Discrimination in online ad delivery

“For example, in an investigation of the delivery of personalized ads for public records doing basic name searches, Sweeney found statistically significant results in the discrimination in ad delivery across two websites based on names alone [15].”

“Sweeney [49] noted that queries for public records on Google resulted in different associated contextual ad-vertisements based on whether the query target had a traditionally African American or Caucasian name; in the former case, advertisements were more likely to contain text relating to criminal…”

“In recent years, numerous studies (see e.g. [Sweeney, 2013; Barocas and Selbst, 2016]) and media articles [Mann and O’Neil, 2016; Angwin et al. , 2016; Levin, 2016] have illustrated and cautioned against algorithmic unfairness.”

“Algorithms may impact society in data-intensive contexts such as web advertising (Sweeney, 2013) and predictive sentencing guidelines (Angwin et al., 2016).”

“Gender bias has been shown to affect several realworld applications relying on automatic language analysis, including online news (Ross and Carter, 2011), advertisements (Sweeney, 2013), abusive language detection (Park et al.”

“Sweeny presented an example of dubious advertisement placement that appeared to exhibit racial discrimination [21].”

“The lack of transparency in data-driven decision-making algorithms can easily conceal fallacies and risks codified in the underlying mathematical models, and nurture inequality, bias, and further division between the privileged and the under-privileged (Sweeney, 2013).”

“Recommender systems are making a critical impact on humans and society, shaping the movies we watch, the news we read, the jobs we seek, and more [32, 59, 60].”

“This is consistent with their responses on defining fairness where the majority made references to objective factors and those that have been focused on for analyses in the most prominent research in the field of fairness, accountability, and transparency [of algorithms].(54,58,59) Maybe the respondents, being student…”

“Past research has successfully identied harmful and discriminatory behaviors in a variety of algorithmic domains, including search engines [52, 54, 60], online advertising [45, 67], facial recognition [10], word embedding [7], and e-commerce [40].”

“…to consumers stemming from the use of AI made by business, such as the risk of replicating societal bias in advertising leading to discrimination [Sweeney, 2013], from the challenges to enforcers in oversight and detection, stemming both from the “black-box” problem, and from background…”

“Although artificially intelligent systems may reduce human error, they may also exacerbate discrimination against minorities by targeting disadvantaged groups or training on data that reflect systematic and persistent biases [20, 21, 22, 23, 24, 25].”

“Bias may come from the c©2016, Copyright is with the authors.”

“Other researchers [17, 46] have demonstrated that these weaknesses are not confined to Facebook, and are present in Google’s ad platform as well.”

“science, technology, engineering, and mathematics) topics, which may affect society for generations; An unfair ad recommendation may even result in racial discrimination [148].”

“…today, algorithmic bias, has been identified as one such problem, where technologies are clearly in conflict with social values and regulations but lack of technological insight are hindering regulation (Sweeney, 2013; Hajian et al., 2016; Zou and Schiebinger, 2018; Clark and Hadfield, 2019).”

“Another example of bias is online advertising that delivers ads of arrest at a significantly higher rate for Black-identifying names than White-identifying names (Sweeney, 2013).”

“Algorithmic audits have exposed biased outcomes in a variety of domains including criminal justice [6, 40, 89], tenant screening [56, 83], healthcare [31, 72], and online advertising [57, 92].”

“In2013,Sweeney’s study on discrimination showed that search results were disproportionately more likely to return results related to arrest when a racially associated name was used as a keyword in online advertisements [Sweeney, 2013].”

“Some examples include Sweeney’s (2013) finding of discrimination in online ad delivery (Sweeney, 2013), Buolamwini and Gebru’s (2018) finding of bias in facial recognition systems that prefer lighter skin colors, and Veale and Binns’ (2017) finding of racial and social bias in using the location of…”

“In Sweeney (2013), to investigate the fairness online of ad delivery, a sensitive feature, race, is predicted from an independent public dataset, the birth records of the state of California.”

“Ads can also be racially discriminatory [45, 54] and involve sensitive categories (contrary to companies’ public statements) [30].”

“advertisers [95], and consider targeting based on demographics to be discriminatory [93].”

“Furthermore, while the operationalization of race using profile pictures is new to the literature, it differs from that of using Black-sounding names (Edelman et al., 2017); digital racial profiling in online ad delivery (Sweeney, 2013); and race-targeted websites (Appiah, 2010).”

“For instance, searches with “black-sounding” names were much more likely to be shown ads suggestive of an arrest record [24].”

“Notably, however, the vast majority is focused on developed-world contexts: online domains such as targeted advertising [48], search engines [19], and facial recognition [4]; and offline domains such as criminal justice [3, 12], child maltreatment [13], and predictive policing [46].”

“…in diverse text data sources (Zhao et al., 2017), and have been proved to propagate to supervised and unsupervised techniques learning from them, including word embeddings (Boluk-basi et al., 2016; Caliskan et al., 2017; Zhao et al., 2018) and end-user applications like online ads (Sweeney, 2013).”

“…issues that emerge in the sociotechnical gap (cf. Ackerman, 2000) between ML models that quantitatively perform well but produce confusing, misleading, biased, on nonsensical results (Barocas & Selbst, 2016; Bucher, 2017; DeVito, Hancock et al., 2018; Eslami et al., 2016, 2015; Sweeney, 2013).”

“And even ad recommenders can exhibit racial discrimination [67].”

“Latanya Sweeney’s classic study of racial discrimination in online ads (Sweeney, 2013) points to the possibility that how a question is asked or where it is asked from could result in inaccurate or harmful answers for marginalized communities (see also Noble, 2018; Benjamin, 2019).”

“Responsible data analysis practices must be transparent [35, 84], allowing a variety of stakeholders, such as end-users, commercial competitors, policy makers, and the public, to scrutinize the data collection and analysis”

“In the era of Machine Learning (ML), gender biases are translated from sourced data to existing algorithms that may re-flect and amplify existing cultural prejudices and inequalities (Sweeney, 2013) by replicating human behavior and perpetuating bias.”

“…advertisements, etc. Bold claims about the potentials of these technologies abound in this new “Summer of AI”, and while they pave the way for innovations, they also entail challenges about issues such as accountability and transparency, discrimination [2, 60], and fairness in decision-making [6].”

“Scholars such as Noble (2018), Benjamin (2019), and Sweeney (2013) insist that user behavior can also train an algorithmic system to produce problematic outputs in the context of machine learning because of the way that machine learning is conducted.”

“and allocational harm (from algorithms working less accurately).(20,21) A study of facial-recognition programs reported that”

“This pattern occurred whether there was an actual arrest associated with the name (Sweeney, 2013).”

“tain types of interests and characteristics of users are used like gender, health and race [20] etc.”

“Data statements were initially proposed by Bender and Friedman [2] in response to growing awareness of the fact that Machine Learning (ML) approaches to language technology bring various risks of harm to both system users and others affected by system use [5, 25, 30].”

“Given current public discourse, we expected the search to yield literature addressing criminal justice, especially related to predictive policing and sentencing procedures [1, 10, 24], in addition to research pertaining to other domains where algorithmic discrimination has prompted controversy: news/journalism, credit…”

“Important references on the subject include Sweeney (2013), O’Neil (2016), Barocas and Selbst (2016), Noble (2018), Eubanks (2018), Buolamwini and Gebru (2018), Ensign, Friedler, Neville, Scheidegger, and Venkatasubramanian (2018), and Benjamin (2019).”

“[68], for instance, exposed discrimination in online ad delivery by Google.”

“Sweeney (2013) concluded their study with ”Why is this discrimination occurring?”

“Sweeny presented an example of dubious advertisement placement that appeared to exhibit racial discrimination [1].”

“Algorithmic awareness - Enhance public awareness and knowledge about ADM systems to empower users [17, 28] - Increase awareness of users [17] - Raise awareness to possible biases inherent in ADM systems [15, 16]”

“The question of fairness in rankings originated from independent audits on recommender systems or search engines, which showed that results could exhibit bias against relevant social groups [57, 33, 21, 40, 35] Our work follows the subsequent work on ranking algorithms that promote fairness of exposure for individual…”

“However, the growing reliance on automated decision-making systems has led to heightened scrutiny concerning issues of fairness and accountability in these models [12, 37, 53].”

“C R ] 1 0”

“Other studies have compared search results for different queries, for example showing that Google searches for first names associated with Black Americans were more likely to yield discriminatory ads that suggested arrest records [9].”

“For example, algorithms can favor mainstream best sellers (Fleder and Hosanagar 2009), promote higher-priced items (Banker and Khetani, 2019), and might even foster social inequalities in markets (Glaser et al., 2021; Sweeney, 2013).”

“First, we briefly sketch the”

“However, if consumers really are performing this cost-benefit analysis and making a conscious decision, then why we do we see such hype and panic around risks and harms caused by technology in the media? Daily news headlines relay injustice [19, 1, 4, 33], personal boundary violations [32], and gloom [26, 18, 14] over…”

“As the proliferation of artificial intelligence continues, algorithmic bias within AI systems has become a popular topic of scientific study [O’Neil, 2017; Friedman and Nissenbaum, 1996; Sweeney, 2013; Hannak et al. , 2014].”

“Effectively searching is not a new information literacy skill, and yet, with search engines and the rise of personalization web tools that create filter bubbles, discerning readers need to understand the architecture of the data they are accessing and exposed to in platforms, in addition to the impact and results of a…”

“Several technical solutions exist to prevent these three types of attacks, such as kanonymity [82,94], differential privacy techniques [28], and the use of synthetic data [95].”

“Another well-known study shows that names used mostly by men and women of colour are much more likely to generate ads related to arrest records [2].”

“Consider, for example, a search engine disproportionately displaying advertisements about criminal records when common African American names are searched [176].”

“Following similar work on racial bias [14, 37, 98], we further generate obviously gendered names for the applicants by selecting the most popular and heavily gender-skewed names from a USA national names database 5 .”

“Further work has made this risk evident as police face databases are biased toward African American faces [28], word embeddings have proven to encode gender stereotypes [10], and even behavioral advertising is codifying unfortunate racial differences [47].”

“For example, names that are used by men and women of color are much more likely to generate ads related to arrest records [3].”

“One example of such discrimination is, when searching the names of people with black last names, the higher likelihood of getting ads suggesting that these people were arrested, or face a problem with justice, even when it did not happen [8].”

“In this vein, numerous comparison-style studies have also been run, touching on diverse subjects such as discrimination in online advertising [18] and the effectiveness of online privacy tools [19].”

“…of bias and discrimination are being debated as ML systems for natural language processing [7], judicial sentencing [22], target advertising [13, 35], image classification [1], and facial recognition [9, 36] have all proven their ability to inherit bias and create disparate treatment across…”