The Mosaic Effect Still Matters in the Age of AI

For years, organisations took comfort in anonymisation. Remove direct identifiers, aggregate the data, and the risk was assumed to be low.

In AI environments, that assumption no longer holds.

The so-called “mosaic effect” refers to the ability to combine multiple datasets, each benign in isolation, to re-identify individuals when analysed together.

As demonstrated by research from Carnegie Mellon University, 87 per cent of the US population can be uniquely identified using just ZIP code, gender and date of birth. The original work by Latanya Sweeney is available via Harvard’s Data Privacy Lab.

This finding predates modern AI. Today’s models significantly amplify that risk.

Traditional privacy assessments are dataset-level

They typically ask:

• Is this dataset anonymised?
• Have direct identifiers been removed?
• Is there a lawful basis for processing?

AI systems do not operate at dataset level. They ingest, combine, transform and infer across multiple sources.

Three shifts matter:

1. Cross-referencing at scale
   Large models can correlate patterns across structured and unstructured data with far greater precision than traditional analytics.
2. Inference as a capability
   AI does not need explicit identifiers to infer sensitive attributes. Correlated signals are often enough.
3. Context expansion
   Data collected for one purpose can gain new identifiability when repurposed inside a model trained on additional datasets.

The practical implication is straightforward: a dataset that appears low risk in isolation may become high risk when introduced into a broader model ecosystem.

Dataset-level privacy assessments are therefore insufficient in AI environments. Risk must be assessed at the system and aggregation level.

Assess aggregation risk before model training

Before introducing a new dataset into model development:

• Map all datasets that will coexist within the same model or pipeline
• Identify quasi-identifiers, not just direct identifiers
• Evaluate how attributes could interact across sources
• Conduct a re-identification risk scenario analysis at the system level
• Document residual aggregation risk within your AI risk register

The critical question is no longer: “Is this dataset anonymised?”

It is: “What happens when this dataset is combined with everything else, we hold?”

In the age of AI, privacy risk rarely sits inside a single dataset. It emerges from the mosaic.