Synthetic data must reflect justice, not just accuracy, in AI systems

A new study challenges one of the fastest-growing assumptions in artificial intelligence research, that synthetic data can serve as a harmless, neutral substitute for real data.

The research, published in Big Data & Society and titled "Synthetic Data as Meaningful Data: On Responsibility in Data Ecosystems," argues that synthetic data is far from a technical convenience. Instead, it represents a deeply ethical and relational artifact that must be governed with moral responsibility, transparency, and social awareness.

Beyond imitation: Redefining synthetic data as a moral entity

Synthetic data, artificially generated datasets used to train AI models when real data are limited, private, or biased, has become central to modern machine learning. Companies and institutions increasingly rely on it to reduce privacy risks, accelerate model training, and overcome data scarcity. However, the author warns that this approach often oversimplifies what data truly represent.

The study rejects the widespread idea that synthetic data merely replicates real data patterns. Instead, it introduces an analogical perspective, a framework that views synthetic data as a distinct yet relational form of meaning-making. Rather than copying reality, synthetic datasets produce new analogies of the world, influenced by the values, assumptions, and intentions of their creators.

The author argues that this analogical relationship carries ethical significance. Every synthetic dataset encodes decisions about what counts as relevant, fair, or desirable information. These decisions, often embedded in algorithmic design, sampling strategies, and validation metrics, determine how synthetic data mediate between human experience and digital representation. As a result, synthetic data must be treated not simply as technical artifacts but as moral and political constructs that shape social reality.

This shift from imitation to analogy redefines the boundaries of responsibility in AI. The author calls for developers and regulators to adopt a context-sensitive evaluation of synthetic data that accounts for its societal implications, not merely its accuracy or fidelity.

The responsibility gap in AI and data ecosystems

The study examines responsibility gaps, situations where no clear actor can be held accountable for the effects of AI systems trained on synthetic data. As generative models increasingly create datasets autonomously, tracing the origins and biases of these datasets becomes nearly impossible.

The paper critiques existing approaches such as data cards, model cards, and watermarking, arguing that these tools, while useful, fail to address the deeper issue of moral responsibility. They document provenance but ignore ethical causality, the social and political consequences of data design decisions.

According to the author, responsibility in AI should not be reduced to traceability or compliance checklists. True accountability, the author argues, requires understanding how data practices affect power, representation, and justice. This is particularly urgent in synthetic data ecosystems, where the source of information is not directly observable but algorithmically constructed.

To address these challenges, the author introduces two key concepts: meaningful human control and algorithmic reparation. The former emphasizes maintaining human oversight in data generation and use, ensuring that synthetic datasets reflect collective human values rather than corporate or technical expediency. The latter calls for active correction of systemic biases through participatory governance, where stakeholders, especially marginalized communities, are involved in defining what counts as fair data.

The author's approach links synthetic data governance to broader debates about data justice and responsible innovation. the author situates synthetic data within a "responsibility ecosystem" that includes not just AI developers but also policymakers, civil society, and end-users. Responsibility, in this framework, is distributed, iterative, and dialogic, a shared moral task rather than a one-time technical adjustment.

From technical substitution to ethical innovation

Synthetic data is both a solution and a symptom of the current data-driven economy. On one hand, it provides a valuable response to data scarcity, privacy restrictions, and ethical concerns surrounding real-world datasets. On the other, it exposes how easily the AI industry's pursuit of efficiency can obscure deeper moral questions about what data mean.

The author identifies a growing paradox: the same systems that generate synthetic data to correct bias can reproduce and amplify bias when used without ethical scrutiny. This phenomenon, often called model collapse, occurs when AI models trained on synthetic data repeatedly regenerate and retrain on their own outputs, leading to distortions and overfitted knowledge.

The paper urges the AI community to move beyond statistical metrics and consider qualitative forms of responsibility. Metrics like fidelity and utility measure how closely synthetic data resemble real data, but they overlook the ethical accuracy, the degree to which synthetic data reflect human values and social diversity.

The author calls for an expanded understanding of meaningful data: data that are not only valid and useful but also just, transparent, and co-created. the author advocates for governance models grounded in responsible innovation, where ethical deliberation is built into every stage of data production, from algorithm design to dataset release.