High-risk AI under the AI Act: what your AI system needs before December 2027

From 2 December 2027, AI systems classified as high-risk under Annex III of the AI Act, such as those used for credit scoring, recruitment or decisions on access to public benefits, must meet the Act’s requirements before they are placed on the EU market or put into use. Their providers have to demonstrate this through a conformity assessment. The requirements include a risk management process, properly governed data, technical documentation, human oversight, and an appropriate level of accuracy, robustness and cybersecurity. AI built into products that already fall under EU product safety law, such as many medical devices, has until 2 August 2028.

Two of those requirements depend directly on training data. Article 10 governs the data a model is trained, validated and tested on, and Article 15 governs how accurate, robust and secure the finished system is. A new Article 4a adds rules on using sensitive personal data for bias checks. Below, we cover what each of these asks for and where synthetic data helps: making training data representative, running bias checks without spreading sensitive records, producing evidence of model performance, and proving that the synthetic data itself is safe.

The new AI Act deadlines after the Digital Omnibus

The 2027 and 2028 dates were set by the Digital Omnibus on AI, which took effect on 27 July 2026 and postponed the original deadlines. The requirements themselves barely changed.

High-risk deadlines after the Digital Omnibus on AI
Type of high-risk AI system Original deadline New deadline
Standalone systems listed in Annex III, such as credit scoring, recruitment and access to public benefits 2 August 2026 2 December 2027
AI in products covered by EU product safety law (Annex I), such as many medical devices 2 August 2027 2 August 2028

Article 10: training data that represents everyone the model affects

Article 10 requires training, validation and testing data that is relevant, sufficiently representative and, as far as possible, complete and free of errors for the setting the system will be used in. Providers must also examine the data for possible bias and identify any gaps, along with how they plan to close them. In practice, the data has to reflect all the people and situations the model will deal with, including the ones that rarely occur.

Real-world data usually falls short because it mirrors the past. Common profiles dominate, while smaller groups and rare events appear too seldom for a model to learn from. Take a bank retraining its credit scoring model. Self-employed applicants make up a small share of past applications, so the model sees too few of them to learn how they actually repay. It ends up judging them by patterns learned from salaried applicants, and irregular income reads as a risk by default.

BlueGen solves this by rebalancing the data during generation. The team decides how often each group appears in the synthetic dataset, for example raising self-employed applicants from 20% to 50% of the records. At that share, the model sees enough of the group to learn its real repayment behaviour. The other fields in each record adjust with it. Income volatility and contract length follow employment type, because BlueGen generates every record as one consistent profile rather than relabelling existing rows. The extra applicants therefore behave like real self-employed people, which is what makes the rebalanced dataset defensible under Article 10.

The same approach works for rare events. Defaults make up a small fraction of most loan portfolios, and in acute kidney injury prediction, a project BlueGen carried out with TU Eindhoven, the patients who deteriorate are the minority that matters most. BlueGen generates additional examples of such cases with more variety than the originals, so the model learns the different forms an event can take instead of memorising a handful of instances. Gaps inside individual records are handled as well: missing values are filled in line with how the columns relate to each other, which supports the completeness requirement.

synthetic data rebalances an underrepresented group from 20% to 50% of records for EU AI Act Article 10.

Article 4a: synthetic data comes first for bias checks

Testing whether a model treats groups fairly often requires the attributes that data protection law guards most closely, such as ethnicity or health. Using that kind of data for bias checks is only allowed under strict conditions. The Digital Omnibus moved those conditions from Article 10(5) into a new Article 4a and extended them beyond high-risk providers to deployers and to other AI systems.

The central condition works in favour of synthetic data. Sensitive personal data may only be used for bias detection and correction if the same result cannot be achieved with other data, and the law explicitly names synthetic and anonymised data as examples of that other data. Organisations must also record why using the sensitive data was strictly necessary.

This makes synthetic data the legally preferred first step, with sensitive data as the fallback. A team that runs its bias checks on synthetic data first benefits either way. If the synthetic data does the job, the analysis goes ahead while the sensitive information stays with the data owner and never reaches analysts or model developers. If it falls short, the team has documented evidence of why, which is the justification Article 4a requires before sensitive data can be used.

an animated flow showing bias checks run on synthetic data first, with sensitive data only as a documented fallback.

Article 15: evidence that the model stays accurate and robust

Article 15 requires a high-risk system to reach an appropriate level of accuracy, robustness and cybersecurity, and to keep that level throughout its use. Providers must also declare the system’s accuracy levels and metrics in its instructions for use. Model performance therefore becomes something you measure, document and defend, including under conditions that differ from the ones the model was trained on.

For robustness, BlueGen’s conditioning lets a team build test sets for specific scenarios. They fix part of a record, such as a region or a customer profile, and the generator completes the rest consistently. The team can then check how the system performs in situations the real-world data barely covers, before it meets them in production. Conditioning builds on the patterns in the real data. The generator is strongest at filling the space between observations, and it can go somewhat beyond them, but it cannot invent behaviour with no basis in that data. Scenarios far outside the source still need domain expertise and validation against real outcomes.

For accuracy, the question is whether adding synthetic data actually makes the model better. At Schneider Electric, it raised prediction accuracy by more than 10%. A gain like that puts hard numbers behind the accuracy a provider declares.

an animated scenario test and a more than 10% accuracy gain from synthetic data at Schneider Electric.

What makes synthetic data safe, and how BlueGen measures it

Synthetic data is safe when it passes the three tests Europe’s data protection authorities use to judge whether data is anonymous, most recently restated in draft EDPB guidelines from July 2026. Only anonymous data falls outside the GDPR.

The three anonymity tests for synthetic data
Risk The question it asks Name in the EDPB’s 2026 draft guidelines
Singling out Can a record be isolated as belonging to one specific person? No Record Isolation
Linkability Can a record be connected to the same person in another dataset? No Linkage
Inference Can details about a specific person be deduced from the data? No Inference

Passing those tests is not automatic. The AI Act lists synthetic data next to anonymised data as an alternative to personal data, but a poorly trained generator can reproduce real records almost exactly. Safety has to be tested, dataset by dataset.

BlueGen builds that protection during training. Differential privacy limits how much any single person can influence the model, and federated learning keeps source data in your infrastructure when it may not leave the organisation.

It then measures the result in an evaluation report that covers both safety and usefulness. For safety, it simulates an attack for each of the three risks and scores every attack against a random guess and against an attacker holding separate real data. That comparison separates genuine leakage from the simple fact that realistic data is informative. Exact copies of real records are removed before release, and a distance test checks that synthetic records are not systematically closer to real people than real people are to one another. For usefulness, the report trains models on real data and on data enriched with synthetic records, compares their performance, and checks whether both rely on the same features. This is how gains like the one at Schneider Electric are measured, and the report doubles as the documentation Article 10 expects for every data preparation step.

That tested privacy protection also supports the cybersecurity part of Article 15. Development and test environments, and the external analysts working in them, are where sensitive data tends to spread. When those environments run on validated synthetic data, a leak exposes far less. Defences against data poisoning and adversarial inputs, which Article 15 also covers, remain a separate job.

Conclusion

The AI Act turns training data from a technical detail into something a provider has to justify during a conformity assessment. Synthetic data can carry much of that load, as long as its effect on the model and its privacy risk are measured and written down. Evidence of that kind is what separates a compliance claim from a compliance file.

Where synthetic data helps, at a glance
Rule What it asks for How synthetic data helps
AI Act, Article 10 Representative, complete training data, checked for bias and gaps Rebalances underrepresented groups, adds varied rare cases and fills missing values consistently
AI Act, Article 4a Sensitive data for bias checks only when synthetic or anonymised data cannot do the job Bias checks run on synthetic data first, with the outcome documented
AI Act, Article 15 Declared accuracy, robustness throughout use and cybersecurity Scenario test sets through conditioning, measured accuracy gains and fewer real records in development and test environments
GDPR Only anonymous data falls outside its scope Attack-based privacy tests help show whether a synthetic dataset is anonymous

The deadline is closer than it looks for teams that still need to rebuild their training data and get sign-off from a privacy officer, since both can take months. Organisations that start now can spend the remaining time improving their models instead of waiting for access to data.

To see how this works on your own data, including a full evaluation report, book a session at bluegen.ai/demo.

Share this article:

Get inspired by our cases.