KG LEGAL \ INFO
BLOG

Synthetic data generation: the technology and its legal framework in the EU and Poland

Publication date: November 16, 2026

Legal status as at 11 November 2026

Synthetic data, meaning text, images and tabular records produced by algorithms rather than collected from the real world, has moved from the research lab into everyday business use. It promises datasets that can be shared, tested and used to train AI without exposing real individuals. The law, however, does not accept that promise at face value. In 2026 the European Data Protection Board stated that synthetic data is not anonymous by default, the AI Act’s transparency duties for synthetic content began to apply, and Poland established its own AI supervisory authority. This article explains how synthetic data is generated and what legal conditions its generation and use must meet.

What synthetic data is and why the method matters legally

The term covers three quite different families of techniques, and the legal analysis depends on which one is used.

  • Rule-based (mock) data. Values are drawn from dictionaries, formats and random generators, without any real dataset as input. Libraries such as Faker and Mimesis, or older utilities such as pydbgen, produce names, addresses, dates or registration numbers for software testing.
  • Simulation. Data are produced by a model of the underlying process, for example Monte Carlo sampling or agent-based modelling (the Mesa framework), where simulated agents interact according to defined rules.
  • Data-derived synthesis. A generative model is trained on a real dataset and then sampled to produce new records that preserve the statistical properties of the original. This is the family that covers large language models, image generators and tabular synthesisers.

The first two families usually raise few data protection questions, because no real person’s data enters the process. The third does: a model that has learned from personal data, copyright works or protected databases carries legal baggage from its training set, and the output may carry it too.

Generating text: from Markov chains to large language models

Natural language processing (NLP) has two branches: natural language understanding (NLU), which extracts meaning from text or speech, and natural language generation (NLG), which produces coherent text from data or a prompt. Both must respect the rules of syntax, vocabulary, morphology and semantics, and generated text should select the relevant information and present it in a logical structure. Typical uses include chatbot and voice assistant responses, suggested replies in e-mail, product descriptions for e-commerce, machine translation, summaries and narrative reports built from financial, business or Internet of Things data, transcription, and paraphrasing.

The methods have developed in four steps.

  • Templates and Markov chains. The earliest systems filled the gaps in predefined text templates with data. A Markov chain is a stochastic model in which the probability of the next state depends only on the current one; applied to text, it predicts the next word from the last word entered, which is how early smartphone keyboards suggested words.
  • Recurrent neural networks (RNNs). An RNN processes a sequence element by element and feeds each output back in as input, so that earlier words are remembered when predicting the next. As sequences grow longer, training breaks down because of vanishing and exploding gradients: the weight updates propagated back through many time steps shrink towards zero, and the network can no longer learn long-range dependencies or produce long, coherent passages.
  • Long short-term memory (LSTM) networks. A variant of the RNN with three gates (input, forget and output) that decide which information to keep, discard and pass on. Selective memory reduces the vanishing gradient problem, but the architecture is computationally demanding and in practice retains context over a few hundred words.
  • Transformers. Introduced in 2017, the Transformer combines an encoder-decoder structure with an attention mechanism that relates every word in a sequence to every other word, regardless of distance. Each word keeps its own vector representation instead of the whole sentence being compressed into a single fixed-length vector, and all words are processed in parallel. This made it possible to model long-range dependencies efficiently and is the basis of all leading language models.

The best-known Transformer models illustrate the main design choices.

  • GPT (generative pre-trained transformer), developed by OpenAI, uses the decoder blocks of the Transformer and predicts the next word from all the preceding ones; the extended sequence is fed back in until the text is complete. GPT-1 (2018) showed that a model pre-trained on unlabelled text could be fine-tuned to specific tasks through transfer learning. GPT-2 and GPT-3 scaled up the data and the number of parameters; GPT-3 was trained on, among other sources, the Common Crawl web corpus.
  • BERT (bidirectional encoder representations from transformers), developed by Google, uses the encoder. It was trained on BooksCorpus and English Wikipedia with two unsupervised objectives: masked language modelling (MLM), in which randomly hidden words are predicted from the context on both sides, and next sentence prediction (NSP). Lighter or optimised variants include DistilBERT, ALBERT and RoBERTa, and BERT underlies many adaptations to Polish.
  • XLNet, from Carnegie Mellon University and Google Brain, combines BERT’s bidirectional context with autoregressive training. Through permutation language modelling it learns from all possible orderings of the words in a sentence, avoiding the artificial masking of inputs, and builds on the Transformer-XL architecture.
  • T5 (text-to-text transfer transformer), from Google, recasts every NLP task as converting one text string into another, with a prefix indicating the task. It was released in sizes from 60 million to 11 billion parameters and trained on the C4 corpus, a cleaned version of Common Crawl of about 750 GB.
  • BART, from Facebook AI (now Meta AI), is a denoising autoencoder with a bidirectional encoder and an autoregressive decoder. It learns to reconstruct text corrupted by token masking, token deletion, text infilling, sentence permutation or document rotation, and performs well in summarisation, translation and question answering.

Since GPT-3, the GPT-4 generation and its successors, together with open-weight model families, have made fluent, multilingual and multimodal generation a commodity. The remaining weaknesses are no longer stylistic but factual: models can produce plausible but false statements and can reproduce fragments of their training data. For synthetic data purposes, language models are used to generate training dialogues, test documents and paraphrases, and to augment scarce datasets, for example in less widely used languages, where progress has historically lagged behind English.

Generating images: autoencoders, GANs, Transformers and diffusion models

Image generation has become accessible to anyone with a phone and an internet connection. Text-to-image generation combines language understanding with image synthesis: the model must interpret the prompt, map its terms to objects and render shapes, colours and overlapping elements consistently. Synthetic images are widely used to supplement, diversify and balance training sets, including in medical imaging, and in super-resolution, where the resolution of an image is increased while its detail is preserved.

  • Variational autoencoders (VAEs). An encoder compresses input images into a latent representation and a decoder generates new images from it. VAEs train quickly and are useful for reducing detail or detecting anomalies, but they render fine detail poorly; today they mostly serve as components of larger systems.
  • Generative adversarial networks (GANs). A generator produces images from random noise and a discriminator learns to tell them apart from real ones; the two networks improve by competing in a zero-sum game until the discriminator can only guess. GANs are powerful but hard to train and prone to mode collapse, where the generator produces only a narrow range of outputs. Conditional GANs (cGANs) add auxiliary information, such as a class label, to steer generation.
  • CycleGAN translates images from one domain to another without paired training examples, by learning two mappings that are inverses of each other. Uses include style transfer, changing the season in a photograph and turning paintings into photographs.
  • StyleGAN, published by Nvidia in 2018, uses a mapping network and a synthesis network to control the style of the generated image at several levels of detail, and is known for highly realistic faces.
  • Transformer-based models. Image GPT (iGPT) from OpenAI treats an image as a sequence of pixels and completes it in the same way a language model completes a sentence, with impressive results at a high computational cost. The first DALL-E, a 12-billion-parameter model trained on 250 million image-text pairs, combined a Transformer decoder with elements of a variational autoencoder to generate images from text descriptions.
  • Diffusion models learn to reverse a gradual noising process and generate an image from pure noise by step-by-step denoising, guided by a classifier or, in classifier-free models such as GLIDE, directly by the text. DALL-E 2 (2022) uses the CLIP text and image encoders to produce an image representation that a diffusion decoder turns into a picture. Google’s Imagen relies on a frozen T5-XXL text encoder and a cascade of diffusion models that increase resolution.

DALL-E 2 and Imagen were initially withheld from the public over misuse concerns, but were later released commercially, and openly available models such as Stable Diffusion put the technology in anyone’s hands. The same approach now extends to video and audio.

Generating tabular data and time series

For structured data, a model of the real dataset is built and then used to draw new values with similar statistical properties, so that an analyst working on the synthetic set reaches the same conclusions as on the original. The methods fall into statistical techniques and machine learning.

  • Monte Carlo methods use repeated random sampling from specified probability distributions to model a probabilistic system; by the law of large numbers, the average result converges on the expected value as the number of trials grows. The approach is simple but only as accurate as the empirical data used to define the problem.
  • Markov chain Monte Carlo (MCMC) and Gibbs sampling generate sequences of dependent observations, which makes it possible to sample from high-dimensional distributions. Gibbs sampling draws each variable in turn from its conditional distribution, which is useful when the joint distribution is unknown or hard to sample directly.
  • Bayesian networks represent the joint distribution of a set of variables as a directed acyclic graph, with nodes for variables and edges for conditional dependencies. They are computationally efficient and remain a popular way of synthesising population data.
  • table-GAN adds a third network, a classifier, to the generator and discriminator in order to preserve the semantic integrity of generated records, and was designed with data protection and leak prevention in mind.
  • CTGAN (conditional tabular GAN) deals with mixed data types, non-Gaussian and multimodal distributions and imbalanced categories through mode-specific normalisation, a conditional generator and training by sampling.
  • CTAB-GAN adds a classifier and an improved encoding of mixed and highly skewed variables; its successor CTAB-GAN+ incorporates differential privacy, so that utility can be traded against a defined privacy budget.
  • TimeGAN preserves temporal dynamics in time series such as share prices, using four jointly trained components: an embedding function, a recovery function, a sequence generator and a sequence discriminator.

More recent work applies diffusion models and language models to tables. Evaluating the quality of synthetic tabular data remains difficult, because fidelity, usefulness for machine learning and privacy have to be measured separately.

Selected tools

  • Faker and Mimesis are Python libraries that generate mock data (names, addresses, dates, payment details and similar fields) for many locales; Mimesis is the faster and more comprehensive of the two. The older pydbgen offers similar functions with output to a Pandas DataFrame, an SQLite table or an Excel file. None of them models relationships found in a real dataset.
  • Mesa is a framework for agent-based modelling, in which simulated agents (cells, animals, people or organisations) interact with each other and their environment according to defined rules; its DataCollector component records model-level and agent-level variables as tables.
  • DataSynthesizer takes a real dataset and produces a structurally and statistically similar one with privacy guarantees. Its DataDescriber module learns the distributions and correlations, the DataGenerator samples from them in random, independent-attribute or correlated-attribute mode, and the ModelInspector compares input and output.
  • Synthetic Data Vault (SDV) is a family of libraries for single-table, multi-table and sequential data, using statistical and deep learning models such as CTGAN, together with metrics for assessing the quality of the result.

Before commercial use, the licence terms of each tool and model should be checked, as some are distributed under source-available rather than fully open licences. Every data-derived method also faces the same trade-off: the more faithfully the synthetic set reproduces the original, the more useful it is, and the greater the risk that it reveals something about real individuals. That trade-off is where the legal analysis begins.

Is synthetic data personal data?

The GDPR does not apply to anonymous information (Recital 26). Whether a synthetic dataset qualifies depends on whether individuals in the source data remain identifiable by means reasonably likely to be used. In EDPS v SRB (C-413/23 P, judgment of 4 September 2025) the Court of Justice confirmed that identifiability is assessed in relation to the entity concerned: the same data may be personal for the party holding the key or the original records and not personal for a recipient with no realistic means of identification.

On 7 July 2026 the European Data Protection Board adopted draft Guidelines 02/2026 on anonymisation, the first EU-level update since the Article 29 Working Party’s Opinion 05/2014. The public consultation closed on 30 October 2026 and the final text is awaited. The draft matters for synthetic data in four respects.

  • No presumption of anonymity. AI model outputs and synthetic data are not anonymous by default. The risk of re-identification need not be zero, but it must be insignificant in reality.
  • Three criteria. Anonymity is tested against record isolation, linkage and inference. For synthetic data the inference test is the critical one: if querying the data or the generator reveals something specific about a person, or reveals that the person was in the source dataset, anonymity is in doubt. General knowledge about a population does not count.
  • Generation is processing. Training a generator on personal data is itself processing and needs a legal basis under Article 6 GDPR and, for health or other special category data, an exemption under Article 9(2), even if the output is anonymous. Data subjects should be informed that their data will be used in this way.
  • Documentation and review. The controller must be able to demonstrate anonymity, keep a record of the tests performed and reassess when technology lowers the cost of re-identification. A contractual ban on re-identification helps but is not sufficient on its own.

The same reasoning applies to the generative model. In Opinion 28/2024 the EDPB concluded that an AI model trained on personal data cannot automatically be regarded as anonymous, because personal data may be extracted from it. Companies that buy or license a synthesiser trained on someone else’s data should therefore ask how it was trained and tested.

The legislative picture is still moving. The Commission’s Digital Omnibus proposal of November 2025 would have narrowed the definition of personal data, but the Council’s negotiating mandate of 21 October 2026 keeps the current definition and instead proposes a specific rule on pseudonymised data and a clearer path for relying on legitimate interest in AI development. The European Parliament has not yet adopted its position, so none of this is law.

Synthetic data under the AI Act

Regulation (EU) 2024/1689 (the AI Act) treats synthetic data as a privacy-preserving alternative, not as an exemption. Three points are worth noting.

  • Data governance. Training, validation and testing datasets for high-risk AI systems must be relevant, sufficiently representative and, as far as possible, free of errors, and their origin must be documented (Article 10). These requirements apply equally to synthetic datasets, which can inherit or amplify the biases of their source.
  • Synthetic data first. Under the new Article 4a, introduced by Regulation (EU) 2026/1744 in place of the former Article 10(5), special categories of personal data may be processed to detect and correct bias only where this cannot be achieved with other data, including synthetic or anonymised data. Article 59 applies the same logic to further processing of personal data in AI regulatory sandboxes.
  • Revised timetable. The same amending regulation, in force since 27 July 2026, postponed the high-risk obligations to 2 December 2027 for Annex III systems and 2 August 2028 for systems covered by Annex I product legislation. Obligations for providers of general-purpose AI models have applied since 2 August 2025.

Synthetic content: marking and labelling duties

The AI Act also regulates the output side. Since 2 August 2026, under Article 50(2), providers of AI systems that generate synthetic audio, image, video or text must ensure that outputs are marked in a machine-readable format and are detectable as artificially generated. Systems already on the market before that date have until 2 December 2026 to comply. Under Article 50(4), deployers must disclose deep fakes and AI-generated text published to inform the public on matters of public interest, unless the text has undergone human review and is under editorial responsibility.

The Commission published a voluntary Code of Practice on marking and labelling of AI-generated content on 10 June 2026 and final guidelines on Article 50 on 20 July 2026. According to the guidelines, marking is not required, among other cases, for outputs that stay within an organisation in controlled business settings, a point of practical importance for synthetic datasets used purely internally for testing or model training. In addition, from 2 December 2026 the AI Act prohibits AI systems that generate non-consensual intimate imagery of identifiable persons.

Copyright and database rights

A generator trained on protected works or databases raises intellectual property questions at two stages.

Training. Reproductions made for text and data mining are permitted under Articles 3 and 4 of Directive (EU) 2019/790, the latter only where the rightholder has not reserved its rights in an appropriate manner. Poland transposed these exceptions in Articles 26² and 26³ of the Copyright and Related Rights Act with effect from 20 September 2024. Providers of general-purpose AI models must in addition maintain a copyright compliance policy and publish a summary of training content (Article 53 of the AI Act). The first case on generative AI before the Court of Justice, Like Company v Google (C-250/25), is pending. In his Opinion of 3 November 2026 Advocate General Szpunar advised the Court to declare the questions on model training inadmissible, while taking the view that real-time retrieval and reuse of press articles in chatbot answers may infringe press publishers’ rights. The judgment is still to come, and first-instance rulings in national courts, such as Getty Images v Stability AI in London and GEMA v OpenAI in Munich, point in different directions.

Output. Purely machine-generated content is generally not protected by copyright in the EU or in Poland, because protection requires a human author’s creative contribution. A synthetic dataset may therefore need to be protected by contract and as a trade secret. Conversely, output that reproduces recognisable parts of a training work may infringe, and extraction of a substantial part of a protected database may engage the sui generis right under Directive 96/9/EC.

Health data and other regulated sectors

Synthetic data is particularly attractive in life sciences, where patient-level data is scarce and sensitive. Health data used to train a generator is special category data, so an Article 9(2) GDPR ground, a data protection impact assessment and robust re-identification testing will normally be required. The European Health Data Space Regulation (EU) 2025/327, whose rules on secondary use start to apply from March 2029, will create a formal route for access to health data for research and innovation. Where synthetic data supports the development or validation of a medical device or a regulated financial model, sectoral regulators will also expect evidence that it is representative of the real population.

The Polish dimension

The Act of 3 July 2026 on Artificial Intelligence Systems, which for the most part entered into force on 11 August 2026, established the Commission for the Development and Safety of Artificial Intelligence (KRiBSI) as the market surveillance authority and single point of contact under the AI Act. KRiBSI can conduct inspections, hear complaints, impose administrative fines and issue individual opinions on the application of the AI Act, and the Act provides a basis for regulatory sandboxes. The President of the Personal Data Protection Office (UODO) remains competent for the data protection aspects, so a project involving data-derived synthetic data may fall under the supervision of both authorities.

Practical checklist

  • Identify the method: rule-based, simulated or derived from real data. Only the last normally triggers the GDPR analysis.
  • Establish the legal basis and purpose compatibility for using the source data to train the generator, and inform data subjects.
  • Test the output and the model against record isolation, linkage and inference, document the results and schedule reassessment.
  • Do not label a dataset as anonymous in contracts or privacy notices unless the assessment supports it.
  • Check the provenance and licence of the generator, the training data and any pre-trained model, including text and data mining reservations.
  • Where synthetic data feeds a high-risk AI system, document its origin, representativeness and known limitations in line with Article 10 of the AI Act.
  • Where synthetic content is released outside the organisation, apply the marking and labelling rules of Article 50.
  • Allocate responsibility in contracts: warranties on training data, re-identification prohibitions, audit rights and liability for infringing output.

Key sources

  1. Regulation (EU) 2016/679 (GDPR), in particular Articles 4(1), 6, 9 and 35 and Recital 26.
  2. Judgment of the Court of Justice of 4 September 2025, EDPS v SRB, C-413/23 P.
  3. EDPB, draft Guidelines 02/2026 on anonymisation (adopted for consultation on 7 July 2026); EDPB Opinion 28/2024 on AI models.
  4. Regulation (EU) 2024/1689 (AI Act), in particular Articles 4a, 10, 50, 53 and 59, as amended by Regulation (EU) 2026/1744.
  5. Code of Practice on marking and labelling of AI-generated content (10 June 2026) and Commission Guidelines on Article 50 of the AI Act (20 July 2026).
  6. Directive (EU) 2019/790, Articles 3 and 4; Directive 96/9/EC; Polish Act of 4 February 1994 on Copyright and Related Rights.
  7. Opinion of Advocate General Szpunar of 3 November 2026, Like Company v Google Ireland, C-250/25.
  8. Regulation (EU) 2025/327 on the European Health Data Space.
  9. Polish Act of 3 July 2026 on Artificial Intelligence Systems.

This text is for information purposes only and does not constitute legal advice.

 

Services

KIELTYKA GLADKOWSKI is a Polish business law firm headquartered in Warsaw and Krakow, with a New York office under the iSTART1 brand. Building on a practice dating back to 1981, we advise clients in the following areas.

Explore our services

Select a practice area to learn more.

Business
& Finance
Technology
& IP
Energy
& Projects
Sectors &
Cross Border
Private Clients
& Disputes
Commercial & Corporate Polish LLCs Employment LawIT, IP, Media & New Tech Trademarks & Brand ProtectionGreen Transition Energy Policy 2040Pharma, Life Sciences & Healthcare Pharmaceutical LawPrivate Client Desk HNW & Next Gen Wealth
Venture Capital & Private Equity Stock Exchange & Securities Financial Instruments & CryptoAI Sector Data CentersEnergy Sector Renewable EnergyMedical Devices Food LawInternational Divorces International Family Law
Agreements & Contracts International Trade & Smart ContractsData Privacy & Protection CybersecurityBiofuels & Biocomponents Offshore WindArt Market FashionCross Border Inheritance Immigration & Polish Passports
Offshore Companies & TaxDistributed Ledger Technology Metaverse & Web3Greenfield Investment Horizon EuropeCross Border Practice Customs & TariffsAsset Tracing in Poland Asset Tracing Business Investigations
Company Dissolution & Liquidation Cross Border BankruptcyDual Use Technology Drones & UAS Defence & SpaceInfrastructure Construction Public Private PartnershipsSworn & Certified Legal TranslationsLitigation in Polish Courts Litigation & Competition

Service directory: kg-legal.eu/services

 

UP