What is it about?

This study examines whether the ads people see online can reveal private information about them. Even when advertising platforms remove explicit options for targeting sensitive groups, the ads delivered to a person may still reflect the platform's inferred understanding of that person. We test whether modern large language models can use these patterns in ad exposure to infer private user attributes from ads alone. Using more than 435,000 Facebook ad impressions collected from 891 users over time, we built a pipeline that analyzes the text and images in ads and then predicts attributes such as gender, age, education, employment status, income, and political preference. We show that off-the-shelf multimodal LLMs can infer these attributes with notable accuracy, even from short browsing sessions rather than only from long-term tracking.

Featured Image

Why is it important?

This work highlights a privacy risk that is largely invisible to users and not well addressed by current platform safeguards. Most privacy discussions focus on what people actively share online, such as posts or photos. Our findings show that passive exposure also matters: simply seeing a sequence of ads can create a rich digital footprint that can be used to profile someone. This is important because the attack does not require privileged access to the advertising platform, advanced AI expertise or a large custom-trained model. With today's widely available AI tools, even a lightweight system such as a browser extension could potentially collect visible ads and use them to infer sensitive information off-platform. Our results suggest that current privacy protections need to go beyond restricting explicit targeting options and start addressing the hidden signals embedded in ad delivery itself.

Perspectives

This paper was motivated by a simple but unsettling question: what do the ads we passively see say about us, especially in a time when generative AI has made powerful inference tools much more accessible? In this work, we show that ad exposure itself can act as a meaningful source of information about private attributes when processed by multimodal LLMs. For me, the most striking part of this work is how the generative AI era changes the nature of the risk: what once required substantial data and technical effort can now be approached with off-the-shelf models. I hope this paper encourages broader discussion of privacy risks in online advertising, especially those that arise not only from the data people actively share, but also from the content algorithmic systems choose to show them.

Baiyu Chen
University of New South Wales

Read the Original

This page is a summary of: When Ads Become Profiles: Uncovering the Invisible Risk of Web Advertising at Scale with LLMs, April 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3774904.3793060.
You can read the full text:

Read

Resources

Contributors

The following have contributed to this page