What does a synthetic data curator actually do?

When people talk about synthetic data, everyone thinks about algorithms, generative models and privacy, but rarely in the key figure who makes it all make sense: the curator of synthetic dataThis professional profile has become essential in AI projects, advanced analytics and data spaces, because it is responsible for ensuring that this "fake" data is, at the same time, useful, realistic and compliant with regulations.

In a context where accessing quality real data is becoming increasingly difficult, and where data protection laws are becoming ever more demanding, The synthetic data curator acts as a bridge between business, technology, and legal compliance. It not only oversees how data is generated, but also decides what can be modeled, what risks exist, what analytical value is preserved, and how all of this is communicated to stakeholders so they trust the results.

What are synthetic data and why do they need curation?

Synthetic data are artificially created data sets These datasets mimic the behavior and distributions of real-world data, but without containing personal or confidential information. They are not simply random data: they are designed to preserve the structure, correlations, and statistical patterns relevant to a specific use case.

This data is mainly used for Develop, test, and validate machine learning modelsAI systems and analytics solutions are especially useful when real-world data is scarce, sensitive, or nonexistent. They are also very useful for simulating rare or extreme scenarios, such as infrequent fraud, security breaches, critical situations in autonomous vehicles, or rare clinical events.

Furthermore, synthetic data allows sharing information between organizations (for example, in public-private data spaces) reducing the risk of exposing trade secrets or violating privacy. In this way, they become a dual technology: they boost the data economy while also acting as a privacy protection tool.

To achieve this, the generation of synthetic data relies on techniques such as probabilistic modeling, simulations, decision trees or generative adversarial networks (GANs)These latter consist of two competing neural networks: one generates synthetic data and the other tries to distinguish it from real data, iteratively improving the quality of the synthesis.

The problem is that, if used naively, these methods can produce unhelpful, biased, or even potentially re-identifiable data. This is where [the solution/approach] comes in. synthetic data curationSomeone has to decide which variables are synthesized, how quality is assessed, what level of anonymization is acceptable, and whether the result actually serves the purpose of the project.

Key functions of a synthetic data curator

The role of a synthetic data curator combines technical, analytical, legal, and communication skills. Their work goes far beyond simply "pressing the data generation button": It's more like that of a content editor supported by creative AI.except that instead of texts it works with complex datasets.

One of their main responsibilities is define the use case and objectives of the synthetic dataData is not generated for each sport, but rather to address a specific need: training a risk scoring model, testing a computer vision system, releasing an educational dataset, or enabling the validation of a medical algorithm without using real medical records. The curator translates these objectives into data requirements: what variables are needed, what distributions must be preserved, and what scenarios must be able to be analyzed.

It also takes care of select and prepare the actual starting data when they exist. This includes cleaning, handling outliers, defining metadata, and exploratory analysis. Tools like MIT's SDV (Synthetic Data Vault), used in environments like Google Colab, require that the actual dataset and its metadata be well-structured in order to properly learn the relationships between variables.

Another crucial function is to determine the degree of synthesis required: fully synthetic or partially synthetic dataIn some contexts, it is feasible to synthesize only the most sensitive variables (identifiers, health data, financial information) while leaving others unchanged; in others, due to the risk of re-identification, it is mandatory that the entire dataset be synthesized. This decision has direct implications for usability and privacy.

The curator must also choose the most suitable generation techniques For each type of data: advanced resampling, probabilistic models, simulations, GANs, or combinations thereof. Synthesizing tabular customer data is not the same as synthesizing medical images, audio, sensor time sequences, or clinical texts. Furthermore, it's crucial to ensure that the selected techniques accurately capture not only means and variances, but also correlations, distribution tails, and potential temporal patterns.

Quality, usefulness and control of synthetic data

A central aspect of the curator's work is to ensure that synthetic data have real analytical valueIf the generated dataset does not allow for conclusions to be drawn similar to those that would be obtained with real data, it is not suitable for the stated purpose. This includes statistical similarity metrics, hypothesis testing, evaluation of models trained with one type of data or another, etc.

Quality refers not only to statistical accuracy, but also to the inclusion of data some diversity and relevant rare casesMany generation algorithms struggle to recreate outliers and anomalies, precisely the elements that are often critical for testing the robustness of fraud detection systems, cyberattacks, or extreme failures in control systems.

To control this quality, the curator combines automatic checks and manual checksAutomated checks allow for the verification of large volumes of data, while manual checks are used to inspect specific examples, validate that they make business sense, and detect strange patterns that an algorithm does not consider problematic but that, to human eyes, are clearly unrealistic.

However, it is always necessary to maintain a balance. quality and privacyTo prevent someone from linking a synthetic record to a real person, it is sometimes necessary to slightly degrade the accuracy of certain attributes, introduce noise, or smooth distributions. The curator must find that balance point where the dataset remains useful for analysis without creating unacceptable risks of re-identification.

In addition, the curator communicates and negotiates the level of trust in the data with stakeholders. Some may show skepticism about the relevance of results obtained with synthetic dataWhile some tend to overinterpret them as if they were a perfect representation of reality. Part of the work involves clarifying limits, assumptions, and margins of error.

Privacy, GDPR and synthetic data governance

The creation of synthetic data is not a “trick” to circumvent data protection regulations. In fact, If one starts with real personal data, the generation itself is a processing operation subject to the GDPR. Therefore, before starting, the controller must ensure that there is an adequate legal basis, that the principle of proactive responsibility is applied, and that the resulting risk of re-identification is assessed.

Within the European framework, standards such as the GDPR and the EU AI Act They demand rigorous data governance practices, especially in high-risk AI systems. This includes requirements regarding the quality of training, validation, and testing data, as well as its traceability, documentation, and human oversight. The synthetic data curator becomes a key figure in demonstrating that these requirements are met.

A basic principle is that synthetic data that is to be considered “non-personal” They must not allow the direct or indirect identification of individualsAlthough generated from data of real people, these anonymizations should only retain aggregated statistical properties and patterns relevant to the analysis. To further enhance this anonymization, additional techniques such as differential privacy or other controlled perturbation mechanisms can be applied.

The curator also evaluates whether it is better to opt for fully or partially synthetic data From a data protection perspective, partially synthetic datasets are riskier because they mix hyper-realistic records with original data, which can facilitate linking attacks if combined with other sources. Therefore, in high-risk contexts, full synthesis is generally recommended.

In any case, before releasing or sharing a synthetic dataset, the curator must carry out an assessment of anonymity and re-identification riskIf the analysis shows that high risks persist, it will be necessary to adjust the synthesis process, apply additional measures, or even resort to other Privacy Enhancing Technologies (PETs), such as strong pseudonymization, controlled access in closed environments, or homomorphic encryption.

Limitations, challenges and risks of synthetic data

Although commercial narratives sometimes present synthetic data as a kind of silver bullet, the curator's work includes to put their feet on the ground and explain their limitationsNot all data problems are solved by synthesizing them, and there are contexts in which this solution is directly inadequate.

One of the main difficulties is the large-scale quality controlManually verifying massive sets of synthetic data is impractical, and automated metrics don't always capture the business aspects that matter. This can result in datasets that appear statistically correct but don't accurately reflect the real-world dynamics of the system or market being modeled.

There are also serious technical challengesGenerating a good imitation of reality requires a thorough understanding of modeling techniques, knowing how to adjust hyperparameters, avoid overfitting, and detect when a generative model is "copying" too much of the original data. Even highly experienced teams struggle to reproduce heavy tails, complex nonlinear dependencies, or unusual interactions between variables.

In addition, there is a component of expectation management and communicationSome stakeholders may view synthetic data as "too artificial" and distrust any analysis based on it; others, conversely, may take for granted its near-perfect accuracy because the generation environment is highly controlled. The curator must clearly explain what this data can and cannot tell us.

Finally, synthetic data can introduce new biases or amplify existing ones If the generation process is not properly supervised, and if the model learns from real-world data that is already biased (for example, in credit decisions, medical diagnoses, or surveillance patterns), the synthetic dataset can consolidate those biases and make them harder to detect. The curator's task is to analyze and, where possible, mitigate these distortions.

Practical applications where the curator is essential

In sectors such as automotive, healthcare, finance, and manufacturing, the use of synthetic data is already commonplace, and The intervention of a curator is crucial for the projects to work.It's not just about generating data, but about aligning that generation with technical, regulatory, and business requirements.

In the case of autonomous vehiclesFor example, millions of different scenarios are needed to train and validate vision and decision systems: extreme weather conditions, atypical pedestrian behavior, traffic signal failures, etc. The curator defines what type of scenes are needed, how they should be distributed, what anomalies should be introduced, and how to assess whether the dataset sufficiently covers critical edge cases.

En biomedicine and genomicsSynthetic data allows working with DNA sequences, medical images, or clinical records without directly exposing patient information. The curator must ensure that relevant epidemiological and clinical patterns are preserved, that the risk of re-identification is low, and that the data remains useful for research, drug development, or training diagnostic algorithms.

En industrial quality controlsSensor readings, maintenance logs, or production data can be synthesized to train early fault detection systems. The curator collaborates with plant engineers to understand which faults are most critical, what signals anticipate them, and how to reflect those behaviors in simulated data.

In the field financial and fraud detectionThe limited availability of real fraud data (due to its rarity and sensitivity) makes synthetic data particularly attractive. The curator defines profiles of suspicious behavior, balances the rates of fraudulent and legitimate events, and validates that the models trained on this data do not generate a flood of false positives or, worse, miss actual fraud.

Synthetic data, data economics, and data spaces

Beyond specific technical cases, synthetic data plays a strategic role in the data-driven economy and the creation of shared data spacesPublic and private organizations are often reluctant to share real datasets for fear of exposing trade secrets, vulnerabilities, or sensitive personal information.

The synthetic data curator helps these organizations to design shareable versions of your dataThis approach preserves the utility for analysis and collaboration while minimizing the risk of leaking critical information. This can be key, for example, for several companies in the same sector to jointly analyze market trends, cyber threats, or systemic risks without revealing fine details of their internal operations.

In the public sector, statistical offices or educational institutions may use synthetic data to publish information useful to researchers, teachers and studentsWhile safeguarding the identity of respondents or individuals included in administrative records, the curator designs processes to ensure that this data can be used for experimentation, learning, and developing analytical skills without posing risks to the individuals involved.

In this context, synthetic data are consolidated as Dual technology: enabling new data-driven business models And at the same time, they act as a privacy-by-design mechanism. The decision to use them or not, however, is never automatic: each case requires a specific assessment of the balance between dataset complexity, modeling capacity, and the risk of re-identification.

When datasets are extremely complex, with interactions that are difficult to model or highly influential outliers, the curator may conclude that the synthesis does not offer sufficient guarantees or that it introduces misunderstandings during critical phases of development, testing, or validation. In these cases, the following must be considered: other alternative or complementary PETs instead of forcing the use of synthetic data.

Parallels with content curation and generative AI

The job of a synthetic data curator is quite similar to that of a content curator powered by generative AIIn both cases, the machine can do the heavy lifting (generating versions, condensing information, producing variations), but the responsibility for selecting, filtering, contextualizing, and validating falls on the person.

For the data, this means that the curator must formulate very precise prompts or instructions to the generation tools: which variables are key, what distributions to expect, what range of outliers to simulate, which extreme scenarios are relevant, and what level of noise is acceptable. Just as an editor gives instructions to an AI writer, the data curator "trains" the generator to work in their favor.

Furthermore, this professional must be very clear the target audience and the objectives for using that dataData science teams, compliance officers, external researchers, product developers, etc. Depending on who will use the data and for what purpose, the curator adjusts the level of detail, the diversity of cases, the format, and the associated documentation.

In the same way that a content curator divides a "mother" document into pieces for social media, newsletters, or blogs, a data curator can derive synthetic subsets specialized: one for stress testing, one for regulatory validation, one for internal training, each calibrated with the appropriate level of realism and anonymization.

Professional profile and future of the synthetic data curator

The synthetic data curator is a hybrid profile that combines Knowledge of data science, statistics, AI, digital law and communicationHe doesn't have to be an absolute expert in everything, but he does need to understand enough about each area to orchestrate multidisciplinary teams and make informed decisions.

In practice, it usually comes from environments such as data science, data engineering, data protection, business analytics, or official statisticsand complements that foundation with specific training in synthetic generation techniques, anonymity assessment, and data governance. The ability to explain complex concepts simply is almost as important as technical expertise.

As AI becomes integrated into more critical processes and regulations such as the EU AI Act gain traction, The demand for these types of profiles is going to grow stronglyOrganizations that currently rely on external consultants to generate synthetic data will tend to incorporate internal data curation and governance teams to maintain control and traceability.

In this scenario, AI does not replace the curator, but rather acts as your advanced assistantIt automates tedious tasks, proposes alternatives, and helps evaluate patterns, but the final decision about what data to use, how to interpret it, and what limitations apply remains human. That combination of judgment, ethics, and creativity applied to data is difficult to automate.

However, the synthetic data curator is becoming a strategic figure in any organization that wants to exploit the potential of AI and advanced analytics without losing sight of privacy, quality and regulatory compliance, turning "invented" data into a reliable tool for innovating, testing, collaborating and making informed decisions.