2. Why Use Synthetic Data?

Organizations face persistent data bottlenecks: (i) privacy, regulatory and copyright limits; (ii) scarcity; (iii) bias, imbalance and unrepresentative samples; (iv) noise and missingness; (v) high collection cost; and (vi) cases where data simply does not exist. Major technology firms have publicly explored synthetic data approaches to meet growing data needs. We classify synthetic data’s solutions to these barriers into five capabilities.

Figure: Capabilities of synthetic data addressing data challenges

1. Privacy-Preserving Data Sharing

Synthetic data enhances privacy by replacing sensitive records with artificial yet statistically representative data. Instead of replicating individuals one-to-one, it aims to capture the general patterns, variability, and uncertainty of the source data. Synthetic data should therefore be viewed as a privacy-preserving approach when combined with safeguards (e.g., regularization, outlier handling, or differential privacy) within broader privacy-enhancing technology frameworks.

Example Applications: Cross-organizational data sharing for research and collaboration, and third-party sensitive data access for testing and development.

Synthetic is not safe by default

Synthetic data (and the AI models for generating synthetic data called synthesizer) can still be vulnerable to privacy risks, necessitating rigorous evaluation and mitigation strategies that we have covered in the Privacy-preserving Data Synthesis chapter.

2. Overcoming Data Scarcity and Accessiblity

Real-world data is often scarce, costly, or inaccessible due to privacy restrictions, proprietary ownership, collection difficulties, or limited sample sizes. Once a synthesizer is trained on available real data, synthetic generation can create substantially more samples. By expanding coverage of rare patterns, balancing underrepresented groups, and reducing the risk of overfitting to small datasets, synthetic data can improve machine learning model performance and enable analytics and research that would otherwise be impossible with limited real datasets.

Example Applications: Medical research with additional samples of rare diseases, and financial modeling with more examples of unusual market conditions or fraud patterns.

Diminishing returns

Creating more samples eventually stops helping for your downstream tasks. If results stop improving, focus on better coverage or quality of synthetic data, or add fresh real data—not just more volume.

Data scarcity is especially acute for training and evaluating Large Language Models (LLMs)¹: performance tends to improve as models and datasets grow together (with scaling laws). With limited human-generated data and concerns about contamination, synthetic data now plays a key role in sustaining LLM progress.

3. Improving Data Representativeness

Many real-world datasets underrepresent certain groups of people or situations based on demographics, race, conditions, or edge cases, leading to biased models that fail to generalize. Synthetic data can be generated with fairness constraints to improve data balance and train more equitable systems. This contributes to both representation fairness (all groups are included in the training data) and algorithmic fairness (model outputs are not skewed against any group).

Example Applications: Augmenting healthcare datasets to ensure rare conditions are represented in predictive models; generating balanced tabular data for loan approval systems across demographic groups; or creating diverse synthetic text corpora to reduce bias in language models.

Distribution shift risks

Generating synthetic data to “balance” representations can introduce distribution shift—a mismatch between the synthetic dataset and the true underlying real-world distribution. This can lead to datasets that no longer reflect real-world conditions. For e.g., artificially balancing gender representation in hiring data might mask actual workplace inequities that need addressing.

It is critical to perform careful downstream task evaluation to ensure these trade-offs serve your intended purpose. Practical approaches include reporting performance metrics for both overall populations and specific subgroups, monitoring fairness–utility trade-offs, and validating that synthetic adjustments align with your ethical and business objectives.

4. Reducing Expensive Data Collection and Labeling

Data collection is often a significant bottleneck, with manual efforts like data labeling taking months and costing millions. Legal and privacy constraints may also prevent the reuse or commercialization of datasets. Synthetic data offers a faster, cheaper alternative by allowing the generation of tailored, task-specific datasets on demand. Additionally, synthetically generated annotations (e.g., preference ratings or intent labels) can offer more consistent and scalable alternatives to human-labeled data.

Example Applications: Bootstrapping a new customer service chatbot with thousands of labeled synthetic conversations categorized by intent (billing, technical support, complaints) before real interactions are collected; generating pre-labeled medical images to help prototype diagnostic models before large-scale expert annotations are available.

Validation still required

Synthetic labels can introduce systematic errors or biases that human annotators might catch. Validation on real, human-labeled holdout data remains essential to verify label accuracy across different subgroups and use cases. For instance, synthetic customer service labels might consistently misclassify complex billing disputes as simple account inquiries. Regular validation helps identify these patterns and maintain quality standards across your specific domain and user populations.

5. Simulating “What-if” Scenarios

Real-world data rarely covers edge cases, future events, or hazardous conditions needed for resilience testing. A synthesizer can create targeted “what-if” datasets by adding conditioning logic and recombining learned patterns under domain constraints. For e.g., it can pair normal traffic flows with extreme weather, or blend typical patient symptoms with a rare complication. This can provide scenario datasets that support autonomous vehicle stress tests and healthcare surge planning, subject to plausibility checks and validation.

Example Applications: Government agencies testing emergency response systems with simulated crisis scenarios (natural disasters, cyberattacks), and AI safety teams generating adversarial prompts to test model robustness against misuse attempts such as jailbreak prompts, hallucination, and bias probes.

Avoid synthetic fantasies

Synthesizers can create scenarios that look plausible but violate real-world constraints. For e.g., a healthcare synthesizer might generate a patient record showing someone recovering from major surgery in just two days, or a traffic simulator might create rush-hour patterns with impossible vehicle densities. These “synthetic fantasies” appear realistic in isolation but fail when deployed in practice. Subject matter experts should validate scenarios against domain knowledge, physical constraints, and operational realities before using synthetic data for critical decision-making or system testing.

---

The next chapter covers different synthetic data generation methods, providing the technical foundation for implementing these capabilities.

Footnotes

1. The authors of the paper “Will we run out of data? Limits of LLM scaling based on human-generated data” project human-generated text data shortage for LLM training by 2026-2032 based on current scaling trends and suggest synthetic data generation as of the solutions to address the data bottleneck. ↩