Preserving Context: The Case for In-Stream Data Annotation

The pursuit of perfectly clean data may be counterproductive, especially in the age of generative AI. As models become more sophisticated, their reliance on stale or incomplete information can lead to confident but incorrect answers.

Instead of solely focusing on cleansing, experts advocate for capturing metadata and context at the source—what’s being called “in-stream labeling.” This approach recognizes that true data quality isn’t about absolute purity but rather about preserving provenance and relevance.

“You can’t pursue perfectly clean data; it’s just not possible,” says David Aronchick, founder of Kubeflow and CEO of Expanso. “Every line should have a mechanism saying where it came from. Otherwise, you’re never really going to know because you can’t reconstruct the context later.”

This philosophy aligns with observations that nearly 75% of CPU work in training frontier models goes toward data cleansing and validation—a process that often strips away valuable contextual information.

Why Context Matters

Consider a wind farm operator: sensor readings from individual turbines are meaningful only when viewed within the context of turbine location, weather conditions, operating speed, etc. Without this metadata, performance anomalies might be misdiagnosed as failures instead of routine variations.

Similarly, IoT digital twin systems tag data back to specific devices so that temperature spikes can be distinguished between critical malfunctions and calibration events—information that could prove valuable for broader analysis down the line.

Ulrik Hansen, co-CEO of Encord, emphasizes this point: “Dirty data often conflates actual corruption with contextual anomalies. Cleansing both types of issues, you lose information that might be useful later.”

The Future of Data Management

The shift toward in-stream annotation reflects a broader trend toward smarter data management—one where metadata is captured early and continuously enriched rather than treated as an afterthought.

This approach not only supports more accurate AI models but also aligns with emerging regulatory requirements like the EU AI Act, which mandates traceability of data used in AI systems.