Academic journal article Northwestern University Law Review

Algorithmic Advertising Discrimination

Academic journal article Northwestern University Law Review

Algorithmic Advertising Discrimination

Article excerpt


In May 2015, Google launched Google Photos, a free service that allowed users to upload unlimited numbers of pictures and later search through those images using words.1 Google photos automatically tags each picture with words describing its content based on predictions generated by Google's artificial intelligence image analysis system.2 But barely a month after rolling out its new service, Google suffered a major public embarrassment when a user discovered that Google photos had labeled images of black people as "gorillas."3 Google immediately apologized and promised to fix the problem, and yet, years later, its solution still has not advanced from simply preventing its software from tagging any image as containing a gorilla.4 Apparently, Google could not find a solution to reliably prevent such racist image mislabeling to occur.5

The world has entered a new era of big data and machine learning, where more and more decisions are being made based on patterns algorithmically extracted from large datasets.6 Machine learning systems ingest large amounts of data and learn to make predictions about some element of interest (e.g., a label for an image) based on that data (e.g., attributes of the image itself).7 If these datasets encode the biases of the humans generating the datasets, then machine learning systems trained on those datasets are likely to replicate those biases or even introduce new biases based on patterns that happen to be present in those data.8 But because machine learning systems are not easily inspected or explained,9 such biases may pass largely undetected.

Big data and machine learning have already transformed advertising. The old model of advertising based on newspapers, billboards, and television is declining in favor of a model in which consumers see ads online, such as on Google and Facebook.10 Companies are now recruiting employees by advertising job opportunities through social media, a practice that will likely become more and more commonplace.11 The old advertising model may have targeted a general audience, but the new advertising model targets individual users with extreme precision.

As social media websites increasingly become platforms for employee recruitment, the mechanisms through which they target job ads to users or allow employers to request that those ads be targeted have faced growing scrutiny. Facebook, for example, has over a billion users and makes money by promising advertisers it will show their ads to users likely to click them.12 Thus, its business depends on its ability to effectively target specific ads to individual users. Doing this manually would take a global army of employees, who would not necessarily be effective. Instead, Facebook uses machine learning technology to predict which kinds of ads particular users might click.13

Though it is illegal to target job ads using statutorily defined protected characteristics (such as sex, race, age, and others),14 Facebook has recently faced criticism and legal action for targeting such ads in these exact ways.15 Indeed, Facebook recently settled several such lawsuits, agreeing to change the way housing, employment, and credit ads can be targeted to its users.16 Yet, as this Note will demonstrate, these changes may be insufficient to prevent discrimination.

Furthermore, Facebook is not the only company at risk of discriminating through its use of machine learning systems. Other social media and online advertising companies, such as Google, Twitter, and LinkedIn, will be susceptible to similar claims to the extent they use machine learning to target advertisements. Although these issues can involve discrimination along any characteristic protected by Title VII,17 this Note specifically focuses on sex discrimination and on Facebook's employment advertising algorithms.

The new era of online employment advertising and machine learning is in fundamental tension with Title VII because Title VII is not designed to deal with the ways in which machine learning systems might discriminate. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed


An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.