News Release

Google Ads algorithm misses Spanish speakers for SNAP benefits

Reports and Proceedings

Cornell University

ITHACA, N.Y. – A Cornell University-led research team has discovered that the algorithm behind Google Ads charged significantly more to deliver online ads to Spanish-speaking people in California about the benefits of SNAP, formerly known as food stamps.

“SNAP is a really important resource to get right,” said Allison Koenecke, lead author of the study and assistant professor of information science. “When faced with an algorithm that has disparate impact, our research asks, how do you pick a strategy to interact with the algorithm to equitably recruit SNAP applicants?”

Californians can apply for SNAP benefits using a website called GetCalFresh, which is developed and managed by Code for America, a civic tech nonprofit that builds digital tools and services for community leaders and governments. Code for America primarily recruits GetCalFresh applicants through Google Ads – for example, spending roughly $400 daily to reach anyone from San Diego County who punches key words and phrases like “how to apply for food stamps” into Google.

However, despite GetCalFresh being offered in multiple languages, Spanish-speakers were filling out proportionally fewer applications than English-speakers. In San Diego County, 23% of families living below the poverty line speak Spanish as their primary language, and yet just 7% had applied for SNAP via GetCalFresh, researchers said.

Koenecke and her collaborators discovered one possible reason: the default, dollar-stretching algorithm behind Google Ads was working too efficiently and disregarding Spanish-speaking people in the process.

When Google Ads is configured to garner the most SNAP enrollments per dollar, it ends up delivering fewer ads to prospective Spanish-speaking applicants because such ads cost more than those for English speakers, the team found. At the time, for every $1 spent on Google Ads to “convert” an English-speaking applicant into a SNAP benefits holder, it cost $3.80 to convert a Spanish-speaking person – nearly four times more. Another bidding option on the Google Ads platform cost 1.4 times more to reach Spanish-speakers versus English-speakers.

Koenecke and her collaborators can’t definitively explain the difference, since Google Ads is a black box – a proprietary machine-learning tool outside of public review. It could be attributed to any number of factors, like supply and demand or a bug in the system, she said.

For GetCalFresh, the research findings pose an important ethical question regarding how to spend its limited online advertising budget: Should they reach as many Californians as cheaply as possible, even if that means fewer Spanish-speaking applicants, or advertise more to Spanish-speakers, even if that yields fewer total applicants?

Trade-offs such as these are at the heart of Koenecke’s research into fairness and algorithmic systems, which are increasingly being used to help with decision-making in areas with real consequences, like health care, banking and child services. But without additional scrutiny, algorithms – including a seemingly harmless one behind an advertising platform – can exacerbate inequality or produce results that run counter to what people actually want or need, she said.

As a result of the team’s findings, Code for America adjusted its online advertising strategy to directly target more Spanish-speaking prospective applicants.

“It’s important for the field and the public to have productive dialogues about the kinds of metrics we should be using in these algorithmic systems,” she said. “The communities most impacted by the algorithms should be given more power in the decision-making process.”

This research was partly funded by the National Science Foundation and Stanford University.

For additional information, see this Cornell Chronicle story.


Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.