Technology Science is an open access forum for any original material dealing primarily with a social, political, personal, or organizational benefit or adverse consequence of technology. Studies that characterize a technology-society clash or present an approach to better harmonize technology and society are especially welcomed. Papers can come from anywhere in the world.

Loading articles...

February 25, 2025

Linguistic Inequity in Facebook Content Moderation

Yi-Ting Lien and Maya Vishwanath

Thematic Figure.png

Translating Facebook content from Mandarin to English using automated tools resulted in significant discrepancies in users’ perception of the content’s harmfulness. In a survey, subjects were presented objectionable (green) and neutral (blue) content in English translation and the original Mandarin and asked whether the content violated Facebook’s community guidelines.

  • Automated Translation and Content Moderation: Facebook's reliance on automated translation tools for moderation of non-English content can lead to significant inaccuracies. This impacts the platform's ability to accurately judge the harmfulness of non-English content, often resulting in wrongful removal or improper oversight of problematic content.
  • Disparities in Perception: The study found that English-speaking reviewers perceived the harmfulness of translated non-English content differently than native language speakers. This may create a linguistic inequity where non-Western users are disproportionately affected by Facebook's moderation practices.
  • Participatory Approach Recommendation: The research suggests that Facebook incorporates a participatory content moderation approach, where local language users can contribute to moderation decisions via surveys. This aims to mitigate the inaccuracies caused by automated translations.
  • Global Challenges in Content Moderation: Facebook faces challenges in moderating content in multiple languages, especially in regions like India and the Mandarin-speaking world. Automated tools and outsourced moderation practices often fail to capture local nuances, leading to further inequity.
  • Survey Results: In the survey, we compared English speakers’ perceptions of translated content with Mandarin speakers' perceptions of the same content in Mandarin. There was a statistically significant difference in how the two groups viewed the permissibility of the content, highlighting the risks of relying on automated translation tools.

February 25, 2025

Facebook’s Search Interventions: Bad in English, Peor en Español

Julia High

Thematic Figure.png

Facebook’s search feature displays interstitial message screens when a user searches using certain terms related to violence, nudity, sexual activity, or the sale of drugs. This study found that English-language searches were much more likely than searches in Spanish to trigger a pop-up intervention of the type shown here.

  • This study explores the following question: How effective are Facebook’s search interventions at responding to harmful queries in non-English languages?
  • Facebook’s search feature is instrumental in separating users from harmful content on the platform, and it is an essential part of the platform’s content moderation ecosystem.
  • The study found that 49.1% of English language words that should violate Meta’s Community Standards triggered search interventions, whereas only 21.1% of similar Spanish words did.
  • When searching in Spanish, it was far easier to access content that violates the Community Standards, including violent, sexually explicit, and harmful content. Furthermore, searches in Spanish rarely triggered the support or deterrence resources offered in response to comparable English searches.
  • Facebook’s search feature allows researchers a straightforward way of exploring the disparity between English and non-English languages without sowing harmful content onto the platform.

February 25, 2025

Risk Perception on Buy-now, Pay-later Platforms

Zoe Brewer and Alexander Arber

Thematic Figure.png

This study assessed how consumers perceived financial risks when using two popular buy-now, pay-later services, Affirm and Afterpay, that are offered at the point of sale by online vendors. Above, a few of Affirm’s merchant partners at the time of the study in 2022.

  • This study investigates consumer understanding of financial risks associated with BNPL services, focusing on Affirm and Afterpay in 2022.
  • BNPL services disproportionately target financially vulnerable consumers, potentially exacerbating debt cycles.
  • In an experiment, most users struggled to comprehend key loan terms, such as payment frequency, number of installments, and the potential impact on credit scores.
  • Authors recommend further research to assess the impact of the May 2024 expansion of Truth in Lending regulations on consumer financial outcomes.

November 10, 2021

The Inaccessibility of Many State Voter Registration Websites

Jinyan Zang, Alyx van der Vorm, and Juan Guzman

Thematic Figure.png

A Map Demonstrating the Accessibility Scores of Each State Voter Registration Website Tested Using VoiceOver on a Mac.

  • Testing with a screen reader tool, VoiceOver on MacOS, revealed that 8 state voter registration websites were inaccessible to a visually impaired user: Alaska, Minnesota, Nebraska, Nevada, New Jersey, New Mexico, North Dakota, Oregon, Kansas, and Mississippi.
  • In addition, 14 state voter registration look-up websites—sites that can be used to look up a voter registration and other election information—were inaccessible with a screen reader: Alaska, Connecticut, Georgia, Mississippi, New Hampshire, New Jersey, New Mexico, North Dakota, Oregon, South Dakota, South Carolina, Texas, Vermont, and Wisconsin.
  • The 5 states that passed all our accessibility tests – California, Maryland, Oklahoma, Virginia and Washington – had strict web accessibility standards as requirements for official state websites.
  • However, 13 out of the 18 states that had either an inaccessible voter registration or voter registration look-up website also had strict web accessibility standards as requirements for state websites.
  • A few simple fixes, such as adding more labels to web forms, next page buttons, and submission buttons, would address many of the problems that made the websites we tested inaccessible when using a screen reader.

October 19, 2021

How Facebook’s Advertising Algorithms Can Discriminate By Race and Ethnicity

Jinyan Zang

Thematic Figure.png

As more stereotypically African-American attributes were added to a customer list used to create a Lookalike audience of North Carolina voters, the sample shares of African-American voters in the corresponding Lookalike audiences offered to Facebook advertisers became more biased, increasing to 93% in 2020 and 94% in 2021.

  • This study examines racial and ethnic biases in the targeting of ads on Facebook’s advertising platform in January 2020 and January 2021, before and after a major July 2020 boycott of Facebook by advertisers over issues of misinformation and civil rights.
  • In 2021, Facebook’s new “African-American Culture” ad targeting option contained 75% fewer White users than the old “African American (US)” option removed the previous year.
  • Facebook’s tools to help advertisers find users similar to their existing customers generated biased target lists that included either more African-Americans or more Whites, depending on which racial group was dominant in an advertiser’s submitted customer list. This was true for both the Lookalike Audience tool and the Special Ad Audience tool that Facebook designed to explicitly not use sensitive demographic attributes when finding similar users.
  • The degree of bias toward African-Americans or Whites in the composition of a Lookalike or Special Ad audience was greater when customer lists of individuals with racially stereotypical names or ZIP codes were used as the basis for each tool.
  • Similarly, lookalike audiences can also be biased toward Asians. An audience generated by using a customer list of Asians with stereotypical names and ZIP codes as the source list was 100% Asian. Lookalike audiences based on source lists of Hispanics also overrepresented Hispanics.

December 17, 2019

Deepfake Bot Submissions to Federal Public Comment Websites Cannot Be Distinguished from Human Submissions

Max Weiss

Thematic Figure.png

Example Deepfake Text generated by the bot that all survey respondents thought was from a human.

  • Publicly available artificial intelligence methods can generate an enormous volume of original, human speech-like topical text "Deepfake Text") that is not based on conventional search-and-replace patterns
  • I created a computer program (a bot) that generated and submitted 1,001 deepfake comments regarding a Medicaid reform waiver to a federal public comment website, stopping submission when these comments comprised more than half of all submitted comments. I then formally withdrew the bot comments
  • When humans were asked to classify a subset of the deepfake comments as human or bot submissions, the results were no better than would have been gotten by random guessing
  • Federal public comment websites currently are unable to detect Deepfake Text once submitted, but technological reforms (e.g., CAPTCHAs) can be implemented to help prevent massive numbers of submissions by bots

December 17, 2019

Investigating the infamous #Pizzagate conspiracy theory

Panagiotis Metaxas and Samantha Finn

Thematic Figure.png

The echo-chamber formed by the highly visible actors promoting the #pizzagate conspiracy theory on Twitter reveals a dense group of accounts who do not question the obvious lies they propagate.

  • Describes the TwitterTrails tool that enables the investigation of rumors on Twitter
  • Using TwitterTrails, analyzes Pizzagate, an infamous conspiracy theory that gained prominence during the 2016 US Presidential election
  • Reveals several previously unknown facts about the story, including the role that a Turkish journalist played in internationalizing the rumor, visualizes the echo chamber of those who promoted it, and describes their identities and connections

November 12, 2018

Information Warfare in the Digital Age: A Study of #SyriaHoax

Sarah P. White

Thematic Figure.png

A Gephi visualization diagram showing the viral and explosive transmission of the #SyriaHoax story on Twitter over a four-day period

  • Examines the spread of two competing narratives on Twitter following the sarin gas attack in Syria in April 2017
  • Finds that false narratives tend to spread via widely networked distribution nodes, many of which are amplifier accounts created for the express purpose of increasing activity around a particular hashtag
  • Russian digital disinformation efforts are the result of new methods applied to old tactics, and tend to target online communities which have the greatest latent potential for agitation
  • Future research would do well to explore the actual impact of such disinformation efforts on either popular opinion or leader decision-making

November 12, 2018

Saying it’s Anonymous Doesn't Make It So: Re-identifications of “anonymized” law school data

Latanya Sweeney, Michael von Loewenfeldt, and Melissa Perry

Thematic Figure.png

Sequence of 4 litmus tests to perform on a Sander Team protocol.

  • Leading data privacy experts produced four protocols that they said were popular ways to render personal information anonymous so it could be shared or sold publicly
  • Experts relied on the HIPAA Safe Harbor, a flawed use of k-anonymity, an enclave, randomization, and standardized statistical values
  • None of their protocols achieved the privacy protection they promised
  • We were able to put names to the records in all of their protocols.

October 08, 2018

A Host of Troubles: Re-Identifying Airbnb Hosts Using Public Data

Aron Szanto and Neel Mehta

Thematic Figure.png

Example of an Airbnb listing that includes first name of the host, town of the listing, and approximate location

  • We randomly sampled 693 Airbnb listings from Wisconsin from urban, suburban, and rural areas with varying population densities
  • We introduced a method to probabilistically re-identify Airbnb hosts using public voter files from Wisconsin and the fuzzed location of Airbnb listings
  • Despite Airbnb’s efforts to protect personal data, we re-identified 94% of cases using the first name and town of resident living closest to the location provided by a listing