Skip to main content
Bureau of AI
Category: Privacy & Data Protection

Personal Data

Also known as: Personal Information, PII (Personally Identifiable Information)
Simply put

Personal data is any information that relates to a specific, identifiable person, whether it identifies them directly (like a name or email address) or indirectly (like an IP address that can be linked back to an individual). The exact scope of what counts as personal data depends on the applicable law or jurisdiction, so definitions can vary in practice.

Formal definition

Personal data is commonly defined as any information relating to an identified or identifiable natural person. Under the GDPR, it is framed as 'any information which is related to an identified or identifiable natural person,' and some formulations extend this to information that is 'linked or reasonably linkable' to such a person, thereby encompassing both direct identifiers (e.g., name, email address) and indirect identifiers (e.g., IP address) that permit identification when combined with other data. The precise boundaries are jurisdiction- and instrument-specific; terms such as 'personal information' and 'personally identifiable information' are frequently used interchangeably in practice but may carry distinct statutory meanings, so practitioners should verify the definition applicable to the governing legal framework rather than assume a single universal standard.

Why it matters

Personal data is the pivot on which most data protection and privacy obligations turn: whether a given piece of information counts as personal data typically determines whether a legal framework applies to its processing at all. Because the scope is defined by the applicable law rather than by a single universal standard, the same data element can trigger obligations in one jurisdiction and fall outside them in another. For organizations deploying AI systems trained on or making inferences about people, misclassifying data as non-personal can leave processing activities without the consent, transparency, or safeguarding controls the governing framework expects.

The distinction between direct and indirect identifiers is where practitioners most often err. Information such as an IP address may not name a person on its own, yet can become personal data when it is reasonably linkable to an individual in combination with other data. Under the GDPR, personal data is framed as any information relating to an identified or identifiable natural person, and some formulations extend the concept to information that is 'linked or reasonably linkable' to such a person. Treating only obvious identifiers like names and email addresses as in scope understates the breadth of the concept and can create compliance gaps.

The interchangeable use of 'personal data,' 'personal information,' and 'personally identifiable information' compounds the risk. These terms are frequently used as synonyms in practice but may carry distinct statutory meanings depending on the instrument. Assuming a single definition travels across borders or across regulatory regimes is a common source of error; the operative definition should be verified against the specific governing legal framework rather than inferred from general usage.

Who it's relevant to

Privacy and Compliance Officers
They rely on the scoping of personal data to determine which processing activities fall within a given legal framework. Correctly identifying both direct and indirect identifiers, and confirming the definition under the governing jurisdiction, is central to establishing applicable obligations.
Data Scientists and Model Developers
Teams building or training AI systems need to recognize when inputs, training data, or inferences relate to identifiable individuals, including indirect identifiers like IP addresses that become personal data when reasonably linkable to a person, so that appropriate handling controls are applied.
Legal Professionals
Counsel advising on data use must verify the specific statutory definition that applies, since 'personal data,' 'personal information,' and 'personally identifiable information' are often used interchangeably in practice but may carry distinct meanings across instruments and jurisdictions.
Auditors and Model Risk Reviewers
Those assessing data governance benefit from confirming whether an organization has correctly classified data as personal under the applicable framework, since misclassification can leave processing without the expected safeguards and controls.

Inside Personal Data

Identifiability
The core criterion in many data protection frameworks: data qualifies as personal data when it relates to an identified or identifiable natural person, whether directly (such as by name) or indirectly (by reference to identifiers that can be combined to single out an individual).
Direct identifiers
Data elements that identify a person on their own, such as full name, government-issued identification numbers, or contact details, as commonly understood in data protection contexts.
Indirect identifiers
Attributes that do not identify a person in isolation but can single out an individual when combined, such as location data, online identifiers, or a combination of demographic characteristics.
Special categories (where applicable)
Certain data protection regimes treat some categories of personal data as more sensitive and subject to heightened conditions. The exact categories and treatment vary by jurisdiction, so the specific list should be confirmed against the applicable law rather than assumed.
Relationship to a natural person
Personal data, as commonly defined, concerns information relating to a living natural person; the scope regarding deceased persons, legal entities, or households differs across jurisdictions and should not be assumed to be uniform.
AI-context relevance
In AI systems, personal data may appear in training datasets, inputs, outputs, and inferred attributes, which can raise governance and model risk management considerations around data provenance, use limitation, and downstream inferences.

Common questions

Answers to the questions practitioners most commonly ask about Personal Data.

Is personal data the same thing as personally identifiable information (PII)?
Not necessarily. Although the terms are often used interchangeably, they originate in different traditions and are not perfectly aligned. "Personally identifiable information" is a term more common in U.S. contexts and various sectoral practices, while "personal data" is the defined term used in frameworks such as the EU General Data Protection Regulation. As commonly defined in the GDPR, personal data refers broadly to information relating to an identified or identifiable natural person, which can be broader than narrower PII conceptions that focus on directly identifying fields. Because scope and definitions vary by jurisdiction and instrument, professionals should rely on the specific legal definition that applies to their context rather than treating the two labels as equivalent.
If data has been anonymized or pseudonymized, does it stop being personal data?
These two treatments are frequently conflated, but they are distinct. Pseudonymization typically replaces identifying elements with substitutes while retaining the possibility of re-identification through additional information; in many frameworks pseudonymized data is still treated as personal data. Anonymization, as commonly described, aims to render re-identification no longer reasonably possible, and truly anonymized data may fall outside the scope of some data protection regimes. However, the threshold for what counts as effective anonymization is contested and can depend on available auxiliary data and re-identification techniques. Whether a given dataset qualifies depends on the applicable legal standard and a factual assessment, so the label alone should not be relied upon.
How should we determine whether the data used to train an AI model counts as personal data?
A practical starting point is to assess whether the training data relates to identified or identifiable individuals under the definitions applicable in your jurisdiction. This typically involves data mapping and inventory work to catalog data sources, fields, and any identifiers or quasi-identifiers, and to consider whether individuals could be singled out or re-identified. Because identifiability can arise from combinations of attributes rather than a single field, the assessment should consider the dataset as a whole and reasonably available auxiliary information. Where classification is uncertain or contested, involving privacy or legal specialists is advisable rather than defaulting to a non-personal-data assumption.
Who in the organization should be accountable for identifying and classifying personal data in AI systems?
Accountability commonly spans multiple roles and lines of defense rather than resting with a single function. Business or model owners in the first line typically hold responsibility for knowing what data their systems use; privacy, compliance, or data governance functions in the second line often set policy, provide classification standards, and perform oversight; and internal audit in the third line may provide independent assurance. This is a governance question about roles, policies, and oversight, and it interacts with but is distinct from model risk management activities. Specific allocation of responsibilities depends on the organization's structure and applicable requirements.
What documentation supports the treatment of personal data in an AI system?
Documentation practices vary by framework and organization, but commonly useful artifacts include data inventories or records describing what personal data is processed, its sources, and its purposes; classification decisions and the rationale behind them; and records of any anonymization or pseudonymization applied, including the basis for concluding data does or does not remain personal data. Some regimes contemplate specific instruments such as records of processing activities or impact assessments; whether these are required depends on the applicable law and should be confirmed against the specific instrument rather than assumed.
How does personal data classification connect to ongoing monitoring of a deployed model?
Classification is not a one-time exercise, because the personal data status of a model's inputs and outputs can change as data sources, features, or use cases evolve. In practice, organizations may periodically re-review data classifications, monitor for scope changes such as new data feeds or expanded processing purposes, and reassess re-identification risk as available auxiliary data grows. It is worth noting that this monitoring addresses data governance and privacy scope, which is distinct from monitoring for model performance degradation; the two may run in parallel but track different concerns. The appropriate cadence and depth of review depend on the risk profile and applicable requirements.

Common misconceptions

Anonymized and pseudonymized data are the same thing and both fall outside personal data rules.
These are typically distinguished. Pseudonymized data is generally still treated as personal data because re-identification remains possible with additional information, whereas data considered genuinely anonymized may fall outside scope. Whether a given dataset meets the relevant anonymization threshold is fact- and jurisdiction-specific and should not be assumed.
Only obvious identifiers like names or national ID numbers count as personal data.
As commonly defined, personal data also includes indirect identifiers and combinations of attributes that can single out an individual, and in some frameworks inferred or derived data. The scope is broader than direct identifiers alone.
"Personal data" has a single universal legal definition that applies the same way everywhere.
The definition, scope, and treatment of personal data vary by jurisdiction and regime. Practitioners should confirm the definition under the specific applicable law rather than relying on a single generalized formulation.

Best practices

Confirm the definition and scope of personal data against the specific applicable jurisdiction and regime rather than assuming a universal meaning.
Distinguish anonymized from pseudonymized data in documentation, and treat pseudonymized data as personal data unless a defensible anonymization assessment shows otherwise.
Inventory where personal data appears across the AI lifecycle, including training data, inputs, outputs, and inferred attributes, to support both governance oversight and model risk considerations.
Account for indirect identifiers and attribute combinations, not just direct identifiers, when assessing whether a dataset contains personal data.
Where a framework recognizes more sensitive categories of data, verify the specific categories and heightened conditions under the applicable law before applying controls.
Document data provenance and use limitations so that downstream inferences and reuse can be traced and evaluated against applicable requirements.