Google Ads Algorithm Adjusted to Provide More Equitable Promotion of SNAP Benefits to Spanish Speakers
A recent study led by a research team at Cornell University has uncovered disparities in the Google Ads algorithm, specifically relating to the promotion of SNAP benefits, previously known as food stamps, to Spanish-speaking individuals. The algorithm was found to charge significantly higher rates for delivering online ads about SNAP benefits to Spanish speakers, resulting in fewer applications from this demographic group. This revelation prompted changes to directly target more Spanish speakers in California who are in need of assistance with food expenses.
The research, titled Popular Support for Balancing Equity and Efficiency in Resource Allocation: A Case Study in Online Advertising to Increase Welfare Program Awareness, was presented at the AAAI Conference on Web and Social Media in June. The objective of the study was not only to quantify the disparities in ad costs but also to address how decision-makers can address these issues and allocate ads in a manner that aligns with the desires of the public.
In California, individuals can apply for SNAP benefits through a website called GetCalFresh, which is managed by Code for America, a civic tech nonprofit organization. GetCalFresh primarily utilizes Google Ads to recruit new applicants for the program. However, despite offering GetCalFresh in multiple languages, Spanish-speaking individuals were submitting proportionally fewer applications compared to their English-speaking counterparts. The research team discovered that the default algorithm behind Google Ads was working too efficiently, inadvertently disregarding Spanish-speaking individuals.
The team found that when the algorithm was configured to maximize SNAP enrollments per dollar, it delivered fewer ads to prospective Spanish-speaking applicants due to the higher costs associated with reaching this demographic. While it cost around $1 to convert an English-speaking applicant into a SNAP benefits holder, converting a Spanish-speaking person cost approximately $3.80. Another bidding option on the Google Ads platform was 1.4 times more expensive to reach Spanish speakers.
The team points out that the reasons for these cost disparities are not definitively known since Google Ads operates as a proprietary machine-learning tool beyond public review. Factors like supply and demand or glitches in the system could contribute to the discrepancy.
This discrepancy in ad delivery costs has raised an ethical dilemma for GetCalFresh regarding how to allocate its limited online advertising budget. Should the focus be on reaching as many Californians as possible at a lower cost, even if that means fewer applications from Spanish speakers? Or should the emphasis shift toward advertising more to Spanish speakers, even if it results in fewer total applicants?
These trade-offs lie at the heart of research on fairness and algorithmic systems conducted by lead author Allison Koenecke and her collaborators. Algorithmic systems are increasingly used in areas with significant real-life consequences, such as healthcare, banking, and child services. However, without proper scrutiny, algorithms, even seemingly innocuous ones behind advertising platforms, can perpetuate inequality or yield results that contradict the desires and needs of the public.
To gain insights into the public’s perspective, Koenecke and her collaborators conducted a survey of approximately 1,500 Americans. Respondents across various demographics, including age groups, gender, race, welfare status, and political affiliations, expressed a preference for more equity in advertising SNAP benefits. They were willing to accept a reduction in total enrollments if it meant increasing enrollments among Spanish speakers. Similar results were found among Code for America staff members, indicating a strong desire for equitable access.
As a result of the study, Code for America has adjusted its online advertising strategy to directly target more Spanish-speaking prospective applicants. Koenecke emphasizes the importance of engaging the communities most affected by algorithms in the decision-making process and fostering productive dialogue around the metrics used in algorithmic systems.
The research serves as a reminder that algorithms, even those used in online advertising, can have profound implications for equity and fairness. It highlights the need for continuous examination and public input to ensure that these algorithms align with societal values and goals.
By shedding light on the challenges posed by algorithmic systems, this study contributes to a larger conversation about the intersection of technology, fairness, and decision-making. It also underscores the potential for bipartisanship when considering issues of fairness and equity. The majority of both Republicans and Democrats expressed a preference for some level of equity in the promotion of SNAP benefits, bridging a potential divide and promoting a more inclusive approach to online applications involving algorithmic distribution.
Moving forward, it is vital to continue exploring these issues, advocating for transparency in algorithmic systems, and enabling communities affected by such systems to have a say in decision-making processes. By doing so, society can strive for a more equitable and balanced approach to resource allocation.
