Categories
Technology

Before Uploading Client Drafts to an AI Writer: Privacy and Export Checklist

Last updated: November 20, 2026

Before uploading a client draft to an AI writer, decide whether you are allowed to share it at all. A model-training opt-out, a vendor’s privacy badge or a paid subscription is not permission to hand an unpublished manuscript, customer list or confidential campaign to an outside service. First establish what the client permits; then check the exact provider, plan and workspace.

This practical checklist is for freelancers, editors and small agencies working with documents they do not fully control. It is not a ranking of AI writing tools and it is not legal advice. Vendor policies and controls can change, so verify them for your actual account before each new workflow.

Stop, review, or proceed: a three-way decision

Decision Situation Practical action
Stop Unpublished client draft under confidentiality rules, with no recorded permission to use outside AI Keep the original in the approved system. Ask for written authorization or use an entirely fictional sample.
Review first Client permits AI assistance, but provider, workspace and data-handling rules are unspecified Identify the actual service and plan, verify its controls and agree on what may be shared.
Proceed with minimum data Specific task, provider and permitted information have been approved Share only necessary material, document the action and verify every generated factual claim.

What the official privacy policies actually say (checked 19 November 2026)

The following table describes the specific products and plan tiers named in each provider’s documentation, not an assurance that any client contract permits the upload. OpenAI’s guidance may be updated more recently; Anthropic’s consumer retention note is dated 1 July 2026.

Product and plan Model training Concrete retention limit / nuance What to check
ChatGPT personal Free, Plus and Pro For personal accounts, model-improvement sharing is enabled by default. To opt out, turn off “Improve the model for everyone” under Settings → Data controls. Temporary Chat may be kept for up to 30 days for safety purposes; regular chat history is not the same as Temporary Chat. OpenAI’s Data controls help. For Business, Enterprise and Edu workspaces, model training is off by default, subject to workspace policy.
Claude Free, Pro and Max (including Claude Code when using those accounts) Consumer training policy: use for training with user consent, a safety-review flag or other explicit opt-in; check the privacy setting. It is not accurate to infer one universal default for all users. Deleted Claude chats are removed from back-end systems within 30 days; if improvement is enabled, de-identified training data may be retained up to 5 years. Safety flags and legal exceptions may extend retention. Anthropic consumer retention policy (1 July 2026); business/enterprise/API terms are different.
Notion AI in a Notion workspace Notion and its AI providers do not train on customer data by default, under Notion’s stated data-processing commitments. For Free, Plus and Business, AI providers retain data for 30 days or fewer. For Enterprise, many models offer zero data retention, but some model exceptions require workspace owner opt-in. Notion AI security and privacy practices; review workspace permissions and the current AI provider list.

Important: the 30-day Claude deletion statement applies when a consumer deletes a chat; it is not a claim that every active conversation is deleted after 30 days. Similarly, Notion’s provider retention statement does not prove that a client’s original workspace content vanishes after 30 days.

Seven checks before a client document leaves your computer

1. Confirm that you have the right to share it

Identify the document owner, whether the text is public, and which confidentiality or client terms apply. An author who drafted the copy may still lack permission to put it into another company’s system. If the client contract is unclear, ask for the permitted processing workflow before making the upload. Nothing in an AI vendor’s pricing page resolves that question for you.

2. Reduce the content to what the task actually needs

Look for names, account IDs, contact details, unreleased features, private rates, strategic plans and third-party information. Remove unnecessary identifiers. Be careful: text can remain identifiable through context even if you delete a client name. The safest alternative is often a fully invented example from which you can learn a writing technique without exposing real material.

Worked approach: rather than pasting an unreleased launch memo, ask the AI to improve a fictional 120-word product announcement, then apply the structural lessons by hand to the original document in the approved editor.

3. Verify training settings for the exact product and plan

For Free, Plus and Pro personal workspaces, OpenAI says model-improvement sharing starts enabled; Settings → Data controls → “Improve the model for everyone” turns it off. The official guidance also says Temporary Chats are not used for model improvement but may be retained for up to 30 days for safety. Business, Enterprise and Edu workspaces have model training off by default; never mix personal and organizational policies.

Anthropic’s consumer training documentation explicitly covers Claude Free, Pro and Max, including Claude Code on those accounts. Under its consumer retention policy dated 1 July 2026, a deleted chat is removed from the back-end within 30 days, while de-identified material used for model training with consent may remain in the pipeline for up to five years. Safety investigations and other exceptions have different terms. Verify the current settings and client permission before uploading.

Notion AI’s official security practices say that its providers do not train on customer data by default; for Free, Plus and Business, providers retain AI input for no more than 30 days. Enterprise generally uses zero-retention models, with some optional exceptions. Those statements do not replace the client’s permission or the workspace’s access rules.

4. Separate training, retention and human access

These are different questions. Does the provider use eligible input to improve models? How long does it keep chats or uploads? Which people or subprocessors might access the data for permitted operations? What does deletion actually cover? Is the storage location relevant to your client? A promise of “no training” is not a promise that content is never stored or processed.

Write down the sources and dates of the answers. If a vendor’s policy does not address a condition your client requires, do not fill in the gap with a guess.

5. Inspect connectors, sharing and workspace permissions

A permitted AI provider can still be connected to an inappropriate workspace, shared notebook or automated integration. Review who can read prompts, outputs and attachments, whether a plugin or connector forwards information, and who controls the permissions. Test access with non-sensitive synthetic text first. If your team changes administrators or tools, redo the permissions check.

6. Test a real export before trusting the chat history

A conversation transcript is not necessarily a usable client deliverable. Verify that the accepted text, its source citations, comments and any required files can be retrieved into the client’s approved system. Anthropic describes Claude export routes, while Notion explains exporting workspace content. Export functions do not guarantee the recreation of every view, permission or integration.

Maintain the approved master draft in the client’s own workspace. Preserve the decisions and factual checks independently of an AI chat.

7. Record the authorization and a fallback path

For repeat client work, record the project, approved AI purpose, provider and plan, data classes shared, policy verification date, approving person, retention questions and a deletion/export plan. This record is useful operational evidence, but is not itself legal compliance. Keep a non-AI editing process available when a permission check fails or a vendor changes its terms.

Three realistic scenarios

Case A: confidential launch draft, no permission

A marketing agency receives an unpublished announcement containing a launch date, strategic partner and private price. No contract clause or written instruction authorizes uploading it to an AI writer. Decision: stop. Editing locally or asking for explicit authorization is safer than assuming a training opt-out solves the confidentiality restriction. Deleting the company name would still leave unreleased details in the text.

Case B: already-public product copy, client agrees

A client asks for alternative headings to a public product page and approves use of a named AI assistant. Decision: potentially proceed after checking the account controls. Provide only the public excerpt needed for the task. Human-review the proposed headlines: AI output can still contain invented specifications or misleading promises.

Case C: agency approved to use a team workspace

A small agency uses Notion for editorial workflows and its client has approved that workspace. Before using Notion AI on draft material, the owner checks Notion’s published AI practices, workspace permissions, connectors and export options. Decision: follow the approved scope; the provider’s broad security statement is not a substitute for knowing who can access each document.

Copy this pre-upload checklist

  • Client approval covers the exact type of text and the external service being considered.
  • Unnecessary personal identifiers and unpublished commercial details have been removed or replaced by synthetic material.
  • Correct account, workspace and plan identified; model-improvement settings reviewed.
  • Retention, provider access, subprocessors, connectors and sharing checked independently.
  • Deliverable and sources can be exported to the client’s authorized system.
  • Policy-check date and authorization are recorded.
  • A local or fictional-example fallback is ready if any essential check fails.

Questions freelance writers ask

Does switching off model training mean I can upload confidential copy?

No. Model-training settings concern one use of eligible data. They do not replace client permission or resolve storage, access, third-party processing and contractual requirements.

Is a paid Business plan automatically acceptable?

No. Different plans may offer different terms and controls, but you still need to verify the actual service and the client’s conditions. Plan labels are not a privacy certificate.

Can I anonymize a client’s original draft?

Sometimes removing details is appropriate, but deleting names alone may leave identifiable strategy or sensitive context. A wholly fictional test brief is a stronger alternative when permission is unclear.

When should I compare features and costs?

After establishing which services you may use for the material. Our separate AI writing tools comparison addresses the later choice among eligible tools; this page is the permission and data-handling gate that comes first.

Bottom line: ask “Which data may this client share, with which exact provider, through what access controls, for how long, and can we recover it?” If you cannot answer, keep the confidential original out of the AI service.

Vendor documentation checked: November 19, 2026. This is practical editorial guidance, not an assurance of GDPR, NDA or sector-specific legal compliance. Reverify the official policies before use.

Leave a Reply