Anonymize a CSV for AI — in your browser
CSV exports from a CRM, help desk, or billing system are full of names, emails, phone numbers, and account IDs. Data Alias parses the file in your browser — delimiter inference, quoted fields, commas and line breaks inside cells all handled — and builds a safe copy that keeps the rows, columns, and repeated-value relationships your analysis needs.
How it works, step by step
Upload your CSV
Drop the file into the workspace. It is parsed locally in your browser tab — up to 10 MB on Free, and up to 100 MB / 200,000 rows with streaming on Pro. Nothing is uploaded to a server at any point.
Review the suggested column rules
Headers like name, email, phone, and customer_id get suggested types with matching rules — alias for people and companies, mask for phone numbers, remove for credentials. Every cell is also scanned for emails, phone numbers, IP addresses, API keys, JWTs, national ID numbers, payment card numbers, and IBANs hiding in free-text columns — the last three checked against their own checksums, so a random digit run is not mistaken for one. Suggestions are never auto-confirmed: you adjust and approve them.
Download the safe copy for the AI tool
Export a .safe.csv with the original column order, every row, and UTF-8 BOM intact — aliases like PERSON_001 instead of identities. Upload that copy to the AI tool, and keep the mapping on your machine to translate the answer back.
Example: customers.csv → customers.safe.csv
Why anonymize before you upload?
Because the answer depends on which tool you use, which account you are signed into, and what its defaults happen to be this month. A copy that never contained real identities does not depend on any of that. Here is what each vendor documents about its own handling.
ChatGPT (OpenAI)
Opting out of training is not the same as no retention
Consumer ChatGPT lets you opt out of model training via Data Controls, but OpenAI's own documentation notes retention for abuse monitoring — up to 30 days even for content you delete. An opt-out changes what training sees, not whether your pasted data reaches and briefly lives on OpenAI's servers.
Source: OpenAI Data Controls FAQ ↗Business-tier defaults are not consumer defaults
OpenAI documents that business tiers default to not training on customer data — but much real-world pasting happens in personal accounts, where the defaults and retention rules differ. A copy that never contained real identities doesn't depend on which account type someone used.
Source: OpenAI enterprise privacy ↗Claude (Anthropic)
The consumer training default changed in 2025
In August 2025 Anthropic updated its consumer terms, presenting the data-for-training choice pre-checked in favor of sharing — consumer users had to actively opt out to keep chats out of training. Defaults like this can change again; a file that only ever contained aliases is indifferent to them.
Source: Anthropic consumer terms update ↗Settings are per-account, your data policy isn't
Whether a given Claude account opted out is invisible to you when a file gets shared onward or pasted by a colleague. Anonymizing before upload is the control that travels with the file rather than with the account settings.
Source: Anthropic consumer terms update ↗Gemini (Google)
Human reviewers can read consumer conversations
Google documents that conversations in the consumer Gemini apps may be read by human reviewers to improve the service. That is a documented process, not a leak — and it is exactly why pasting a customer list into a consumer chat deserves an anonymized copy instead of the original.
Source: Google Gemini Apps privacy notice ↗A reviewed conversation should contain aliases, not identities
If a conversation can be sampled for review, the safest version of your file is one where the reviewer would see PERSON_001 and EMAIL_002 rather than real people. The analysis quality is the same; the exposure is not.
Source: Google Gemini Apps privacy notice ↗These are each vendor's own documented behaviors as of our research (July 2026). Policies and defaults change — an anonymized copy is the control that doesn't depend on them.
What to know about csv files
Aliases stay consistent across rows
The same source value always maps to the same alias within a session — John Smith is PERSON_001 in row 8 and in row 8,000. Joins, group-bys, and repeat-customer patterns survive anonymization, which is exactly what one-shot find-and-replace breaks.
Structure is preserved on export
The safe copy keeps original column order, all rows, empty cells, and proper CSV quoting, and adds a UTF-8 BOM so it opens cleanly in Excel. Numeric columns are preserved by default so totals and aggregates still compute.
Common questions
- Will the same person get the same alias in every row?
- Yes. Within a session the same source value always maps to the same alias (PERSON_001, COMPANY_002, and so on), so joins and repeated identifiers remain analyzable. With the Pro alias vault you can keep that mapping consistent across files, too.
- Does the safe copy keep my columns and rows intact?
- Yes. The export preserves original column order, every row, and empty values, quotes fields correctly, and includes a UTF-8 BOM. The file downloads as name.safe.csv next to your original — the original is never modified.
- What about sensitive values hiding in free-text columns?
- Every cell is scanned for emails, phone numbers, IPs, national ID and payment card numbers, and credential patterns — regardless of the column name. Text columns with detections get an in-place "alias detected values" suggestion, and the pre-download check warns about detected values you chose to keep.
An honest note on detection
Data Alias reduces exposure risk but cannot catch every sensitive value, and no automated detector can. That is why nothing is applied without your review: check the safe copy before you share it with an AI tool or anyone else.
No account needed. Files stay in your browser — verify it yourself.
Related guides
Every format is covered in the full guide index.