Same word, different industry
On a crypto site, tokenization usually means putting an asset on a blockchain, the subject of asset tokenization. Data tokenization has nothing to do with blockchains. It is a data security technique, and it started with card numbers.
Security vendor Thales says on its tokenization page that the technique was “originally defined for payments” and swaps sensitive data for “a unique string of characters or numbers known only to the tokenization system.” Google’s cloud documentation applies the same idea to customer IDs, email addresses and other personal data, so staff and software can handle records without seeing the values in them.
Three ways to make a token
The documentation for Google’s Sensitive Data Protection service, read on November 12, 2026, lists three methods and notes that “the same cryptographic key that can generate new tokens can also reverse tokens.”
| Method | Keeps the original length and character set | Can be reversed |
|---|---|---|
| Deterministic encryption with AES-SIV | No | Yes, with the key |
| Format-preserving encryption (FPE-FFX) | Yes | Yes, with the key |
| Cryptographic hashing (HMAC-SHA-256) | No, always the same length | No |
A fourth approach needs no key, only a table. The UK’s Information Commissioner’s Office says in its pseudonymization guidance that tokens can be random numbers stored in an indexed sequence. In that case, the regulator writes, “there is no mathematical relationship between a token and an original identifier.”
Tokenization or encryption?
The difference lies in where the secret sits. The European Data Protection Board (EDPB) sorts the methods into two classes, cryptographic algorithms, where the secret is a key, and lookup tables, where the table itself is kept secret.
The PCI Security Standards Council, which sets card security rules, says in its 2011 tokenization guidance that a token derived from a card number by reversible encryption “is an encrypted PAN” and may bring extra obligations under the PCI DSS card security standard.
Systems that keep only a key, and no table, are known as vaultless tokenization. Each design moves the risk to a different place, so the question for any vendor is who holds the key or the table.
In the EU and the UK
Both regimes define pseudonymization, the legal term that covers tokenization of personal data, in Article 4(5) of the GDPR, kept in UK law as the UK GDPR. The data must no longer be linkable to a person “without the use of additional information”, and that information must be “kept separately.”
The ICO says that “pseudonymised personal data is in scope of data protection law.” It also points to section 171 of the Data Protection Act 2018, which makes it an offense to knowingly or recklessly re-identify de-identified personal data without the consent of the controller that de-identified it.
The EDPB takes the same line in its Guidelines 01/2025, adopted on January 16, 2025. Pseudonymised data that could be attributed to a person with additional information “is therefore personal”, the board writes, even when that information sits with someone else.
The mistake
The usual error is to treat a tokenized dataset as anonymous and share it freely. “Take care not to confuse pseudonymisation with anonymisation,” the ICO warns. Before sending tokenized records to a partner or a test team, check whether anyone, including you, can reverse the tokens, and record where the key or the table is held.