The uncomfortable moment in an AI privacy review is not when the team admits it collected too much data. It is when someone notices the model has created a new field nobody put in the intake form: likely income, possible health condition, inferred race, or another sensitive characteristic. Old data maps have a neat place for collection. They are less good at asking what the system manufactures after collection, then stores or uses. ## The new personal data may be the output Bloomberg Law framed the issue directly in its July 6, 2026 article, AI-Powered Inferred Data Poses New Threats for Consumer Privacy. Dykema's summary of that Bloomberg Law piece says the article examines AI tools that generate inferences about consumers, including health, race, income, and other sensitive characteristics, then use those inferences to influence decisions involving employment, housing, lending, and other areas. That is the useful counterweight to the usual minimization script: privacy risk can be created after the user has finished typing. For builders, the practical lesson is not that every prediction is forbidden. It is that an inference used in a consequential workflow should be treated as a governed data asset, not as harmless analytics exhaust. If a model assigns a consumer to a risk bucket, income band, or likely condition, the product team needs to know whether that value is stored, refreshed, shared, or used to change an outcome. The privacy register that stops at raw inputs is now politely lying. ## Why collection notices do not solve inference TechPolicy.Press makes the architecture problem explicit. Ikenna Ogbogu writes that foundation models learn latent statistical representations from large quantities of human generated information, turning data into inferential capabilities. The same TechPolicy.Press essay says those capabilities allow AI systems to generate sensitive inferences about individuals from information that was never explicitly disclosed. That matters because notice and consent are built around the idea that a person can understand what they gave you. Inference breaks that comfort. A consumer may disclose purchase behavior, social media activity, or conversational patterns, while the system produces a sensitive conclusion the consumer never wrote down. The compliance question moves from, did we collect this field, to, did our system create it, keep it, and use it in a way we can explain. ## Privacy by design now has to inspect the pipeline Slaughter and May's note on the Information Commissioner's Office guidance quotes Elizabeth Denham as saying, "the underlying data protection questions for even the most complex AI project are much the same as with any new project. Is data being used fairly, lawfully and transparently? Do people understand how their data is being used and is it being kept secure?" The same note says data minimization can be especially challenging where machine learning decides what information is necessary from large datasets. That is the dry sentence privacy teams should tape above the model review board. Plainly translated, the inference pipeline needs its own controls. A product review should identify inferred attributes, the inputs used to produce them, the model or rule that generated them, the retention period, and the downstream decision they affect. Vendor contracts should ask the same questions: what inferences are produced, whether the vendor may reuse them, how deletion or access requests are supported, and whether the customer receives enough explanation to describe the processing honestly. Calling the output a score does not make it less personal if it functions like a sensitive profile. ## The small business excuse is getting thinner Dykema's summary of the Bloomberg Law article adds a second uncomfortable point. Matthew Hays, a Dykema data privacy attorney, explains that inexpensive, widely available AI tools have made sophisticated consumer profiling accessible even to small businesses that may lack resources or awareness. In other words, profiling capacity has moved faster than governance capacity. That is a familiar compliance smell. Dykema also says lawmakers are increasingly focused on transparency, limiting discriminatory outcomes, and giving consumers greater control over how inferred data is used. Those are not identical obligations in every jurisdiction, and builders should resist the online habit of flattening them into one universal AI rule. The safer design move is more concrete: make inferred attributes visible inside the company before regulators, consumers, or counterparties ask where they came from. The next privacy review should therefore start one step later than usual. Do not only ask what the app collects at signup, checkout, onboarding, or chat. Ask what the model produces after that, whether anyone relies on it, and whether the user would be surprised to learn it exists. That is where AI privacy design is becoming less about prettier notices and more about product architecture that can survive being read aloud in an enforcement file. ## Sources - AI-Powered Inferred Data Poses New Threats for Consumer Privacy

Sources