We thought the danger was confined to cookies or our browsing history. However, a new threat to our digital privacy has just been highlighted by a team from the University of New South Wales (UNSW). Their study proves that a language model (LLM) can draw up a disturbingly precise identikit of an Internet user based on a hitherto underestimated element: the advertisements displayed on their screen.
How can an AI create a profile from simple advertisements?
The experiment carried out is revealing. Scientists collected more than 435,000 ad impressions from 891 Australian volunteers. Each advertisement, text and image, was then subjected to a artificial intelligence like Google’s Gemini. The AI first analyzed each ad individually before reconstruct a complete profile by crossing all the advertisements seen by the same person.
The results go far beyond mere chance. The model guessed gender with 59% accuracy, employment status at 48%, education level at 43% and even political preference at 35%. These scores match, or even surpass, those of human annotators, but at a cost 223 times lower and 52 times faster. This is proof that our advertising streams are real digital footprints.
Why are current protections like VPNs ineffective?
The most worrying detail of this study is that usual protections are useless. A VPN, for example, does not protect you at all. The reason is simple: advertisements are targeted upstream by agencies like Meta or Google, based on the profile that the platform has already meticulously built about you.
Changing your IP address does not change this underlying profile. Worse still, the attack described by the researchers does not even need to access the platform’s servers. Any browser extensioneven the most innocuous one, with permission to “read page content” could capture this advertising flow and send it to an external LLM for analysis. The study also revealed advertising bias significant, with men seeing 2.25 times more gambling advertisements.
What are the implications for our security and regulation?
This new profiling capability opens a gaping hole for personalized and sophisticated scams. A cybercriminal could exploit this technique to infer a target’s financial status or vulnerability without ever having to hack their device. The question of the private life is therefore more central than ever.
On a legal level, this scenario directly calls into question European law. The Court of Justice of the EU has already ruled that data “ likely to indirectly reveal » sensitive information must be treated as such within the meaning of the GDPR. A profile reconstructed by AI would therefore potentially fall under this very strict regime. The researchers are clear: a prolonged follow-up is not necessarya few minutes of navigation are enough for a successful attack.
