On this page
A privacy control has greater value when it determines what information a system needs to receive, rather than trying to conceal information after it has already travelled through the system.
Start with the flow of information
Many useful datasets contain information about people. Customer records, transactions and contact details may contain names, addresses or identifiers alongside analytical measures. Some of those fields are needed to group or compare records; that does not mean the identity of a person needs to be exposed to every component performing that work.
This distinction shaped the privacy architecture in determinIQ. Rather than treating personal information as an afterthought, the design asks what can be identified and protected before ordinary analytical ingestion, and what representation the analytical engine can work with afterwards.
A system can be mathematically deterministic while still handling personal information badly. Correct analysis does not excuse unnecessary disclosure.
Protection does not have to discard analytical meaning
Removing every identifying column entirely may prevent useful analysis. For example, a business question may require grouping records by a consistent customer identifier. What is often needed is not the person’s name, but a stable way to recognise that separate rows relate to the same analytical subject.
Supported determinIQ protection flows can substitute tokens or pseudonymous representations for identity-bearing values. The governed analytical process can then use the permitted representation rather than unrestricted original values. Where an authorised user needs to recognise a name, a separate identity-recovery path can support presentation without recalculating the analytical result.
There is a fundamental distinction here: pseudonymised is not the same as anonymous. If the system retains an authorised means of restoring an original value, it would be wrong to suggest that the person can no longer be identified. The purpose is controlled separation and minimisation, not a change of terminology that makes the risk disappear.
AI needs a defined data boundary
It is tempting to send a complete dataset to a model because doing so is convenient. But the language tasks needed by an analytical product are narrower. Interpreting a possible question may require field names and supported analytical capabilities. Explaining a governed result may require a prepared result and appropriately controlled context. Neither task necessarily requires unrestricted original rows containing personal information.
The determinIQ design therefore separates the AI intent and narrative boundaries from the raw dataset. Its documented intent path uses governed schema metadata instead of raw dataset rows and example values; its narrative path prepares context to exclude raw row data and personally identifiable information. These are specific implemented design boundaries, not a claim that no information of any kind is ever passed to an AI service.
The distinction matters because blanket statements are easy to write and difficult to prove. A useful privacy description should say which information is protected, at which boundary, for which operation, and how that behaviour has been tested.
Revealing information is itself part of the design
Identity protection is not finished once values have been tokenised. The application may need to reveal original labels locally for an authorised user. That introduces another responsibility: a name shown in a table, a chart axis, a tooltip or an exported image is still personal information.
Recent determinIQ development has specifically addressed the behaviour of protected chart labels during local name reveal. The documented design keeps the underlying numerical series and governed evidence unchanged while refreshing labels only when a user has already revealed the corresponding identity in the controlled interface. It also provides a route back to protected presentation.
The November implementation note describes focused automatic checks, but identifies further browser and data-flow acceptance work. I would therefore present this as an evolving capability, not a finished independent privacy certification.
What has been established, and what remains
The architecture demonstrates a practical separation between analytical use of data, use of AI for language tasks and authorised identity presentation. It also shows that retaining analytical value does not always require retaining unrestricted visible identity throughout the user journey.
What remains necessary is ongoing verification across every supported input and output route. A system’s privacy behaviour can be undermined at a visualisation boundary, in a log, through an unexpected export or through a change elsewhere in the data lifecycle. Effective controls are a property of the whole flow, not simply the storage format of one mapping.
This work connects directly to the wider research programme: useful analysis should be accessible without giving any component more authority—or more personal information—than its task genuinely requires.
Research context
Research programme: Making data analysis understandable, reproducible and accountable
Topic: Governed analytics
Explore further
Supporting material is listed below.
Further reading is listed below.