© Vitaly Gariev (Unsplash)
Australian researchers fed an AI with Facebook ads from 891 users. Result: it guessed their gender, occupation, and political opinions without ever accessing their browsing history.
We thought the problem of online privacy was cookies, trackers, and browsing history. A team from the University of New South Wales (UNSW Sydney) has just shown that the danger lies elsewhere. In a study, the researchers proved that an AI model could reconstruct an internet user's demographic profile based solely on the ads displayed on their screen. No need to dig through personal data, cookies, or history. Ads are enough.
How an AI rebuilds your profile using your Facebook ads
The team collected 435,314 Facebook ad impressions from 891 Australian volunteers via a browser extension. Each ad (text and image) was submitted to Gemini, Google's model. The AI first analyzed each ad individually, then reconstructed a composite user profile by cross-referencing all the ads viewed during a session. Result: 59% accuracy on gender, 48% on employment status, 43% on education level, and 35% on political preference. Scores that consistently exceed chance and match (or even surpass) those of human annotators, for 223 times cheaper and 52 times faster.
Ads are targeted even before they appear on your screen
The most alarming detail of the study is captured in one sentence: a VPN does not protect. The reason is simple. Ads are targeted by ad agencies (Meta, Google) in advance, based on the profile that the platform has already built. Changing your IP address doesn't change this profile. And the attack described by the researchers doesn't even need to access the platform. Any browser extension with the permission « read the page content » could capture the ad stream and transmit it to an external LLM. The study also reveals marked advertising biases: men see 2.25 times more ads for gambling games than women, and the unemployed 2.89 times more.
In European law, the issue is far from theoretical. The EU Court of Justice ruled in 2022 that data « susceptible of indirectly revealing » sensitive information (political orientation, health status) must be treated as such under the GDPR. A profile reconstructed by AI from ads would therefore potentially fall under the strictest regime of the regulation. So far, no data protection authority has yet taken a position on this specific scenario. However, researchers from UNSW are categorical: « Prolonged tracking is not a prerequisite for a successful attack. » A few minutes of browsing are sufficient.
👉🏻 Follow tech news in real-time: add 01net to your Google sources, and subscribe to our WhatsApp channel.
Source: UNSW
- Google Play Store
- 4K Video Downloader Plus
- VLC Media Player
- Adobe Reader
- uTorrent
- WinRAR
- Xender
- Google Chrome
- CCleaner
- Mozilla Firefox
- Avast Gratuit
- ChatGPT
- Microsoft 365
- Opera One
- Avast One
