DL Seminar | Lessons from Auditing the Hidden Societal Impacts of Ad Delivery Algorithms

Individual reflections by Laura Hoyos Salazar and Chi Wei Hsiung (scroll below).

By Laura Hoyos Salazar

Cornell Tech

In today’s seminar, Aleksandra Korolova, Assistant Professor of Computer Science and Public Affairs at Princeton University, presented her investigation into ad delivery algorithms and their significant impact on shaping information access and opportunities.

Professor Korolova began by introducing her research hypothesis, suggesting that ad delivery algorithms may inadvertently “skew” delivery, particularly along gender and racial lines, potentially leading to discriminatory outcomes.

She shared her experiment methodology, which involved immersing herself in the advertiser's journey on Facebook, from sign-up to campaign execution. Her findings revealed a two-phase ad delivery process: advertisers selecting their target audience and uploading the creative content, followed by the ad delivery algorithm bidding on behalf of the advertiser to select the winning ad slot. Professor Korolova learned that the winning ad is determined by the highest total value made up of the actual bids, the estimated action rate, and the user value.

To test her hypothesis, Professor Korolova conducted a series of experiments involving a selection of advertisements. For instance, she ran one ad for a Software Engineer role without specifying gender and another for bodybuilding, among others. To mitigate potential confounding factors, she ensured that all ads were deployed simultaneously to the same audience. The primary variable under investigation was the role of the ad algorithm. The findings revealed a consistent bias in ad delivery, with the algorithm favoring specific demographics when delivering the ads.

Discussing her experiments, Professor Korolova concluded that Facebook’s algorithms for ad delivery exhibit gender discrimination highlighting the need for urgent action. She proposed platform auditing as an approach to address this issue, yet acknowledged its challenges concerning privacy and business interest preservation.

The seminar concluded with three key recommendations:

1. Prioritize ad delivery approaches centering on the needs of both end users and advertisers.

2. Restrict advertiser targeting capabilities, especially in societally important areas such as gender and politics.

3. Advocate for greater scrutiny and transparency from advertising platforms to ensure accountability.

This seminar shed light on an important issue that while targeted advertising can generate productive outcomes for advertisers, it can also create discriminatory impacts on information access and opportunities. Therefore, as ad delivery algorithms continue to evolve, it is imperative to adopt an ethical lens to ensure accountability in business practices.

By Chi Wei Hsiung

Cornell Tech

The seminar dives into the complex issue of discrimination through optimization in digital

advertising. The researchers meticulously demonstrated that ad delivery algorithms on platforms like Facebook could lead to discriminatory outcomes, even when advertisers set neutral targeting parameters.

The focus of the study was on how these platforms distribute ads, especially for sensitive

categories like employment and housing, which are legally required to be free from

discriminatory biases. The research revealed that despite neutral settings by advertisers, the ad delivery systems skew the visibility of ads across different demographic lines, particularly

affecting gender and racial groups. This skew is largely a byproduct of the platforms' algorithms aiming to maximize engagement and revenue by predicting and enhancing relevance of ads to users.

The methodology involved running multiple ad campaigns that objectively appeared neutral but resulted in varied delivery skewed towards particular demographic groups. This was shown through extensive data collection and analysis, which highlighted significant biases without direct input from the advertisers to target or exclude based on sensitive characteristics.

The findings from this seminar are particularly significant in the context of digital ethics and the role of technology in society. As digital platforms increasingly become gatekeepers of

information and opportunities, the algorithms that underpin these platforms play a crucial role in shaping social dynamics and individual opportunities. The inherent bias in algorithmic decisions, as demonstrated by the seminar, underscores a critical challenge in digital ethics: ensuring that technology serves the public good while avoiding harm.

The broader implications of such algorithmic biases are profound. They not only perpetuate

existing social inequalities but also complicate the legal landscape around discrimination and

equal opportunity. For instance, skewed ad delivery could influence job prospects, housing

opportunities, and even credit offerings, disproportionately affecting marginalized communities. This raises urgent questions about accountability and transparency in algorithmic decision-making.

Moreover, the seminar's findings contribute to ongoing discussions about the need for regulatory frameworks that can effectively oversee digital platforms. Policymakers and regulators are called upon to develop more robust mechanisms that can prevent such biases from occurring, ensuring that digital advertising platforms comply with both the spirit and the letter of anti-discrimination laws.

The seminar not only highlighted the problem but also suggested potential paths forward,

including the development of algorithmic auditing practices, more transparent model

architectures, and continuous monitoring of ad delivery practices by independent bodies. These steps are crucial in fostering a digital ecosystem that promotes fairness and equality.

This seminar serves as a vital reminder of the dual-edge nature of technology. While digital

platforms have the potential to connect and enrich, they also carry risks of reinforcing societal biases. The responsibility lies with technologists, policymakers, and society at large to vigilantly guide the evolution of digital practices towards more equitable outcomes. As we continue to navigate the complexities of a digital-dependent world, ensuring that our technologies reflect our ethical commitments becomes not just a technical challenge but a moral imperative.