Concept
Synthetic Data
Artificially generated data that mimics the statistical properties of real data - used to augment scarce training sets, preserve privacy, simulate rare events, and train AI systems when real-world data collection is impossible or prohibited.
Added May 18, 2026
High-quality labelled data is the primary constraint on machine learning performance for many tasks. Annotating medical images, transcribing rare languages, recording edge-case driving scenarios, or collecting examples of dangerous industrial failures is expensive, slow, and sometimes impossible. Synthetic data is generated (by humans, procedural systems, or other AI models) to supplement or replace real data for training.
The spectrum of synthetic data approaches is broad. Rule-based simulation generates data from explicitly programmed simulators: autonomous driving training uses game engines (CARVE, CARLA) and physically accurate simulators (Waymo's carcraft) to generate millions of synthetic driving scenarios including rare events (pedestrians running into traffic, vehicles appearing from blind spots) that would take years of real driving to accumulate. Industrial quality control uses 3D rendering to generate realistic images of defective parts in various lighting and camera conditions, addressing the severe imbalance between defect-free and defective training examples.
Data augmentation is a mild form of synthetic data: applying transformations to existing examples (cropping, flipping, rotation, colour jitter for images; paraphrasing, word substitution for text; time warping, noise addition for audio) to generate new training examples that improve generalisation. Mixup and CutMix create new examples by interpolating or splicing existing ones, producing synthetic training points that regularise model behaviour.
Generative model-based synthetic data uses GANs, VAEs, or diffusion models to generate new examples drawn from the learned distribution of real data. Synthetic faces generated by StyleGAN are indistinguishable from real faces to human observers. Synthetically generated tabular data (using Tabular GANs or diffusion models) can match the statistical properties of sensitive financial or healthcare records, enabling model training and algorithm testing without exposing real records. Synthetic medical images (CT scans, histopathology slides) augment training sets for rare conditions where real cases are few.
LLM-generated synthetic data has become particularly impactful for language model training. Rather than relying entirely on human-written text (limited in volume and coverage of specialised topics), LLMs generate synthetic conversations, Q&A pairs, code examples, and domain-specific text. Microsoft's Phi series of small but capable models was trained primarily on high-quality LLM-generated synthetic data ("textbook quality" educational text), demonstrating that carefully curated synthetic data can partially substitute for large diverse web corpora.
The fundamental risk in synthetic data: model collapse. If models are trained on data generated by other models, which is then used to train more models, statistical biases and errors from the generator amplify in each generation. Real-world data remains necessary to anchor synthetic data to actual distributions. Research on model collapse (Shumailov et al., 2024) formalised this risk: iteratively training on synthetic data degrades performance across generations.
Analogy
Flight simulators producing realistic training scenarios for pilots without requiring expensive real aircraft hours. A pilot can practice hundreds of emergency procedures, instrument failures, and rare weather events in a simulator that cannot be safely replicated in real aircraft. The simulator's synthetic experiences are not identical to real flight (some fidelity is always lost), but they provide sufficient training signal to dramatically accelerate skill development. Synthetic data does the same for AI: artificial examples that approximate the properties of real data well enough to accelerate training without requiring expensive real-world data collection.
Real-world example
Waymo generates over 15 million miles of synthetic driving data per day using their proprietary simulation environment (Waymo Simulation City). Real-world fleet data identifies challenging scenarios (complex intersections, unusual vehicle behaviour, pedestrian edge cases). These are then simulated with systematic variations: different times of day, weather conditions, pedestrian speeds, and vehicle trajectories. The resulting synthetic data covers a distribution of scenarios far broader and more safety-critical than could be collected from real driving alone.
Why it matters
Synthetic data is increasingly central to AI development, addressing the data scarcity that limits performance in specialised domains, the privacy constraints that prevent use of real data, and the long-tail data distribution problems that make real-world collection insufficient for rare scenarios. Understanding synthetic data - its mechanisms, its quality assessment challenges, and the model collapse risk from excessive use - is essential for building AI systems where real data is scarce, sensitive, or dangerous to collect.
In the news
Related concepts