot is not an approved recipient by default.
Confidential business material
Keep non-public company information out of an unapproved service. Relevant examples include source code, proprietary business logic, trade secrets, unreleased product plans, internal investigations, contracts, procurement documents, vendor terms, evaluation materials, draft policies, pricing strategies and confidential correspondence.
Microsoft advises developers to understand their organization’s policy before sending internal-systems code to an external AI service. A code snippet can contain more than code: comments may reveal an unreleased project, a variable may expose a customer identifier, and an error trace may include a hostname or token. Remove the surrounding context as well as the obvious secret.
Recommended Free Tools
Official, classified and government-sensitive information
Do not submit official, classified, export-controlled or otherwise government-sensitive material to a public AI application unless it is already public or has been cleared for publication. The UK AI Playbook states: “Never enter any official information directly into public AI applications or APIs unless it’s already publicly available or cleared for publication.” Classification labels, handling caveats and contractual restrictions continue to apply when information is copied into a prompt.
Why deletion or “not used for training” is not a complete answer
Training use is only one part of the risk. A service may retain prompts and outputs for abuse monitoring, troubleshooting, analytics, backups or legal obligations. Human reviewers or support personnel may be able to access some logs. Integrations can pass the content to retrieval systems, plug-ins or other vendors. A deletion control may have limits, exceptions or a delay.
Microsoft recommends governance for prompts, responses, memory events, retrieval events and tool outputs. That broader view matters because a sensitive fact can persist in a conversation history, a saved memory, an audit record or an external system even when a visible chat is removed. Ask what is collected, why it is collected, who can access it, where it is processed and how deletion is verified.
Public tools and approved enterprise deployments are different categories
An organization may permit a specific enterprise deployment while prohibiting the consumer version of a similar product. The relevant question is not “Is this provider secure?” but “Is this exact tenant, account, configuration and data flow approved for this classification?” The UK Playbook recognizes that some commercially licensed applications can have different data-handling terms, but requires users to verify those terms and organizational approval.
Before submitting non-public material, obtain clear answers to the following control questions:
| Control to verify | Question to ask |
|---|---|
| Permitted classifications | Which data classes may be entered, and which are prohibited? |
| Model training | Are prompts, files or outputs used to train or improve a model? |
| Retention and deletion | How long are content and backups retained, and can deletion be enforced? |
| Human and vendor access | Who can view logs for support, safety or administration? |
| Tenant isolation | How are your users, data and retrieval indexes separated from other customers? |
| Residency and subprocessors | In which countries is data processed, and which subprocessors receive it? |
| Auditability | Can administrators review prompts, outputs, memory and tool activity? |
| Incident response | What notification, containment and evidence commitments apply after a breach? |
| Contractual protection | Is there an applicable data-processing agreement and confidentiality commitment? |
If your privacy, security or data-governance owner cannot answer these questions, treat the tool as unapproved for sensitive material. Do not infer approval from a free trial, a personal account, a browser extension or an employee’s prior use.
Why removing a name may not anonymize data
Deleting a person’s name is not automatically de-identification. NIST warns that masking tools may provide insufficient protection and distinguishes direct identifiers, quasi-identifiers, re-identification, pseudonymization, synthetic data, differential privacy and governance.
Quasi-identifiers are details that identify someone in combination. An unusual job title, exact date of birth, small-town location, distinctive transaction, appointment date or sequence of events can point back to one person even when the name is gone. A spreadsheet with “Customer A” can therefore remain personal data if the surrounding facts make the customer obvious.
For a lower-risk transformation, start with the minimum information needed for the task. Replace names, account numbers, dates and amounts with synthetic values; generalize locations and time periods; remove free-text notes; and aggregate results where individual rows are unnecessary. Pseudonyms are labels, not erasure. For high-stakes work, have a privacy or security reviewer validate the transformation and the possibility of re-identification.
Best Value
A send-or-stop decision test
Run this check before pressing Send:
- Assess harm: Would disclosure harm a person, customer, patient, employee, supplier, partner or employer?
- Check obligations: Is the material covered by confidentiality, privacy, health, financial, employment, legal, export-control or government rules?
- Look for access: Does it contain anything that could grant entry to an account, network, database or signing system?
- Identify the deployment: Is this a public service, or has your organization approved this exact deployment for this data class?
- Minimize: Can the task be completed with public, synthetic, aggregated or fully de-identified information?
If any answer raises doubt, stop. Remove or replace the sensitive material and ask your privacy, security, legal or data-governance owner for an approved path. A request for “just a small sample” does not change the classification if the sample remains identifiable or confidential.
Safer ways to get useful AI assistance
Use a public template instead of a real record
Ask the tool to create a generic budget structure, contract checklist, test fixture or policy outline. Supply fictional names, invented amounts and generic descriptions. You can map the output to real records locally without disclosing those records.
Ask for transformation logic, not the underlying data
For a spreadsheet task, request a formula or pseudocode that you can run inside your approved environment. For a coding issue, describe the error pattern and provide a sanitized, minimal reproduction rather than copying an internal repository.
Aggregate before analysis
If the question concerns trends, provide totals or broad categories instead of row-level entries. Remove exact dates, rare events and free-text fields when they are not needed to answer the question.
Use an approved private workflow
Where sensitive analysis is necessary, use the enterprise service, private deployment or internal tool that your organization has authorized. Confirm that access controls, logging, retention and contractual terms match the sensitivity of the information; approval is specific to the configured workflow, not merely to a product label.
What to do after an accidental submission
- Contain access: If the message included a password, token, key or connection string, revoke and rotate it immediately through the issuing system.
- Preserve facts: Record what was sent, when, from which account and to which service. Preserve relevant screenshots or audit records according to your incident policy.
- Notify the right owners: Tell your security, privacy, legal or compliance contact promptly. They can determine whether customers, regulators, partners or an employer must be informed.
- Request provider action: Use the service’s deletion or support process, but do not assume that a deletion request removes backups, logs, downstream copies or prior human access.
- Prevent recurrence: Update your data-classification guidance, block risky destinations where appropriate, and train users on approved tools and redaction.
OpenAI’s usage policy says people’s privacy must be respected and prohibits unauthorized attempts to aggregate, monitor, profile or distribute private or sensitive information. That provider rule does not replace your own legal, contractual or organizational duties; report the event even when the service has a privacy policy.
Bottom line for everyday users
Keep secrets, identifiable personal records, health and financial information, confidential business material and official sensitive information out of public AI tools. Use only the minimum data necessary, prefer synthetic or aggregated examples, and treat anonymization as a technical process that can fail—not as deleting a name. When sensitive work is unavoidable, use an explicitly approved deployment after verifying training use, retention, access, residency, audit and contractual controls. If you cannot verify the path, do not submit the data.
Free tools Windows power users keep installed
One-click scans. No signup required.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
