AI tools can draft, search, translate, and organise work quickly. They can also receive prompts, files, account information, and connected-app data that you did not mean to disclose. AI privacy is therefore a workflow decision: send the minimum necessary input, choose the right account boundary, and check what the provider currently says.［4］

Key Takeaways

Send the minimum input and remove secrets or identifiers.

Check the provider’s controls for the exact account, plan, and workspace.

A VPN protects the network path, not the provider’s handling of your prompt.

Keep high-risk decisions under qualified human review.

Privacy boundary: A VPN can protect the network path. It cannot stop an AI provider from processing what you submit under its own policy.

1. AI privacy: classify the input before sending it

Start with the least sensitive version of the task. Remove passwords, API keys, identity documents, customer records, health information, private addresses, unreleased financial data, and proprietary code. Replace names and identifiers with placeholders, and keep the mapping outside the AI service.

Input	Safer first step
Public article or product note	Summarise only the section you need
Internal draft	Remove names, identifiers, and secrets
Customer, health, or financial data	Use an approved controlled workspace or do not upload it
Production code or credentials	Keep secrets out; use a reviewed local or enterprise workflow

Ask whether you have permission to share the material, whether the complete file is necessary, and whether the result could expose another person. A paid plan or a VPN does not automatically make sensitive input appropriate.

2. Check the provider controls for the exact account

Do not copy a setting from one product, plan, or workspace to another. Read the current provider page and record the account type and date you checked it.

ChatGPT

Review Data Controls, Temporary Chat, Memory, connected apps, shared links, and account security. Personal accounts and managed workspaces can have different training, retention, and administrator settings. OpenAI’s privacy guidance is the source of truth; verify the current control instead of assuming that a plan name or a deleted chat means every copy is gone.［1］

Gemini

Check Gemini Apps Activity, Keep Activity, auto-delete, manual deletion, connected apps, and personal context. A personal Google account and a managed Workspace account may follow different policies. Use Google’s current privacy and activity guidance for the interface you actually use.［2］

Claude

Separate the consumer app, API, and organisation workspace. Check Anthropic’s current privacy policy, retention explanation, account controls, and any organisation agreement. “Not used for training by default” is not the same as “not processed” or “deleted immediately.”［3］

Other tools

For image, audio, video, coding, search, or research tools, check whether prompts or uploads are public, how generated content is shared, whether connected services receive the data, and what plan or workspace controls exist. Do not infer a rule from ChatGPT, Gemini, or Claude.

3. Treat retention and deletion as separate questions

Ask four questions: what is stored, for how long, who can access it, and what deletion actually removes. History, backups, safety logs, abuse reviews, shared links, and exports may have different lifecycles. Save the provider’s current explanation with the date and scope; if it does not answer the question, mark the workflow as unverified.

Before uploading a file, remove unnecessary pages and metadata. After the task, revoke a connected app, delete a shared link, clear local downloads, and remove the source from the workspace where appropriate. Do not promise a complete erasure unless the provider’s policy supports that exact claim.

4. Keep personal and managed work separate

A work account may have administrator access, retention rules, regional storage, audit logs, or approved connectors. A personal account may have different defaults and no contractual protection for company data. Follow your organisation’s policy, use least privilege, and get approval before connecting email, drives, code repositories, or calendars.

If the task affects health, law, money, employment, identity, or safety, keep the final decision with a qualified person. AI output is a draft, not evidence of compliance or professional advice.［4］

5. Protect the network layer without confusing it with data privacy

A VPN can reduce local-network snooping and hide the device IP from the service. It cannot hide prompts from the AI provider, change account eligibility, or replace provider controls. On an untrusted network, use a trusted connection, HTTPS, device updates, and MFA; use a VPN only where lawful and useful for the network-layer risk.

Practical privacy checklist

Define the task and remove everything the task does not need.
Redact secrets, identifiers, and third-party personal data.
Check the exact provider, account, plan, workspace, and connected apps.
Record the current retention and deletion scope.
Review the output and revoke access when the task ends.
Escalate high-risk decisions to a qualified human.

A safer review workflow

Privacy is not a single switch. Treat every prompt, file, result, connected app, browser session, and workspace as a separate data path. Before using an AI service, write down what you are trying to accomplish, which data is necessary, who owns it, and what would happen if it were exposed. This small inventory makes it easier to remove unnecessary details before they reach the service.

1. Classify the material before sending it

Use a simple four-level rule:

Public: information already approved for public release.
Internal: ordinary work material that still belongs to a person or organisation.
Confidential: customer records, private correspondence, unpublished plans, source code, contracts, or non-public financial information.
Restricted: passwords, API keys, recovery codes, identity documents, health records, payment data, or information protected by a specific legal or contractual duty.

Public material may be suitable for a draft. Internal material needs a clear purpose and an approved workspace. Confidential material needs the organisation's data agreement, retention and access review. Restricted material should not be pasted into a public AI service. Redaction is not just replacing a name: dates, locations, account numbers, writing style, screenshots, and combinations of harmless fields can identify someone.

2. Check the exact account and surface

Review the settings for the account and product surface you are actually using. A web chat, mobile app, API key, team workspace, connected drive, browser extension, and shared project may expose different controls. Record which account is active, whether other people can access the workspace, whether history or activity logging is enabled, and whether connected tools can read or write data. Do not assume a setting in one surface applies to another.

3. Separate provider controls from your own controls

Look for the current privacy, data-use, retention, deletion, training, administrator, and connected-app documentation. Record the scope and date of each setting you checked. A toggle can limit one use of data without deleting existing records, removing a backup, or changing what a workspace administrator can see. A deletion request may also have exceptions or a processing delay. If the documentation is unclear, treat the behaviour as unknown and keep the input minimal.

4. Use a staged test instead of real sensitive data

When you need to understand a feature, start with a synthetic example that has the same shape but no real person, customer, secret, or production identifier. Check the response, logs or history that are visible to you, and any connected-tool action. Only then decide whether the task is appropriate for the real workspace. Never test a privacy control by uploading a live password, an unredacted document, or a real customer record.

Question	Evidence to record	Safe conclusion
Who can see the input?	Account, workspace, role, and sharing setting	The recorded surface has this access boundary
How long is it retained?	Current retention or deletion documentation	The policy states a period or leaves it unknown
Is it used for improvement?	Current data-use or training control	One documented control has the stated scope
Can a connected tool act?	Permission list and action log	The listed permission exists; review still matters
Can a VPN help?	Only network-path observations	It does not change provider-side data handling

5. Minimise and redact deliberately

Send the smallest excerpt that can answer the question. Replace names with stable placeholders, remove unique identifiers, shorten dates and locations when precision is unnecessary, and strip hidden metadata from files. Preserve enough context for the output to remain useful, but do not keep an unredacted copy in the chat merely to make comparison easier. Store the approved prompt and the redaction decision separately when the work needs an audit trail.

6. Have an incident response path

If sensitive material was sent by mistake, stop adding data, record the account and time, revoke exposed keys, remove shared access where possible, and follow the organisation's incident process. Do not rely on deleting the visible chat alone. Ask the provider or administrator what retention and deletion controls apply, and document what you could and could not verify. If the material belongs to a customer or another person, follow the applicable notification and legal process rather than making a private guess.

7. Keep the VPN boundary explicit

A VPN can protect part of the path between your device and the service, depending on the connection and local policy. It does not prevent the AI provider from receiving the prompt after delivery, change workspace permissions, remove history, control training or retention, or make a sensitive upload appropriate. Use network protection together with data minimisation, account controls, approved workspaces, and human review.

8. Keep a review record

For recurring work, keep a short record with the task, account or workspace, data classification, settings checked, source links, review date, and the condition that would trigger a new check. Revisit it when the provider changes a product name, account type, privacy policy, retention rule, model, or connected-app permission. A current timestamp is more useful than a permanent claim that a service is “private.” If you cannot determine the answer from the provider's current documentation, mark it unknown and choose a lower-risk workflow.

9. Decide whether the task belongs in AI at all

Do not use an AI service merely because it is convenient. A qualified reviewer, a local script, a document editor, or an approved internal tool may be safer for a high-impact decision or confidential record. Keep the final decision with the person who owns the risk, and record why the selected tool and data scope were proportionate.

10. Do not confuse convenience with permission

A service answering quickly or offering a convenient integration does not mean it is authorised to receive a particular material. Check the workspace purpose, your organisation's agreement, and the data owner's requirements. When in doubt, ask the responsible person before testing the boundary with a real file.

If the policy, settings, or activity record is unavailable, do not replace missing evidence with an assumption. Reduce the data, save the source and review date, and choose a workflow whose access boundaries you can confirm. Refusing a risky test is part of a correct review, not a sign that the task was done badly.

On repeat work, update the record instead of copying old confidence: platform conditions, names, and permissions can change.

The goal is not to prove that a provider is universally safe. The goal is to make a specific use proportionate to the sensitivity of the material, the people affected, and the controls you can actually verify. If those controls are missing, change the task or use another channel.

Summary

Minimise input data, verify provider controls for the exact account, separate personal and managed workspaces, and treat retention as a documented question. A VPN protects one network layer; it does not change how an AI service handles your prompt.

FAQ

Does a paid plan make an AI tool private?

No. It may offer different controls or contractual terms, but check the exact plan and workspace policy before sending sensitive data.

Does Temporary Chat or deleting a chat erase every copy?

Not automatically. Check history, memory, backups, safety logs, shared links, and retention exceptions in the provider’s current documentation.

Can a VPN stop an AI provider from seeing my prompt?

No. It protects the network path, not the provider’s application-layer processing.

Should I upload a whole company document?

Usually not. Redact it, send the smallest necessary excerpt, or use an approved enterprise or local workflow.

What should I do when the policy is unclear?

Do not upload the sensitive material. Mark the task unverified and ask your security, legal, or privacy owner.

Disclaimer: This article is general information, not legal, medical, financial, or professional advice. Follow the provider’s current terms and your organisation’s data policy.

For “Is Your Data Safe in AI Tools An AI Privacy Guide”, AethoVPN cannot replace checks beyond the network path.

Sources:

OpenAI Help Center — Data usage and privacy — https://help.openai.com/en/articles/7730893
Google Gemini Apps Privacy Hub — https://support.google.com/gemini/answer/13594961?hl=en
Anthropic — Privacy Policy — https://www.anthropic.com/privacy
NIST — AI Risk Management Framework — https://www.nist.gov/itl/ai-risk-management-framework

Sources checked 22 August 2026.

Related Articles:

Is Your Data Safe in AI Tools? An AI Privacy Guide