Source-linked AI summary
Discrimination in Online Ad Delivery
Latanya Sweeney
TL;DR
The paper asks whether personalized arrest-suggestive ads are delivered disproportionately for racially associated names. It tests this with real full names across Google AdSense-hosted sites and finds statistically significant unequal delivery, while noting sampling and replicability limits.
Problem
The paper investigates whether online ads implying an arrest record are delivered differently for Black- and white-associated names, an issue with potential consequences for people without such records.
Method
The study harvested real full names from professionals and online communities, then tested ad delivery for a sample of racially associated names across Google and Reuters.
Results
25% more likely: on Reuters.com, a Black-identifying name received an ad suggestive of an arrest record, with the difference statistically significant.
Takeaways & Limitations
The findings reject the no-difference hypothesis and report discrimination in delivery of these ads across the studied sites and sample.
Takeaways & Limitations
The paper says more research is needed, and its full-name harvesting was uneven, with some names contributing substantially fewer names than others.
Abstract
from arXiv · showhide
A Google search for a person's name, such as "Trevon Jones", may yield a personalized ad for public records about Trevon that may be neutral, such as "Looking for Trevon Jones?", or may be suggestive of an arrest record, such as "Trevon Jones, Arrested?". This writing investigates the delivery of these kinds of ads by Google AdSense using a sample of racially associated names and finds statistically significant discrimination in ad delivery based on searches of 2184 racially associated personal names across two websites. First names, assigned at birth to more black or white babies, are found predictive of race (88% black, 96% white), and those assigned primarily to black babies, such as DeShawn, Darnell and Jermaine, generated ads suggestive of an arrest in 81 to 86 percent of name searches on one website and 92 to 95 percent on the other, while those assigned at birth primarily to whites, such as Geoffrey, Jill and Emma, generated more neutral copy: the word "arrest" appeared in 23 to 29 percent of name searches on one site and 0 to 60 percent on the other. On the more ad trafficked website, a black-identifying name was 25% more likely to get an ad suggestive of an arrest record. A few names did not follow these patterns. All ads return results for actual individuals and ads appear regardless of whether the name has an arrest record in the company's database. The company maintains Google received the same ad text for groups of last names (not first names), raising questions as to whether Google's technology exposes racial bias.
Introduction
The paper frames arrest-suggestive online ads as a potential source of racial discrimination, including when they appear for people without arrest records. It investigates this concern as a socio-technical problem involving technology design and unequal effects.
- Introduction: Online ads can imply a person has a criminal record even when that person has none, potentially affecting employment, housing, lending, and other opportunities.The paper notes that such implications may appear alongside a person’s accomplishments and may not appear for competitors.
- Introduction: The legal protections discussed do not straightforwardly resolve ads that suggest an arrest record without an actual record.The paper distinguishes malicious prosecution and Title VII, noting limits on when either approach applies.
- Introduction: The paper asks whether unequal delivery of commercial ads can constitute racial discrimination when people are not equally affected.This question remains relevant even under the assumption that the ads are protected commercial speech.
- Introduction: Institutional or structural racism can arise from systems whose procedures or patterns produce discriminatory outcomes, even without intentional discrimination.The paper applies this framing to online activity and technology design.
- Introduction: The investigation examines whether arrest-suggestive ads appear more often for one racial group than another among racially associated names.The paper describes online ad delivery as a socio-technical construct requiring sociology and computer science.
Problem Statement
The paper begins with examples in which arrest-suggestive ads cluster around names associated with Black people, while neutral or absent ads occur for names associated with white people. These examples motivate a broader statistical assessment independent of whether a database contains an arrest record.
- Problem Statement: “Latanya Farrell,” “Latanya Sweeney,” and “Latanya Lockett” produced arrest-suggestive ads, although the database showed no arrest record for Farrell or Sweeney.The ads appeared on Google and on Reuters, a newspaper website supplied by Google.
- Problem Statement: Searches for “Kristen Haring,” “Kristen Sparrow,” and “Kristen Lindquist” produced no Instant Checkmate ads, despite database records for all three names.The database listed arrest records for Sparrow and Lindquist.
- Problem Statement: Searches for “Jill Foley,” “Jill Schneider,” and “Jill James” produced neutral Instant Checkmate ads without the word “arrest,” despite database arrest records for all three.These examples contrast ad wording with the records reported in the company database.
- Problem Statement: Google image results appeared predominantly Black for “Latanya” and “Latisha” and predominantly white for “Kristen” and “Jill,” providing a race-associated proxy for the names.The paper uses these image patterns to characterize the names’ racial associations.
- Problem Statement: Handpicked examples suggest arrest-related ads tend to accompany Black-associated names, while neutral or absent ads tend to accompany white-associated names.The paper presents these examples as a suspected pattern requiring scholarly and statistical assessment.
Google AdSense
Google AdSense dynamically matches ads to search-related information through interactions among sponsors, Google, and host websites. Its auction and delivery system can show different ads to different readers and share click revenue with hosts.
- Google AdSense: Different readers, or the same reader returning to a site, may see different ads because online ad space is dynamic and tailored to available information.This contrasts with fixed newspaper and magazine advertising.
- Google AdSense: Google AdSense places dynamic advertisements from millions of sponsors on millions of websites.The paper identifies AdSense as the program delivering the ads studied.
- Google AdSense: Sponsors provide search criteria, possible ad copy, and a bid, while Google conducts a real-time auction among bids for matching criteria.Google may choose not to show an ad when the bid is too low or showing it exceeds a stated limit.
- Google AdSense: A website sends information such as a reader’s search criteria to Google and receives corresponding ads in return.Reuters is presented as an AdSense host displaying Google-delivered ads.
- Google AdSense: When a reader clicks an ad, the sponsor pays its bid, which Google splits with the host website.The paper illustrates this arrangement with a hypothetical click on the “Latanya Sweeney” ad on Reuters.
Search Criteria
The ads studied were tied to exact first-and-last-name searches rather than merely first names. The paper therefore constructs a sample of real full names to test delivery across racially associated name groups.
- Search Criteria: Google’s explanations indicated that the ads matched the exact first-and-last-name combination searched.The paper examined the explanations for ads in the relevant figures.
- Search Criteria: The associated search criteria required both first and last names, used names of real people, and could favor people with an online identity.These criteria define the operational conditions summarized by the paper.
- Search Criteria: The study’s racially associated name list was a qualified sample for testing ad delivery, not evidence that advertisers or Google used such a list.The paper explicitly distinguishes its sampling list from the systems’ bidding or delivery inputs.
Black and White Identifying Names
The study defines racially identifying first names using prior birth-record research and augments those lists with observed names to create a 63-name test set.
- Black- and white-identifying names are defined by significantly higher frequency in one racial group than another.The study draws on heavily cited prior academic work for exemplars.
- 50% more interview callbacks went to white-identifying names than black-identifying names in Bertrand and Mullainathan’s otherwise-identical resume experiment.That Job Discrimination Study used Massachusetts birth-name frequencies from 1974–1979 to classify names.
- Fryer and Levitt’s lists of blackest- and whitest-identifying names were based on California birth records covering more than 16 million births from 1961–2000.Their work also reported a change in Black naming patterns beginning in the 1970s.
Full Names of Real People
The study harvested real full names by pairing racially associated first names with web-based searches of professionals and netizens, then assessed associated images and sample composition.
- Full-name harvesting: Testing ad delivery required real first-and-last names, so the study harvested names of professionals and people active on social media and blogs.Google searches supplied professional names, while PeekYou supplied netizen names.
- Professionals: Professional-name harvesting searched first names with PhD, MD, JD, or MBA, collected up to three result pages, and recorded linked-page images as black, white, or other.Additional degree searches were used when fewer than 10 names were found.
- Netizens: PeekYou harvesting searched each racially associated first name, collected up to two pages while avoiding duplicates, and recorded the apparent race of associated images.PeekYou ranks people by the size of their online footprint.
- Harvest results: 2184 full names were harvested, and the first names predicted apparent race at 88% for black-associated names and 96% for white-associated names.The names were collected from September 24 through October 22, 2012.
- Harvest results: White first names contributed more names on average, while the least-contributing names included Hakim (17), Rasheed (17), Precious (12), Nia (11), and Kenya (4).The study reports an overall average of 35 names per first name, with a median of 30 and standard deviation of 16.
- Sample composition: Of the 2184 names, 835 (38%) were associated with black first names and 1349 (62%) with white first names; 1075 (49%) were male and 1109 (51%) female.Images were discernible for 1428 names, including 508 black, 881 white, and 39 other images.
- Predictive validity: Some black-associated names predicted black images perfectly, while Jamal (48%) and Leroy (50%) were the weakest black-name predictors.The study reports that 12 of 31 white-associated names also made perfect predictions.
Ad Delivery
Across 2,184 racially associated names searched on Google and Reuters, public-record ads were widespread, but delivery, placement, copy, and arrest suggestiveness varied substantially by site and name-associated race.
- Ad volume: 84% of 2,184 names received at least one ad, while 342 names received none; Reuters showed ads for 1,826 names versus 622 on Google.A total of 5,337 ads appeared: 4,473 on Reuters and 864 on Google.
- Ad volume: 78% of captured ads concerned public records, and 1,705 of 2,184 names received at least one public-record ad.These records may include addresses, phone numbers, criminal history, and professional or business licenses.
- Ad placement: Instant Checkmate occupied the top Reuters ad position in 892 of 1,826 searches with ads, or 49%.PublicRecords was next, appearing in the topmost position 142 times.
- Race-associated delivery: Public-record ads appeared disproportionately more often for black-identifying names: PeopleSmart 41% versus 29%, PublicRecords 54% versus 44%, and Instant Checkmate 72% versus 69%.The comparison is reported across black- and white-identifying first names, regardless of company.
- Arrest-suggestive copy: On Google, 92% of Instant Checkmate ads for black-identifying names were arrest-suggestive, versus 80% for white-identifying names; the difference was statistically significant.Overall, 90% of the 432 Instant Checkmate ads on Google were arrest-suggestive; X2(1)=7.71, p < 0.01.
Conclusion and Future Work
The findings reject the hypothesis of equal ad delivery: arrest-suggestive ads appeared more often for black-identifying first names, while the mechanisms behind this pattern remain unresolved. The paper calls for further research because online advertising is dynamic and can change easily.
- Findings: A greater percentage of ads containing “arrest” appeared for black-identifying first names than for white-identifying first names, with statistically significant Chi-Square results.On Reuters.com, a black-identifying name was 25% more likely to receive an arrest-suggestive ad, X2(1)=14.32, p < 0.001.
- Open questions: The study leaves unresolved whether advertisers, Google’s optimization, cloud caching, or user clicks produced the observed disparity.The paper explicitly raises these possibilities without determining which mechanism explains the pattern.
- Future work: The paper reports additional captures from 50 hits on 2184 names across 30 websites, while noting that analyzing these data may clarify ad-occurrence distributions.The author states that the basic message remains unchanged: there is discrimination in delivery of these ads.
- Open questions: Instant Checkmate representatives asserted that Google received the same ad text for groups of last names, but not first names.This claim raises questions about how Google’s technology may expose racial bias in ad delivery.
- Broader implications: The paper argues that technology controls the information context in which name-search results, news stories, and dynamic ads appear.It highlights concern when arrest-suggestive ads appear alongside content about children or high school athletes.