ClaudeAI & privacy

Claude privacy: remove personal data before it reaches Claude

Claude privacy comes down to two questions: what Anthropic does with your conversations, and what you upload in the first place – long contracts, case files, project knowledge. This guide summarizes Anthropic's official policies on training and retention, the settings worth changing, and how to anonymize documents with placeholders before they reach Claude.

AI assistantNew chat
Microsoft Word

Tenancy_agreement_anonymized.docx

DOCX · 4 pages · 14 placeholders

Summarise [PERSON_2]'s obligations under this tenancy. When can [PERSON_1] end it at the earliest?

Summary for the tenant [PERSON_2]:

• Rent £1,050 per month by standing order to [IBAN_1]

• Deposit £1,210, protected within 30 days

• Keep the flat in good repair (clause 9)

The landlord [PERSON_1] can serve a section 21 notice no earlier than 4 months into the term, with at least 2 months' notice.

Draft a short question to [PERSON_1] about using the storage unit.

Dear [PERSON_1], clause 1 includes storage unit 4. May I also keep bicycles there? Kind regards, [PERSON_2]

Message AI assistant …

Only placeholders leave your device – the assistant never sees the real names.

Your formats and the AI tools you work with afterwards

  • Word
  • Excel
  • PowerPoint
  • Outlook (emails)
  • PDF
  • Scans & photos
  • TXT Text & CSV
  • ChatGPT
  • Claude
  • Microsoft Copilot
  • Gemini
  • Mistral
  • DeepL
  • Perplexity

Key takeaways

  • Consumer plans (Free, Pro, Max): your Model Improvement setting decides whether chats are used for training. Incognito chats are never used, according to Anthropic.
  • Commercial plans (Team, Enterprise, API): Anthropic says it does not train on your inputs or outputs by default, unless you send feedback or opt in.
  • With Model Improvement on, chats may be kept in de-identified form for up to five years; deleted chats leave Anthropic's back-end within 30 days.
  • Thumbs up/down feedback sends the whole conversation and is stored for up to five years – avoid it on sensitive chats.
  • For client, patient or HR data, the safest approach is to anonymize files before upload – ideally with consistent placeholders like [PERSON_1].

What to remove before working with Claude

  • Names of people and companies in contracts, reports and correspondence
  • Contact details: addresses, email, phone numbers
  • Bank and tax data: account numbers, IBAN, credit card, tax ID
  • ID card, passport and social security numbers
  • Case, customer and employee numbers
  • Comments, tracked changes and document properties that contain author names
  • Files that will stay in Claude Projects as project knowledge

How it works

  1. 1

    Clean the files

    Upload the documents to Diskretio and pick the placeholder strategy. Long PDFs are processed as a whole, within your plan’s page limits.

  2. 2

    Upload to Claude

    Add the cleaned files to the chat or as project knowledge. Placeholders like [PERSON_1] stay consistent across the document, so Claude can follow who is who.

  3. 3

    Map the result

    Claude analyzes structure and content; you reinsert real names only where the output is shared.

Does Claude train on your data?

It depends on your plan. For the consumer Free, Pro and Max plans, Anthropic says it uses chats and coding sessions to improve Claude when you allow it in your privacy settings (the Model Improvement setting), when a conversation is flagged for safety review, or when you explicitly opt in, for example through a testing program. Incognito chats are not used, even if Model Improvement is on.

For commercial products – Claude for Work (Team and Enterprise), the Anthropic API and Claude Gov – Anthropic states that it does not use your inputs or outputs for training by default. The exceptions are feedback or bug reports you explicitly submit, or an opt-in.

One detail people miss: when you rate a response with thumbs up or down, Anthropic stores the entire related conversation for up to five years. It is de-linked from your user ID before use and may be used for training – regardless of your Model Improvement setting.

PlanUsed for training?Typical use
Free, Pro, MaxOnly with Model Improvement on, safety flags or feedbackPersonal use
Free, Pro, Max – incognito chatNoOne-off sensitive questions
Team, Enterprise (Claude for Work)Not by default, except feedback or opt-inCompanies, firms, teams
Anthropic APINot by defaultYour own apps and automations

How long does Anthropic keep your Claude chats?

According to Anthropic's privacy help center, for consumer plans: with Model Improvement on, new or resumed chats may be kept in de-identified form for up to five years in training pipelines. If you switch it off, future chats are no longer used for training, but data already in training runs or trained models stays there.

Deleted chats disappear from your history immediately and are removed from back-end storage within 30 days; they are also excluded from future training. Conversations flagged for usage policy violations may be kept for up to two years, and classification scores for up to seven. Feedback is retained for five years. Commercial customers have separate retention terms set out in their own help article and the commercial terms.

Policies as checked on Anthropic's help pages in November 2026. Anthropic updates them from time to time – check privacy.claude.com before starting a sensitive project.

Who can see your Claude conversations?

According to Anthropic's help center, if you allow your chats to be used to improve Claude, access is limited to a small number of personnel involved in model training, who see the data only after it has been separated from your account identifiers. Conversations flagged by safety classifiers follow a different path and may be used for trust and safety work, such as enforcing usage policies. In Team and Enterprise plans, organization admins also control data settings, such as whether members can submit feedback (the Rate chats setting under Organization settings › Data and Privacy).

That is another reason to keep real names out of sensitive material. If a conversation is reviewed, stored or leaked, an anonymized document exposes far less than the original.

Is Claude GDPR compliant?

There is no blanket answer – it depends on plan, contract, data and purpose. For businesses, Anthropic's commercial terms incorporate a Data Processing Addendum with standard contractual clauses for international transfers, as described for example by the German law firm Heuking (June 2026). For professional use with personal data, a commercial plan is generally the right setting, not a personal Pro account.

Regulated professions face a second layer. The ICAEW, for instance, warned in November 2026 that entering client data into public AI tools can risk unauthorized disclosure and “can be akin to posting it publicly” if the tool isn't in a secure environment. Germany's Federal Chamber of Tax Advisers says personal and identifying data must be anonymized or pseudonymized unless the AI provider is bound to confidentiality.

General information, not legal advice. Check your specific use case with your data protection officer or compliance team.

Claude privacy settings checklist

Whatever you upload, a few settings make a difference:

  • Turn off Model Improvement in your privacy settings if you don't want chats used for training
  • Use an incognito chat for one-off sensitive questions
  • Don't rate sensitive conversations with thumbs up or down
  • Delete chats and projects you no longer need
  • On Team or Enterprise, review whether members may submit feedback
  • Write down which documents may only go into Claude after anonymization

How to anonymize documents before uploading them to Claude

Manually, the approach is: in Word, use Find and Replace (Ctrl + H) to swap each name for a placeholder, check headers and footers, delete comments, accept all tracked changes and run Inspect Document to strip authors and properties. PDFs need proper redaction in Acrobat Pro or a free tool. That works for a short letter.

It breaks down with the material Claude is typically used for: contract bundles, litigation files, years of meeting minutes. The longer the document, the more people and spellings appear, and one forgotten name in an appendix is enough. Diskretio detects names, companies, addresses, bank details, tax and ID numbers, case numbers and more automatically in Word, Excel, PowerPoint, PDF and scans, and removes comments, tracked changes and metadata in the same step.

StrategyExampleWorks well with Claude?
Black redaction█████ resigned on 3 MayLimited – Claude can't tell who is meant
Placeholders[PERSON_1] resigned on 3 MayYes – roles and references stay clear
Fake data (Pro plan)Jonas Berger resigned on 3 MayYes – reads naturally, good for examples and tests

Claude Projects, long context and the API

Files added as project knowledge stay in the project until you remove them, not just for a single chat. Clean them before you add them. With long documents, consistency matters: Diskretio plans replacements across the whole document, so a person gets the same placeholder whether they are named in full or by surname only. The mapping is never stored; putting real names back happens on your side only.

For developers: if you call Claude via the API, anonymize text through our REST API first (Pro plan and above). An MCP connection that lets Claude clients call anonymization as a tool is in preparation.

FAQ

Questions. Answered briefly.

Try it with your own document

10 pages per month are free. No subscription, no credit card.