Calling a dataset synthetic describes its construction. It does not tell us what privacy protection the dataset has.
Generating clinical dialogue without real patient records can remove one major source of exposure: the source records never enter that generation workflow. But synthetic data can be produced in many ways, and the method determines the privacy claim. NIST warns that many synthetic-data techniques satisfy neither differential privacy nor any other formal privacy property.
A privacy review has to ask which private data entered the pipeline, what the generator could reproduce, and which checks ran before release.
Source-data exposure and output privacy are different
The first boundary is the input. If a team generates a fictional patient and consultation without using real clinical records, it avoids collecting those records for that task. That is a meaningful reduction in exposure.
The second boundary is the output. A generator may have been trained on data its user cannot inspect. A synthetic record can also contain identifier-like strings, rare combinations of attributes, or text that should not be published. Calling the row fictional does not remove the need to scan and review it.
Formal privacy methods make a stronger, testable promise. Differential privacy, for example, defines how much a person's presence in source data can affect the released result. Meddies Consultant does not claim differential privacy, and we should not imply an equivalent guarantee.
What the Consultant release establishes
The public card describes Meddies Consultant as synthetic training data. It contains 109,005 English and 58,064 Vietnamese multi-turn consultation rows, plus QA and question-only configs. The dataset is published under CC BY-NC 4.0 and is explicitly described as a research artifact, not medical advice or deployment approval.
The card also names duplicate, parsing, structural, naturalness, empathy, framework-alignment, and safety checks. These controls help assess data quality. They are not a privacy audit.
The public repository does not disclose the full generation pipeline, source models, reviewer setup, or a privacy attack evaluation. We can verify that the publisher describes the release as synthetic. The documentation does not establish zero leakage risk.
De-identification still has a separate job
Real clinical records cannot be made synthetic after the fact by changing a few names. When a workflow must use real records, it needs an appropriate legal basis, access controls, and de-identification or another approved privacy process for the intended use.
Even a recognized de-identification standard does not mean zero risk. US HHS guidance, for example, says that properly de-identified health data retains a very small but non-zero risk of re-identification. That is a US regulatory example, not a statement of Vietnamese law, but the technical lesson is useful: privacy claims need a defined method and threat model.
Meddies PII is our research model for detecting and extracting structured identifier spans from multilingual clinical text. The model does not remove or redact those spans; a downstream de-identification process must do that work. Its public card also states that the release is not a privacy certification and does not replace validation on the data where it will be used.
What a privacy claim must document
For a synthetic clinical dataset, the construction contract should be explicit. Document whether real patient data entered the pipeline. Record the source model and generation inputs. Scan outputs for identifiers and secrets. Review rare or linkable attribute combinations. State the license and intended use. Publish the checks that were run and the failures they cannot detect.
Meddies Consultant reduces one privacy risk if its generation path does not use real patient records. Its public documentation does not yet let an outside reviewer verify that complete path or infer a formal privacy guarantee.
Synthetic generation can make privacy work easier by avoiding unnecessary access to real records. It does not make privacy disappear.