AI Glossary / Governance, ethics and safety / Beginner

Personal data

Definition

Personal data is information relating to an identified or identifiable person. Identification can be direct, such as through a name, or indirect, through information that can distinguish or connect the person in context. Legal definitions and their application depend on the relevant jurisdiction.

What it means

Personal data is broader than a list of names and contact details. An account history, a location pattern or an assessment attached to an identifiable individual can also relate to that person. Removing the name from a record therefore does not automatically remove its personal character.

AI systems may encounter this information in prompts, uploaded documents, training examples, retrieval stores, logs and generated responses. Representations such as embeddings should be assessed in context rather than assumed anonymous because people cannot read them as ordinary prose.

The widely used GDPR formulation focuses on information relating to identifiable natural persons. Other frameworks use terms such as personal information or personally identifiable information with their own definitions. Those terms are related but should not be substituted without checking the applicable scope. This entry provides a conceptual starting point, not a determination about a particular dataset.

How it works

Begin with what the information is about and how it is used. A field may describe a person directly, record an interaction or support an inference about them. Consider the combinations of attributes and other information reasonably relevant to identification, rather than examining each column in isolation.

Then map where the information travels. A support message may pass into a model prompt, a debugging log and a later evaluation dataset. Each copy or transformation can create a new processing context that needs an appropriate purpose, controls and retention decision.

Distinguish pseudonymization from effective anonymization. Replacing a name with a code can reduce exposure, but an organization holding the linking information can still connect the record to the person. Whether information is anonymous in a particular recipient's hands requires a contextual assessment. Encryption and access controls can protect personal data without making it non-personal. Review the assessment as available information, system capabilities and uses change.

Three practical examples

• An AI meeting assistant processes a transcript containing speakers' names and comments about their work responsibilities.
• A support classifier uses ticket text linked to account identifiers, even though the model input excludes the customers' names.
• A recommendation system stores inferred preferences associated with an identifiable user profile; the inferences may matter even if they are uncertain or mistaken.

Think of it this way

Think of replacing names on paper files with numbered tabs while retaining the index in a locked cabinet. The files are harder for a casual reader to attribute, but someone with the index can still connect them to people.

The analogy covers one form of pseudonymization. Digital identification can also arise through combinations of information without any single index, so removing the obvious lookup table is not a complete anonymity test.

Worked example

Why deleting one column is not enough

An invented employee-training dataset contains 200 rows with name, unique staff code, course history and instructor comments. An analyst deletes the name column and describes the result as anonymous.

The organization still has a staff directory connecting each unique code to a named employee. All 200 records remain linkable through that directory. Removing one column has reduced one obvious identifier, but it has not removed the organization's ability to identify the people.

The analyst instead describes the dataset accurately, restricts access to the linking information and assesses what the project actually needs. Further changes may reduce risk, but no fixed number of deleted columns guarantees anonymity. Free-text comments also need attention because they can contain identifying context not captured by a structured-field checklist.

Illustrative case study

Fictional case: Maple Research tests a document assistant with customer correspondence. Its preparation script removes email addresses, but the team overlooks signatures, case references and descriptions of unusual incidents in the text.

A review follows several records through the prompt store and evaluation export. The team revises the data inventory, limits unnecessary content and tests the handling of free text. It stops calling the material anonymous merely because direct email fields were removed.

The scenario illustrates classification and data-management reasoning. It does not establish that a particular redaction method is legally sufficient or that every remaining record has the same privacy risk.

Common misconceptions and mistakes

• Only secret information is personal data. Public availability does not by itself determine whether information relates to an identifiable person.
• A random identifier makes records anonymous. Linkability and context still matter.
• Personal data must be accurate facts. Opinions and inferences can also relate to identifiable people, even when disputed or incorrect.

Business use

Businesses should inventory personal data across the AI workflow, including less visible stores such as logs, evaluation files and search indexes. Ask which information is necessary for the purpose and which people or systems can access it. Do not let the model's technical format obscure the underlying relationship to individuals.

Use the relevant legal framework when deciding obligations, rights and permissible processing. A general glossary definition cannot settle those questions for every country or sector. Keep assessments documented and proportionate to the context, especially when linking datasets or introducing new uses that change identifiability or consequences.

People also ask

Are embeddings automatically anonymous?

No. Their status depends on what they represent, their links and the context in which they can be used.

Is encrypted personal data still personal data?

Encryption protects access; it does not inherently remove the relationship to an identifiable person.

Does information about a company always count?

Company information alone is different from information about a natural person, but a business record may also identify or describe individuals.

Check your understanding

A training file contains coded customer records, and the organization retains the code-to-customer table. Has replacing the names established anonymity?

Answer: No. The retained link means the organization can still connect the records to individuals. Appropriate protections may reduce risk without changing that fact.

Related terms

• Anonymization
• Privacy
• Pseudonymisation
• identifiability
• data minimisation
• personal information

Learn next

Read pseudonymization and data minimization. Trace one personal-data field through prompts, logs, storage and evaluation exports, then identify unnecessary copies and the people able to link them.

More in Governance, ethics and safety

• Opt-out - Opt-out is a mechanism through which a person declines, stops or excludes themselves from a specified activity or use.
• Out-group homogeneity bias - Out-group homogeneity bias is the tendency to perceive members of a group outside one's own as more alike than members of one's own group.
• Outcome - An outcome is a possible or observed result of a process, action or experiment. In probability, an outcome is one element of a defined sample space.
• Popularity Bias - Popularity bias is a tendency for already popular items to receive disproportionate exposure while less popular items receive little representation.
• Population Bias - Population bias is a systematic mismatch between the people represented in a dataset and the population about which conclusions or predictions are intended.
• Post-hoc explanation - A post-hoc explanation is an explanation produced by analyzing an already developed AI model or its behavior.

Sources

The Aalto Dictionary of Machine Learning, personal data: https://aaltodictionaryofml.github.io/ADictML_Main.pdf. UK ICO, What is personal data?: https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/personal-information-what-is-it/what-is-personal-data/what-is-personal-data/. ICO, Pseudonymisation: https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/pseudonymisation. Relevant returned text was reviewed; UK guidance notes ongoing legislative updates. Examples are original and jurisdictional distinctions are intentional.

Previous: Personal cloud Next: Personal knowledge base