03 · What You Need to Know
How Synthetic Data May Protect Privacy and Where Their Limitations Begin
Why does participant privacy matter in research data sharing?
Research datasets may contain information that participants reasonably expect to remain confidential.
Even when direct identifiers such as names or student numbers are removed, combinations of other variables may reveal an individual's identity.
For example, a dataset containing a participant's age, academic program, enrollment year, and unusual educational circumstances might make that person recognizable within a small institution.
Disclosure could expose sensitive information or create risks of stigma, discrimination, or other harms.
Researchers therefore need to balance the scientific benefits of data access with obligations arising from informed consent, institutional ethics requirements, data-use agreements, and applicable privacy law.
Synthetic data offer one possible approach to reducing direct exposure of participant-level records, but their effectiveness must be demonstrated rather than assumed.
How can synthetic data reduce privacy risks?
Synthetic-data generation creates artificial records instead of simply releasing the original observations.
Some methods learn statistical patterns from an empirical dataset and generate new records intended to preserve selected characteristics without reproducing each original participant's information.
For example, a synthetic student dataset might approximate the distribution of academic programs, engagement scores, and learning-platform usage without publishing the original questionnaire responses.
This separation can reduce direct exposure of genuine records.
However, the degree of protection depends on whether the generator has learned general patterns or reproduced information about particular participants.
A synthetic dataset that closely duplicates identifiable source records may offer little meaningful protection for those individuals.
Are synthetic data the same as anonymized data?
No. Synthetic-data generation and anonymization describe different concepts.
Synthetic generation concerns how artificial records are produced. Anonymization concerns whether information can reasonably be used to identify individuals under the applicable legal and technical framework.
A synthetic dataset may qualify as anonymous under particular circumstances, but that status does not follow automatically from artificial generation.
Synthetic data
Artificially generated records created through statistical, computational, or AI-based methods.
Anonymized data
Information processed so that individuals are not identifiable under the applicable anonymization standard and reasonably relevant circumstances.
For example, a synthetic record might reproduce a rare combination of characteristics associated with a known participant.
Although the record was generated computationally, it could still reveal information about that person.
The United Kingdom Information Commissioner's Office (ICO) emphasizes that anonymization requires consideration of identification risks in context.
Researchers should therefore assess whether synthetic outputs satisfy the relevant privacy requirements rather than using synthetic and anonymous interchangeably.
What types of privacy risks can remain in synthetic data?
Different generation methods can create different disclosure risks.
| Privacy risk |
How it may arise |
Why it matters |
| Record memorization |
A generator reproduces an original observation or distinctive combination of attributes. |
Information about a real participant may appear in the released data. |
| Membership inference |
An attacker attempts to determine whether a person's record contributed to model training. |
Participation in a sensitive study may itself be confidential. |
| Attribute inference |
Released information helps an attacker infer a sensitive characteristic about an individual. |
Private information may become more predictable. |
| Record linkage |
Synthetic outputs are combined with external information. |
Distinctive characteristics may increase identification or disclosure risk. |
| Model exposure |
A trained generator is shared or queried by others. |
Repeated access may create additional opportunities for information extraction. |
These risks are possibilities rather than inevitable outcomes. Their likelihood depends on the model, source data, release mechanism, and adversary's capabilities.
The related question of whether synthetic data can leak information about real participants requires examining these mechanisms in greater detail.
Why can rare participant characteristics increase disclosure risk?
Rare combinations of attributes may make individuals more recognizable.
For example, a university dataset might contain only one student enrolled in a particular program, age category, and scholarship arrangement.
If a synthetic generator reproduces that distinctive combination alongside sensitive information, an external observer with relevant background knowledge might infer something about the original participant.
The risk does not require the release of a name or student identification number.
However, rarity alone does not prove that a synthetic record corresponds to an identifiable individual.
Researchers should assess whether the released information meaningfully increases the likelihood of identifying someone or learning confidential information about them.
Does removing names before synthetic-data generation guarantee privacy?
No. Removing direct identifiers can reduce some risks, but it does not necessarily eliminate information that could identify participants indirectly.
Demographic characteristics, dates, locations, educational histories, or unusual combinations of variables may remain informative.
Furthermore, the synthetic-data generator itself may process sensitive source information during training.
Researchers should therefore consider privacy throughout the workflow, including source-data preparation, model training, output generation, and release.
Removing names is one safeguard, not a complete privacy evaluation.
Can differential privacy strengthen synthetic-data protection?
Potentially. Differential privacy is a mathematical framework for limiting how much the output of a computational procedure can reveal about the presence or absence of an individual's information.
When correctly implemented, differentially private synthetic-data generation can provide a formal privacy guarantee associated with the specified mechanism and parameters.
However, differential privacy does not guarantee that the synthetic dataset is statistically accurate or suitable for every research task.
Privacy guarantees also depend on the privacy budget, implementation, composition of repeated releases, and other relevant assumptions.
Researchers should distinguish a formally differentially private generation method from a tool that merely claims to produce privacy-preserving synthetic data.
The foundational work of Dwork and colleagues provides the mathematical basis for differential privacy, while subsequent research has investigated its application to synthetic-data generation.
What is the privacy-utility tradeoff?
Synthetic datasets are often evaluated according to both their usefulness for analysis and the privacy risks they create.
A generator designed to reproduce source records very closely may preserve important analytical characteristics, but close reproduction can also raise disclosure concerns.
Conversely, stronger privacy protections may reduce the fidelity of certain statistical relationships or rare cases.
This does not mean privacy and utility always move in opposite directions by the same amount. Their relationship depends on the generation method and intended application.
Researchers should evaluate the tradeoff using criteria appropriate to their data and research purpose.
A synthetic dataset may be sufficiently useful for software testing while being inadequate for estimating subgroup effects or developing a prediction model.
Can synthetic data be shared publicly?
Potentially, after appropriate assessment.
Public release increases the range of people who can access and analyze the data, including individuals with external information or capabilities not anticipated by the research team.
Researchers should therefore evaluate disclosure risks under the intended release conditions.
A dataset considered suitable for restricted access among approved collaborators may not be appropriate for unrestricted publication.
Data-use agreements, licensing conditions, repository requirements, and institutional policies may also affect whether release is permitted.
Researchers should not assume that artificial records can be published freely simply because the dataset contains no original rows.
Can synthetic data be used in educational research?
Yes. Educational research may benefit from synthetic datasets when access to genuine student information is restricted.
For example, artificial learning analytics records could help researchers develop dashboards, test data-processing pipelines, or demonstrate statistical methods without distributing original student-level activity logs.
However, educational records can contain sensitive information about academic performance, attendance, disability accommodations, or personal circumstances.
A generator trained on such records may retain patterns that create disclosure risks.
Researchers should assess whether the synthetic outputs protect confidentiality and whether the intended analyses remain valid.
For studies making claims about actual students, synthetic data should not automatically replace appropriate empirical evidence.
Does generating synthetic data require ethical approval?
Requirements depend on the source information, research activities, jurisdiction, and institutional rules.
A synthetic dataset generated entirely from hypothetical assumptions may not involve human-participant data processing.
However, training a generator on identifiable or confidential research records may constitute further processing of participant information.
Such processing may be subject to consent conditions, ethics approval, data-use agreements, or privacy law.
In the Philippines, researchers should consider the Data Privacy Act of 2012 and relevant National Privacy Commission guidance when processing personal information.
The fact that the intended output is synthetic does not automatically exempt the underlying processing activity from applicable requirements.
Can researchers upload confidential participant data to a public AI system to generate synthetic records?
Not without establishing that the disclosure and processing are authorized and adequately protected.
Uploading participant-level records to an external AI service may expose information to a provider operating under terms, retention arrangements, and security controls that differ from the research institution's requirements.
Even if the requested output is synthetic, the input may still contain genuine personal or confidential information.
Researchers should examine institutional approval, contractual protections, data-processing arrangements, and applicable privacy requirements before using an external service.
Where appropriate, an institutionally approved local or controlled environment may offer stronger governance over sensitive source data.
The intended privacy benefit of synthetic-data generation does not justify an unauthorized disclosure during generation.
How should researchers evaluate privacy before releasing synthetic data?
Privacy assessment should consider the source dataset, generation method, output characteristics, intended recipients, and reasonably foreseeable misuse.
Researchers may examine whether synthetic records closely resemble original observations, whether distinctive individuals could be exposed, and whether the generator is vulnerable to relevant inference attacks.
The assessment should also consider the utility required for the intended scientific task.
No single similarity score or privacy test establishes safety under every circumstance.
Researchers should document the evaluation approach and any residual risks.
Where the data are highly sensitive or the consequences of disclosure are substantial, independent privacy or institutional review may be appropriate.
Can synthetic data eliminate the need for access controls?
Not necessarily.
Even when a synthetic dataset presents lower disclosure risk than the original records, access restrictions may remain appropriate.
For example, a synthetic dataset derived from highly sensitive clinical or educational records may be shared only with approved researchers under specified conditions.
Access controls can complement generation methods by limiting who can obtain the data and how they may use them.
Other safeguards may include data minimization, controlled computing environments, disclosure review, and restrictions on redistribution.
Researchers should select controls based on the assessed risk rather than assume that synthetic generation removes all confidentiality obligations.
How should privacy-preserving synthetic data be reported?
Researchers should explain why synthetic data were generated and what privacy objective the method was intended to achieve.
Relevant documentation may include the generation approach, categories of source data, privacy safeguards, evaluation procedures, and limitations.
If differential privacy was used, the relevant mechanism and privacy parameters should be reported at a level sufficient for appropriate evaluation.
Researchers should avoid unsupported claims that a dataset is fully anonymous, impossible to re-identify, or entirely free from disclosure risk.
They should also describe the analytical utility of the generated data where that utility is relevant to the study.
The broader requirements for reporting synthetic data in research papers apply alongside privacy-specific documentation.
