TLDR: Synthetic data, artificially generated information, is increasingly used in AI for privacy, training, and generalization. This paper explores how synthetic data complicates the traditional concept of “ground truth” by not referring to real-world observations. It discusses how synthetic data can paradoxically improve AI model performance, even with “implausible” inputs, and outlines a typology of synthetic data. The paper also examines how the role of data shifts from representation to imitation, and from evidence to a training signal, leading to new ways of defining and understanding “ground truth” in the context of AI.
In the rapidly evolving landscape of artificial intelligence, a new kind of data is challenging our fundamental understanding of truth: synthetic data. This innovative form of data, artificially generated for various purposes like privacy protection, training AI models, or simply providing convenient access to quasi-realistic information, is prompting a re-evaluation of what we traditionally consider ‘ground truth’.
Dietmar Offenhuber’s paper, Synthetic Data and the Shifting Ground of Truth, delves into this complex subject, exploring how synthetic data mimics real-world observations without directly referring to external features. This lack of a direct representational relationship, surprisingly, doesn’t hinder its utility. In fact, researchers are finding that synthetic data can often lead to better model performance than realistic data. It can help compensate for known biases, prevent overfitting, support generalization, and make models more robust against unexpected outliers. This counter-intuitive finding challenges the long-held belief of “garbage in – garbage out,” suggesting that even noisy or implausible data can be beneficial for AI training.
Understanding Synthetic Data
The paper categorizes synthetic data into four main types:
- Algorithmic: Idealized datasets and simulations used as benchmarks in fields like physics or network science.
- Obfuscated: Real observations intentionally obscured for privacy protection, often using noise injection for differential privacy.
- Mimicry, Hybridized: Data generated based on one or more existing datasets, combining them to increase resolution, enlarge datasets, or reduce known biases while preserving statistical properties.
- Generated Training Data: Entirely artificial datasets created solely for training purposes, not based on real data, with the only objective being to improve learning when injected into other observations.
The discussion primarily focuses on the latter two categories, where the connection to real-world observations is most tenuous.
Why Synthetic Data?
The motivations behind using synthetic data are diverse. A common thread is addressing the inherent biases and limitations of observational training data. Real-world datasets often contain stereotypes, lack sufficient outliers, or are scarce for underrepresented cases. Synthetic data can fill these gaps by:
- Maximizing diversity, even if it doesn’t accurately reflect demographic variations, to ensure models perform well across all populations. For example, Microsoft’s Face Synthetics dataset uses rendered 3D characters to train facial recognition models without real people.
- Generating data for rare conditions, such as rare brain tumors, where real medical data is scarce, allowing models to be trained to recognize them effectively.
- Modeling “black swan events” or phenomena that have never been observed, helping manage risks in complex systems.
Also Read:
- Crafting Digital Lives: How AI Generates Realistic Smartphone Usage Data
- AI Models Learn to Adapt by Deceiving Distribution Shift Detectors
Shifting Concepts of Data and Truth
The rise of synthetic data leads to profound “figure-ground inversions” in how we perceive data and truth:
From Representation to Imitation: Traditionally, data is seen as a representation of reality. However, synthetic data doesn’t represent; it imitates. It behaves like real data, substituting and augmenting in specific application contexts. Its quality is no longer measured by how well it mirrors reality, but by its effectiveness for a particular purpose, such as preserving privacy or improving model predictions.
From Evidence to Training Signal: The conventional view requires data to demonstrate a clear chain of custody back to its origin to establish accuracy and reliability. With synthetic data, this arrow is reversed. Datasets are constructed based on a desired outcome. If a model is biased, synthetic data is created to correct that bias. The focus shifts from data’s correspondence to reality to the performance of the model it trains.
The Evolving Concept of Ground Truth: The paper highlights three ways “ground truth” is being redefined:
- Best Available Measure of External Validity: This traditional view sees ground truth as the most reliable information, ranging from positivist surface observations (like in remote sensing) to ethnographic observations that contextualize big datasets.
- Whatever Labelers Label: Here, external truth is less central. Ground truth becomes an operational definition, often simply what human annotators (or even other AI models) label. The reliability of these labels can be uncertain, especially with outsourced labeling or machine-generated labels.
- Disagreement as Productive: This perspective rejects a single “gold standard” and instead treats disagreements among annotators as valuable signals. It acknowledges that human uncertainty can capture nuances and ambiguities, making models more robust and better at identifying when not to make a prediction.
In conclusion, synthetic data marks a significant departure from the evidentiary paradigm. It challenges the long-standing association of data with representation, accuracy, and evidential grounding. Instead of passive artifacts, synthetic data acts as an active agent, shaping model learning without necessarily describing external states of affairs. Its value lies not in fidelity to the real world, but as an instrument of variation and transformation, prompting us to rethink data as interventions that generate and reconstitute the very foundations of knowledge claims.
