What Counts as PII? Examples, Grey Areas and AI Chats

PII stands for personally identifiable information. It sounds like something only a compliance team needs to worry about, but the question underneath is one anyone can ask: could this be used to work out who someone is?

It matters more now that so much text gets pasted into AI tools. Before you can clean a document, you need to know what you're looking for, and the answer is broader than "names and social security numbers".

How the rules define personal information

Different laws use different words for roughly the same idea. In the US, the term is PII, and NIST guidance (SP 800-122) describes it as information that can be used to distinguish or trace a person's identity, plus anything linked or linkable to them, such as medical, financial or employment records. Canada's PIPEDA keeps it short: information about an identifiable individual. Europe's GDPR is the broadest:

“"Personal data" means any information relating to an identified or identifiable natural person. GDPR, Article 4(1)”

The important word in all three is identifiable. Data doesn't need a name on it to count. If it can be connected to a specific person, with reasonable effort, it's personal.

Direct identifiers: the obvious ones

These point at one person on their own. Nobody argues about them:

• Full name, especially next to anything else
• Personal email address and phone number
• Home address
• Government numbers: Social Insurance Number, Social Security number, passport, driver's licence, health card
• Bank account, IBAN and payment card numbers
• Photos of a face, fingerprints and voice recordings
• Account usernames that are tied to a real person

Indirect identifiers, and why they add up

The less obvious category is information that doesn't identify anyone by itself but narrows things down fast when combined: birth date, gender, postal code, job title, employer, the school someone's child goes to. Privacy researchers call these quasi-identifiers.

How fast do they narrow it down? In 2000, Latanya Sweeney at Carnegie Mellon used US census data to show that a five-digit ZIP code, gender and full date of birth were enough to single out 87% of the US population. Three fields that look harmless on their own, and most people are unique.

The lesson for everyday use: when you strip names from a document but leave someone's birthday, postal code and job title, you may have removed far less than you think.

The grey areas people argue about

• IP addresses. In Europe, the Court of Justice ruled in the Breyer case (2016) that even a dynamic IP address can be personal data when a website operator has a legal way to link it to a subscriber. In a server log, next to timestamps and page views, treat it as personal.
• Work email addresses. priya.raman@company.com points at one person, so it counts. A shared inbox like support@company.com usually doesn't.
• Job titles. "Account manager" at a bank with 80,000 staff tells you nothing. "Head of finance" at a 12-person startup is one specific person.
• Device and cookie IDs. They don't contain a name, but their whole purpose is to recognise the same person over time, and GDPR lists online identifiers explicitly.
• Free-text descriptions. "My neighbour, the retired judge on the corner" contains no name, no number and no address, and still identifies someone to anyone who lives on that street.

When unsure, ask the original question again. Could someone who knows the context figure out who this is? If they could, handle it as personal data.

Redacting is not the same as anonymising

This is where tools help and where they stop. A redactor can reliably find details that have a format, and our PII Redactor catches names from context as well. What it can't do is understand that a description points at one person. Run this note through it and see what's left:

The patient is protected here, because her name and birth date are gone. The referring dentist is not, and in a town that size neither might the patient be. Proper anonymisation, the kind used before data is published for research, involves generalising details (a birth year instead of a date, a region instead of a town) and checking that no combination is unique. That's a much bigger job than cleaning an email before asking ChatGPT for help.

A practical rule for AI chats

For day-to-day work you don't need a legal opinion on every field. Remove the direct identifiers, skim for descriptions that point at one person, and leave in what the AI actually needs to do the task. An order number, a product name or a date of an event usually stays. A name, an email and a birth date usually go.

If you want a checklist, what not to paste into ChatGPT has one. If you'd like to keep using real emails and documents with AI and get the names back afterwards, the walkthrough on redacting personal information before using AI shows the whole process.