Source-linked AI summary
Automated Experiments on Ad Privacy Settings: A Tale of Opacity, Choice, and Discrimination
Amit Datta, Michael Carl Tschantz, Anupam Datta
TL;DR
Web tracking and personalized advertising raise questions about whether Google’s Ad Settings accurately describe user profiles and provide meaningful control. The paper introduces AdFisher, which runs automated controlled experiments and analyzes ads and settings with machine learning and statistical tests. It finds opacity, user choice, and seemingly discriminatory ad differences, while limiting attribution and generalization because the advertising ecosystem is only partly observable.
Problem
Google’s Ad Settings provide information and some control over inferred profiles, but the paper investigates how completely they describe those profiles and whether they provide transparency, choice, and nondiscrimination.
Method
AdFisher automates randomized, controlled experiments that vary browsing behavior or Ad Settings, collect ads and settings, and analyze group differences using machine learning and statistical tests.
Results
The experiments found opacity, meaningful ad choice, and statistically significant evidence suggestive of discrimination in Google’s targeted advertising.
Takeaways & Limitations
AdFisher provides a scalable way to investigate subtle differences between ads and settings and to explain which ad features distinguish experimental groups.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper investigates whether Google’s Ad Settings accurately describe profiling, provide meaningful control, and avoid discriminatory ad outcomes. It introduces AdFisher, an automated framework for controlled experiments, statistical analysis, and machine-learning explanations of differences in ads and settings.
- Motivation: Google’s Ad Settings display inferred demographics and interests, but provide limited information about how completely they describe a user’s profile.The paper frames transparency, choice, and discrimination as central properties to examine because they affect how users understand and control personalized advertising.
- Approach: AdFisher automates randomized, controlled browser experiments that vary user behaviors or settings and collect resulting ads and Ad Settings.It uses machine learning to identify differentiating patterns and statistical analysis to test whether differences between agent groups are significant.
- Findings: Visiting substance-abuse webpages changed the ads shown, while the corresponding Ad Settings pages showed no differences, revealing opacity about this profiling factor.The experiment found no significant or otherwise observable settings difference despite different ad distributions.
- Findings: Removing an online-dating interest changed the top ads shown, indicating that Ad Settings provide users a degree of choice over advertising.Both groups visited an online-dating site, but only one removed the related interest before collecting ads.
- Findings: Setting gender to female rather than male led to fewer ads from a career-coaching agency promising large salaries, a finding suggestive of discrimination.The study reports statistically significant evidence of an instance of discrimination when demographic information was supplied through Ad Settings, while acknowledging disagreement about the normative meaning of discrimination.
- Scope: The authors limit their claims to statistically sound detected effects and cannot assign responsibility within the distributed advertising ecosystem.They do not claim the findings generalize or represent widespread issues, and note that Google or advertisers could be responsible.
2 Prior Work
Prior studies examined Google’s ads, settings, behavioral targeting, or statistical methods, but generally did not combine rigorous large-scale analysis of ad changes with how personalized settings affect those ads. The paper positions AdFisher as addressing that combined scope.
- Google ads and settings: Wills and Tatar found opacity in Google’s ad settings, including ads changing without corresponding changes in Ad Settings, but their study was mostly manual and small scale.Their work is identified as the closest subject-matter comparison.
- Google ads and settings: Other studies focused on how webpage visits change ads, whereas this paper also examines interactions between ads and personalized ad settings.The distinction concerns the broader relationship among behaviors, ads, and settings.
- Google ads and settings: Barford et al. analyzed many ads, websites, and simulated interests, but examined settings mainly to determine whether an interest was successfully induced.This paper instead rigorously studies how settings affect the ads shown, including user choice.
- Methodological precedents: Tschantz et al. developed a rigorous methodology but ran small-scale studies and did not study how Ad Settings changes related to received ads.The paper builds upon their methodology while extending the analysis to ads and settings together.
- Methods: Related machine-learning and correlation-based tools identify ad differences or associations, whereas AdFisher uses machine learning within statistical analyses aimed at causal relations.The comparison includes Liu et al.’s content-matching approach and XRay’s correlation analysis.
- Methods: The paper also distinguishes its larger agent populations and rigorous statistical analyses from prior small-agent or independence-assumption studies.These differences are stated relative to Guha et al. and Englehardt et al.
3 Privacy Properties
The paper examines opacity, discrimination, and two forms of user choice in personalized advertising. Its tests focus on whether browsing or settings changes affect ads and whether those changes are reflected meaningfully in Ad Settings.
- 3.1 Discrimination: Discrimination is defined operationally as class membership causing a change in ads, without claiming that every such difference is unjust.The paper reserves judgments about the morality of observed differences for discussion rather than presenting a scientific test for unjust discrimination.
- 3.1 Discrimination: The methodology compares many instances of surrogate classes because the researchers cannot create actual people varying by traditional discrimination attributes.For example, gender effects are studied by altering a gender setting or visiting gender-associated webpages.
- 3.2 Transparency: Opacity occurs when browsing activity significantly changes served ads without changing the corresponding ad settings.The paper studies this extreme form of non-transparency rather than attempting to assess complete transparency.
- 3.3 Choice: Effectful choice requires that changing an Ad Settings value has some effect on the ads shown, even if that effect is not meaningful or related to the setting.A lack of observed difference cannot establish that effectful choice is absent because relevant ads may not exist in the repository.
- 3.3 Choice: Ad choice requires a meaningful, setting-related effect, such as fewer ads containing keywords associated with a removed interest.The fixed-direction requirement permits testing both compliance and violations of ad choice.
4 Methodology
The methodology uses randomized browser-agent experiments to test whether specified inputs cause changes in ads or settings. It combines blocking, machine-learned test statistics, held-out testing, and permutation tests, while acknowledging limits on detection and generalization.
- 4.1 Background: Significance Testing: Randomly assigning control and experimental treatments to browser agents enables comparisons of how inputs affect ads and settings.Agents execute their assigned actions before the experimenter measures outputs such as advertisements.
- 4.1 Background: Significance Testing: The permutation test compares the observed test statistic with values generated by randomly reassigning treatment labels.A result is statistically significant at the 5% level when fewer than 5% of random assignments produce a statistic at least as large as observed.
- 4.2 Blocking: AdFisher uses blocking to scale beyond ten comparable agents by running groups of nearly identical agents in parallel.Because exact permutation testing grows exponentially with the number of blocks, the method estimates p-values by sampling reassignments.
- 4.3 Selecting Test Statistics: For experiments seeking unexpected effects, AdFisher trains a classifier on one subset of measurements and evaluates its accuracy on disjoint testing data.The classifier predicts treatment group from served ads, so accuracy above the 50% guessing rate indicates systematic differences that can be tested for significance.
- 4.4 Statistical Soundness: The analysis avoids assuming that agents or advertisements are independent and identically distributed, matching the experiment’s online-system setting.The reported p-values are upper bounds of 99% confidence intervals for estimated permutation-test p-values.
- 4.5 Scope: The experiments may miss some information use and do not establish that findings generalize to all users or identify which ecosystem participant caused them.The ecosystem is treated as a black box, and detection can be affected by experiment-specific behavior such as a shared IP address.
5 AdFisher
AdFisher automates randomized treatments, browser-based measurements, classification, and explanations for experiments on personalized advertising. It supports configurable data collection and feature representations for analyzing ads and settings across websites.
- AdFisher: AdFisher lets experimenters specify treatments, measurements, and classifiers, while supporting extensions for additional platforms and analyses.Its stock treatments include changing interests, gender, and age settings or visiting webpage lists.
- AdFisher: Each agent starts from a fresh browser, receives a random group assignment, performs its treatment, and has outputs such as ads or settings measured afterward.Control agents can receive a null treatment while experimental agents perform the specified actions.
- Measurements: AdFisher collects Google text ads from supported news sites, with configurable reload counts and waiting intervals.The experiments primarily use the Times of India and repeat some studies on the Guardian.
- Classification: Its automatic classification pipeline splits measurements into training and testing subsets and selects classifiers that distinguish treatment groups from ad data.Implemented feature sets include URLs, URL-title pairs, and word stems; logistic regression with an L2 penalty over URL+title features performed consistently well.
- Explanation: AdFisher explains group differences by identifying the features most useful to the trained classifier.This provides analysts with a qualitative account of how the ads differed between groups.
- Results summary: Table 2 summarizes experiment duration, collected-ad counts, source site, and concluded result.The summarized ad sources are the Times of India and the Guardian.
6 Experiments
AdFisher's experiments found that Google Ad Settings could obscure behaviorally used information, enable users to change ad exposure, and produce gender-associated differences in job advertising. Results were statistically supported in several experiments, but some findings were sensitive to ad-serving changes and the ecosystem's responsibility could not be assigned.
- 6.1 Nondiscrimination: 93% test accuracy and adjusted p-value < 0.00005 showed statistically significant differences between male and female agents in the May gender-and-jobs experiment.The classifier used ads from 900 agents for training and 100 for testing across 100 blocks.
- 6.1 Nondiscrimination: Male agents received 1852 impressions of two $200k+ executive-position ads, compared with 318 impressions for female agents.These URL+title pairs had the highest coefficients for identifying the male group.
- 6.1 Nondiscrimination: The gender-discrimination finding was sensitive to ad-serving changes: the focal ads fell from 2170 to 48 impressions in July and were too infrequent for statistical significance.One ad disappeared completely, although the remaining ads were shown only to males.
- 6.2 Transparency: 81% test accuracy and adjusted p-value < 0.00005 showed that substance-abuse webpage visits changed ads even though none of 500 experimental agents had interests listed in Ad Settings.The settings showed no differences between experimental and control agents, making the observed ad response opaque in this case.
- 6.2 Transparency: The substance-abuse experiment's strongest distinguishing ads were Watershed Rehab ads, shown 3309 times, or 16% of experimental-group ads, and never to controls.Related rehab terms were absent from the control group's top distinguishing ads.
7 Discussion and Conclusion
Across 21 experiments with 17,370 agents and over 600,000 ads, AdFisher found instances of discrimination, opacity, and choice in Google’s targeted ads. The findings raise concerns about opaque profiling and seemingly discriminatory customization, while the authors emphasize that causation and policy violations cannot be assigned conclusively.
- 21 experiments using 17,370 agents and over 600,000 ads found instances of discrimination, opacity, and choice in Google’s targeted advertising.
- The authors cannot determine whether Google, advertisers, or interactions among ecosystem participants caused the observed discrimination, which may also have been inadvertent.
- Male agents received high-paying career-coaching ads more frequently than female agents, which the authors describe as discrimination and connect to concerns about the gender pay gap.
- Adjusted p-value < 0.00005 showed that substance-abuse visits changed ad distributions without corresponding changes in Ad Settings.The result was robust to variations and was described as an extreme case of opacity.
- The black-box experiments did not establish whether suspected remarketing caused the opacity finding, and the authors do not claim generalization or policy violations.
- Ad Settings provided users with a degree of choice about the ads shown, operating as expected in this respect.
8 Future Work
The authors propose extending AdFisher to other advertising systems and media, and using it to investigate price discrimination and behavior-based content differences. They also stress that assigning blame and controlling confounding factors remain difficult, motivating further research into safeguards for large-scale advertising algorithms.
- Future work includes studying Facebook, Bing, and Gmail, analyzing image or flash ads, and detecting price discrimination on sites such as Amazon or Kayak.
- Assigning blame for gender-based ad differences depends on whether Google, the advertiser, or an interaction enabled or produced the targeting.
- Substance-abuse opacity may reflect remarketing, targeting by Google, or content correlated with the visited webpages, so controlling confounding factors remains an open problem.
- The authors hope future research will develop machine-learning algorithms that automatically avoid unacceptable discrimination and provide appropriate transparency.
A Details of Methodology
AdFisher’s methodology models randomized experiments through assignments, test statistics, permutation-based inference, blocking, sampling, and confidence intervals. Classifier-based statistics quantify how well observed responses distinguish experimental from control groups.
- Randomized experiments assign units to treatments through a uniformly random assignment from a set of possible assignments.
- A test statistic maps ordered observations to ordered numbers, allowing the methodology to compare experimental outcomes under alternative assignments.
- Permutation tests compare the observed statistic with values obtained by permuting assignments, under the null hypothesis that treatment assignment should not matter.
- Blocking: Blocking partitions units into equal-sized blocks and restricts assignments to permutations within each block.
- Sampling: When enumerating all assignments is difficult, the method samples a random subset and estimates the permutation-test value from the sampled arrangements.
- Classifier and confidence intervals: The classifier statistic counts correct experimental-versus-control classifications, while sampled large-statistic counts support binomial confidence intervals.
B Holm-Bonferroni Correction
The Holm-Bonferroni correction orders hypotheses by p-value and compares each with a position-dependent significance threshold. It identifies which hypotheses have statistically significant evidence while controlling for multiple testing.
- The correction orders hypotheses from smallest to largest unadjusted p-value before evaluating significance.
- For hypothesis H_k, the adjusted significance level depends on α, the number of hypotheses m, and its ordered position k.
- Hypotheses preceding the cutoff are treated as statistically significant, while hypotheses at or after it are not accepted as significant.
- Adjusted p-values are reported for hypotheses passing the correction, and their values depend on both unadjusted p-values and rank.