LLMs Infer Private Attributes From Ads
A new study shows researchers used a multimodal LLM pipeline on about 435,000 Facebook ad impressions from 891 users to infer private attributes from passive ad exposure. Gemini 2.0 Flash processed image and text, enabling predictions of age, gender, employment, education and party preference with accuracies like gender ~59%, employment ~48% and party ~35%. The study finds that benign-looking browser extensions can harvest visible ads to perform scalable off-platform profiling, posing privacy and regulatory challenges.
Key Points
- 1Demonstrate LLM-based profiling using 435,000 Facebook ad impressions and Gemini 2.0 Flash
- 2Reveal demographic patterns in ad delivery that leak age, gender, employment, education, and party
- 3Warn that browser extensions can harvest visible ads, enabling scalable off-platform profiling
Scoring Rationale
Demonstrates practical, scalable profiling risks with multimodal LLMs, but relies on one dataset and limited peer review.
Sources
Public references used for this report.
Practice with real Ad Tech data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Ad Tech problems
