For decades, protecting privacy has meant one thing: stop personal data from leaking out. Gartner thinks that idea is about to break.
By 2029, most privacy incidents will come not from leaked personal data but from what AI infers about people, the research firm predicted this week. The risk is no longer only the records a company holds. It is the conclusions a model can draw from scraps that look harmless.
From data exposure to insight exposure
Gartner analyst Bart Willemsen calls it a shift “from data exposure to insight exposure.” A model can now reconstruct deeply personal things, like a health condition or a behavioural pattern. It pulls them from data that looks anonymous, aggregated or harmless, and never has to breach a database to do it.
There is an irony in this. Companies are storing less personal data, pushed by regulation and cost. That is meant to reduce risk. But if an AI can infer the sensitive detail anyway, holding less data does little to protect the person it describes.
Attacks that leave no leaked record
These inference attacks are hard to catch. A normal breach leaves a trail: a stolen file, an exposed record, an alert that fires. An inferred conclusion leaves none of that.
“Inference attacks are particularly dangerous because they often evade conventional detection mechanisms,” Willemsen said. “Individuals can be exposed through AI-generated conclusions rather than leaked records, creating privacy risks that undermine data integrity and are difficult to detect, explain and mitigate.”
The threat also sits outside most privacy law, which largely governs personal data that companies collect, store and share. A guess a model makes is none of those things. It is a growing blind spot as everyday tools quietly record and analyse more of our lives.
What Gartner wants companies to change
Gartner’s advice to security chiefs is to govern what AI concludes, not just what it stores. It expects spending on data “integrity” protections to reach parity with data confidentiality by 2028, as firms respond to inaccurate, biased or unauthorised AI-generated profiles.
Its recommended fixes are practical. Bake privacy checks into how teams build AI systems. Adopt privacy-enhancing tools such as differential privacy and synthetic data. Minimise the data on hand. And keep a human in the loop to sign off before an AI acts on a sensitive inference.
That last point matters because the exposure can be quiet and accidental, not just malicious. TNW has covered how private AI chats ended up indexed by search engines, no breach required.
A prediction, not a certainty
The usual caveats apply. This is a Gartner forecast, and Gartner sells the research and the conference seats behind it. “Most privacy incidents” by a given year is the kind of claim that is easy to state and hard to prove.
But the underlying worry is real. Researchers have shown for years that models can re-identify people and guess private traits from public data. Regulators are only starting to catch up, with the EU’s AI Act still bedding in. Gartner’s point is simply that the industry has spent years guarding the wrong door.
