Redact your CSV: remove personal data before AI sees it

A CSV has no hidden sheets, no cell comments and no formatting. Everything in it sits in plain view, which makes it simple to clean and just as simple for an AI tool to read in full. The risk lives in the columns nobody looked at and in the notes field.

This page covers which columns go, how to swap names for tokens that line up across rows, and how to check the result. In my judgement, that check outweighs any clever token scheme. If your data starts life in a workbook, tidy the spreadsheet in Excel first and export it afterwards.

Which columns to delete, tokenise or keep

Begin with the header row, not the data. An export from a CRM, a booking system or payroll often writes every field the record holds: internal IDs, the account owner’s email, a last-contact date nobody asked for. For each column, ask whether the question you will put to the AI needs it.

Direct identifiers are the easy part. Names, emails, phone numbers, card and account numbers, national ID numbers: if the analysis runs without them, delete the column. If rows must still be told apart, swap the value for a token. The longer list of details that point to a person covers the less obvious ones.

Keeping a column “just in case” is my first cut. One that never leaves your computer needs no token, no key and no second check. Staff and health files deserve the hardest cut; for a US payroll export, what HR can pass to an AI tool is a sensible first stop.

Free-text fields: where names hide in a notes column

Column rules stop at the notes field. “Spoke to Mrs Penhaligon, her son Rafe will bring the forms” puts two people in one cell, and neither sits under a Name header. Ticket bodies, call summaries and order comments behave the same way, and they often hold the only health detail in the file.

Searching works only for details with a fixed shape. Emails, card numbers, phone numbers written with a country code and some national ID formats can be found by pattern. Names, addresses and nicknames cannot. Paste a row from your masked copy into the box below to see what still has a recognisable format.

If the question doesn’t need the notes, I’d cut the column rather than read four thousand of them. When it does, cut the file down to the rows that matter before reading. Free text is also how client details reach AI tools through routes nobody planned for.

Consistent tokens, and a lookup table kept somewhere else

A token stands in for a value: P-8304370a instead of Aurelio Penhaligon. It must be the same on every row, and in every file you send in that conversation, or the AI cannot tell that rows 4 and 19 are one customer. Totals per person, repeat visits and churn all depend on it.

NIST’s guidance on de-identifying datasets says that when a lookup table maps people to pseudonyms, the pseudonyms must be assigned at random and the table must be highly protected. So never build a token from initials, a birth date or the last digits of an ID: anyone holding the original could rebuild it.

Under the GDPR (EU and UK law) this is pseudonymisation, and Recital 26 treats such data as still relating to an identifiable person. The key, to my mind, belongs in another folder, with fewer people able to open it. Why de-identified data isn’t automatically anonymous goes into the legal side.

Does hashing a column make the data anonymous?

Hashing looks like a shortcut: run each email through SHA-256 and you get a string that never changes. That stability is the flaw. In July 2024 the FTC’s Office of Technology wrote that hashes still track a person, and that hashed emails and phone numbers can be reversed by guessing.

We ran that guess on a fictional number: a plain Python loop, one thread on an Apple M4 computer, hashed every 11-digit number starting 0770, the shape of a UK mobile, ten million of them, and found our target in under four seconds. Emails are easier still: hash a mailing list and compare.

NIST adds that an unkeyed hash generally gives no security, and it advises against hashing direct identifiers at all: one leaked key exposes the whole file. To my eye, a hash is the weaker pick where a random token would do. A keyed HMAC is a fallback for when no lookup table can be kept.

Generalising dates of birth, postcodes and ages

Remove every name and a row can still point at one person. A birth date, a postcode and a sex together can be enough. NIST calls these quasi-identifiers and measures the risk with k-anonymity: how many rows share each combination. A combination that only one row holds is a name in disguise.

The fixes are blunt and they work. Cut dates of birth to the year or a five-year band, keep the first part of a postcode and group ages above a ceiling. The US HIPAA safe harbor is one worked example: years only, three-digit ZIPs for areas over 20,000 people, everyone over 89 grouped.

For a file of a few hundred rows, I’d generalise harder than feels comfortable. Our rule of thumb, as in our Excel guide, is at least three rows per combination; it is not a legal threshold. Whether the result counts as anonymous where you work is for a lawyer to judge.

Opening a CSV in Excel without losing leading zeros

Double-clicking a CSV hands it to Excel, which guesses a type for every column. Microsoft’s support page says Excel removes leading zeros, shows long numbers in scientific notation such as 1.23E+15, and keeps only 15 significant digits, so a 16-digit card or account number has its last digit turned into a zero.

That bites twice. Customer number 000412 becomes 412 and no longer matches your key file, and if you save the masked copy from Excel, the change is written into it. Import instead: Data > From Text/CSV, then Edit, then set the identifier columns to Text, the route Microsoft documents for this.

Scripts guess too. In our test, pandas read 000412 as 412 and turned a plan literally called NA into an empty cell; with dtype=str and keep_default_na=False, both survived. Before any program re-saves the file, I’d check one row against the original in a text editor.

Masking a CSV with Python: pandas and csvkit

For a file you mask every month, a script beats hand edits: same steps each time, and a record of them. The one below uses pandas and Python’s secrets module, which Python’s docs prefer over random for tokens. Try any script on ten invented rows first; I fear a silent failure more than a crash.

It reads every column as text, drops four columns, gives each distinct name a random token, cuts the date of birth to its year and the postcode to three characters, then writes the masked copy and the key separately. Exact matching treats “A. Penhaligon” as a new person, so tidy spellings before you run it.

Without Python, csvkit does the column work: csvcut -n lists headers, csvcut -C drops columns, and csvgrep -r finds notes with Mr, Mrs or an @. It makes no tokens, so for names the script wins. If the CSV feeds an app that calls a model, masking inside the API call is the other half.

import secrets
import pandas as pd
df = pd.read_csv("clients.csv", dtype=str, keep_default_na=False)
df = df.drop(columns=["email", "phone", "account_no", "notes"])
people = df["name"].unique()
key = pd.DataFrame({
    "name": people,
    "token": ["P-" + secrets.token_hex(4) for _ in people],
})
df["name"] = df["name"].map(dict(zip(key["name"], key["token"])))
df["date_of_birth"] = df["date_of_birth"].str[:4]
df["postcode"] = df["postcode"].str[:3]
df.to_csv("clients_masked.csv", index=False)
key.to_csv("key_keep_apart.csv", index=False)

csvcut -n clients.csv
csvcut -C email,phone,account_no,notes clients.csv > clients_cut.csv
csvgrep -c notes -r "Mr|Mrs|@" clients.csv

Checking the masked file before ChatGPT, Claude or Copilot sees it

Open the masked copy in Notepad, TextEdit or any plain editor, not in a spreadsheet. What you see there is exactly what the AI receives, every comma included. Count the fields in a few rows against the header: one stray quote mark can shift a value under the wrong heading.

Then search for @, for a plus sign followed by digits, for long runs of digits, for Mr, Mrs and Dr, and for every surname in your key file. One surname left in a note undoes that person’s token everywhere. What OpenAI does with an uploaded file is the next question.

I’d keep the masked copy, dated, with the job it was made for. If anyone later asks what went to the assistant, you open a file instead of reconstructing it from memory. Retention after upload depends on the provider and plan; how Anthropic handles chats for training has its own settings.

The tool on this page, the app, or a script of your own

The redaction tool above this article takes .csv and .tsv files up to 200 KB, plus .txt, .md, .odt and .rtf. It sends the file to our server, which finds names and identifiers and returns a redacted copy with a list of what it replaced. Our security page explains how uploads are handled.

Untick a type in that list to leave it visible, then copy the result or download it. A word processor file is a different animal: an .odt or .rtf can carry edits that were only hidden, never removed, so settle those before uploading.

For a 40-row file, installing anything is overkill in my opinion. A weekly export, or one over the limit, is where the Nonimo app makes sense: the swap runs on your own computer, so the column of real names stays there. Get it from the Nonimo home page.

Sources

• FTC Office of Technology, No, hashing still doesn’t make your data anonymous. 24 July 2024: hashes stay a unique signature; hashed emails and phone numbers can be reversed by guessing.
• NIST SP 800-188, De-Identifying Government Datasets. Random pseudonyms from a protected lookup table; unkeyed hashes give no security; k-anonymity, dates and top-coding.
• Microsoft Support, Keeping leading zeros and large numbers. Leading zeros removed, scientific notation, 15-digit precision, From Text/CSV with Text columns.
• Microsoft Support, Opening CSV UTF-8 files correctly in Excel. A UTF-8 CSV opens normally when saved with a byte order mark.
• pandas, read_csv. dtype=str to keep values uninterpreted; keep_default_na=False so no strings become NaN.
• csvkit, csvcut. -n lists column names; -C excludes the columns given.
• Python documentation, secrets. Preferred over random for security tokens; token_hex.
• EUR-Lex, Regulation (EU) 2016/679 (GDPR). Article 4(5) and Recital 26 on pseudonymised data.
• EDPB, Guidelines 01/2025 on Pseudonymisation. Public consultation draft, feedback until 14 March 2025.
• 45 CFR 164.514, via the Legal Information Institute. Safe harbor: year only, three-digit ZIPs over 20,000 people, ages over 89, codes not derived from the person.

Accessed 10 November 2026.