A new study into targeted ads has found that the mechanism of delivering such ads may be biased because fewer women than men were shown online ads that promised them help getting jobs paying more than $200,000.
Experiments run by Carnegie Mellon University researchers using an in-house developed software dubbed AdFisher has raised questions on not only the fairness of the targeted ads mechanism, but also the underlying machine learning algorithm that handles operations of the recommendation engine.
According to Anupam Datta, associate professor of computer science and of electrical and computer engineering at CMU said that the discrimination was real, but they were not sure what actually caused this bias. Was it the preference of advertisers? Or was it the unintended consequence of machine learning algorithms that drive online recommendation engines? Datta questions.
The findings were completely “out of the blue” the researchers said as they were not investigating the gender bias. Datta said that the findings were part of a larger study that was looking into the operations of Google’s Ad Settings Web page, previously known as Ad Preferences.
Researchers say that tools such as AdFisher should be used on ad networks to monitor the online advertisement ecosystem and an oversight is needed to ensure that they are not compromising our core values.
AdFisher creates hundreds of simulated users, enabling researchers to run browser-based experiments in which they can identify various effects from changes in preferences or online behavior. AdFisher uses machine learning tools to analyze the results and perform rigorous statistical analyses.
Using the AdFisher tool, researchers ran 21 experiments evaluating Ad Settings web page of Google by creating 1,000 simulated users — half designated male, half female — and had them visit 100 top employment sites. When AdFisher then reviewed the ads that were shown to the simulated users, the site most strongly associated with the male profiles was a career coaching service for executive positions paying more than $200,000.
“The male users were shown the high-paying job ads about 1,800 times, compared to female users who saw those ads about 300 times,” said Amit Datta, a Ph.D. student in electrical and computer engineering. By comparison, the ads most associated with female profiles were for a generic job posting service and an auto dealer.
Researchers are quick to point out that they have found no evidence that Google is doing anything illegal or that it is violating its own policies and AdFisher has simply pointed out discrepancies. The tool can’t explain why they occur without a look inside the black box, he added. Such discrepancies could come from the advertiser or Google’s system selecting to target males.
“We can’t look inside the black box that makes the decisions, but AdFisher can find changes in preferences and changes in the behavior of its virtual users that cause changes in the ads users receive,” said Michael Carl Tschantz, a Ph.D. alumnus of Carnegie Mellon who is now a researcher at the International Computer Science Institute in Berkeley, Calif. Previous researchers have been able to show only a correlation, not causation, between changes in Ad Settings and ads displayed to users.
Anupam Datta is currently working with Microsoft Research to get an inside look at Microsoft’s ad ecosystem. He said he hopes other organizations will use tools such as AdFisher to monitor the behavior of their ad targeting software and that regulatory agencies such as the Federal Trade Commission will use the tool to help spot abuses.
In addition to the findings regarding gender discrimination, another experiment showed that some changes in ads presented to users based on their browsing activity are not transparently explained via the Ad Settings page.
When the simulated users visited Web pages associated with substance abuse, Google subsequently began sending them more ads for a drug rehabilitation center.
Even so, this change was not reflected in Ad Settings, making it impossible for the user to change preferences and halt the substance abuse-related ads. Anupam Datta ( said that is likely because the change in ads was the consequence of remarketing. In remarketing, Google enables advertisers to reach users who have already visited their page. This finding demonstrates that the Ad Settings page doesn’t provide a complete picture of the inferences Google has made about a user. The authors noticed Google started highlighting this limitation on the Ad Settings page a few weeks ago.
In another experiment, the researchers found that adjusting Ad Settings can enable users to avoid some classes of ads they may dislike. They found that simulated users visiting online dating websites could remove that interest on Ad Settings to receive fewer ads related to online dating, giving users some choice over the ads that Google shows to them.
The study is published in the Proceedings on Privacy Enhancing Technologies.
