[Paper Review] Automated Experiments on Ad Privacy Settings: A Tale of Opacity, Choice, and Discrimination

Amit Datta, Michael Carl Tschantz+1 more|arXiv (Cornell University)|Aug 27, 2014

Privacy, Security, and Data Protection14 references77 citations

TL;DR

This paper presents AdFisher, an automated tool that conducts randomized, controlled experiments to study transparency, choice, and discrimination in Google's ad privacy settings. Using machine learning and statistical significance testing, it reveals that Google's Ad Settings are opaque about profile changes from sensitive browsing (e.g., substance abuse), offer limited user control, and can lead to gender-based ad discrimination—such as fewer high-paying job ads for women—without clear accountability due to the black-box nature of the ad ecosystem.

ABSTRACT

To partly address people's concerns over web tracking, Google has created the Ad Settings webpage to provide information about and some choice over the profiles Google creates on users. We present AdFisher, an automated tool that explores how user behaviors, Google's ads, and Ad Settings interact. AdFisher can run browser-based experiments and analyze data using machine learning and significance tests. Our tool uses a rigorous experimental design and statistical analysis to ensure the statistical soundness of our results. We use AdFisher to find that the Ad Settings was opaque about some features of a user's profile, that it does provide some choice on ads, and that these choices can lead to seemingly discriminatory ads. In particular, we found that visiting webpages associated with substance abuse changed the ads shown but not the settings page. We also found that setting the gender to female resulted in getting fewer instances of an ad related to high paying jobs than setting it to male. We cannot determine who caused these findings due to our limited visibility into the ad ecosystem, which includes Google, advertisers, websites, and users. Nevertheless, these results can form the starting point for deeper investigations by either the companies themselves or by regulatory bodies.

Motivation & Objective

• To investigate whether Google's Ad Settings provide meaningful transparency and user control over personalized ads.
• To assess whether user behaviors—especially on sensitive topics—lead to detectable changes in ad content or settings.
• To examine whether the ad system exhibits discriminatory behavior based on gender or sensitive interests.
• To evaluate the effectiveness of automated tools in uncovering hidden behaviors in large-scale, decentralized ad ecosystems.
• To identify systemic issues in ad privacy mechanisms that may require regulatory or platform-level intervention.

Proposed method

• AdFisher automates browser-based experiments using simulated user agents to visit webpages or modify Ad Settings.
• It collects ads and corresponding Ad Settings data from Google for experimental and control groups.
• Machine learning is used to detect statistically significant differences in ad content and settings between groups.
• A significance test is applied to each detected difference to ensure statistical rigor and avoid false positives.
• Experiments are designed with randomization and control to isolate causal effects of user behaviors on ad delivery.
• The tool operates as a black-box analysis, treating the ad ecosystem as opaque, without access to internal logic or data flows.

Experimental results

Research questions

• RQ1 To what extent are Google’s Ad Settings transparent about changes to a user’s inferred profile based on browsing behavior?
• RQ2 Does modifying Ad Settings lead to measurable, statistically significant changes in the ads shown?
• RQ3 Are there instances where user behavior on sensitive topics (e.g., substance abuse) results in discriminatory ad targeting?
• RQ4 Can automated experimentation detect unintended consequences of ad targeting algorithms, such as remarketing based on sensitive content?
• RQ5 Who is responsible when discriminatory or opaque behavior emerges in a decentralized ad ecosystem?

Key findings

• Visiting websites related to substance abuse led to changes in the ads shown, but the Ad Settings page did not reflect these changes, indicating opacity in profile transparency.
• Setting gender to female resulted in significantly fewer instances of ads related to high-paying jobs compared to setting it to male, suggesting potential gender-based discrimination.
• The ad system showed signs of remarketing behavior after visits to substance abuse sites, despite Google’s stated policy against targeting sensitive health information.
• The tool detected statistically significant differences in ad content and settings, demonstrating that automated experimentation can uncover hidden behavioral patterns in ad delivery.
• Google’s Ad Settings provide limited user control, as changes were not always reflected in the ads shown, and the system’s internal logic remains largely inaccessible.
• The findings suggest that even without explicit advertiser misconduct, large-scale machine learning systems can produce discriminatory outcomes due to indirect correlations and lack of oversight.