Threats to Democracy
Entry type: Single project
Country/area: United States
Publishing organisation: The New York Times
Organisation size: Big
Publication date: 2022-01-27
Authors: Ella Koeze, Denise Lu, Charlie Smart, Karen Yourish, Danielle Ivory, Aaron Byrd, Weiyi Cai, Nick Corasaniti, Meg Felling, Rumsey Taylor, Jonathan Weisman, Jennifer Valentino-DeVries, Steve Eder, Ashley Wu
This project included journalists working for the Graphics Desk, the Investigations Desk, the Politics Team and the Video Desk.
Examining threats to American democracy:
[Can You Gerrymander Your Party to Power?](https://www.nytimes.com/interactive/2022/01/27/us/politics/congressional-gerrymandering-redistricting-game-2022.html) (Jan. 27): Journalists created a game on gerrymandering to show how redistricting can favor one party.
[Over 370 Republicans Have Cast Doubt on the 2020 Election](https://www.nytimes.com/interactive/2022/10/13/us/politics/republican-candidates-2020-election-misinformation.html) (Oct. 13): Journalists showed that most Republican candidates questioned or denied the 2020 results.
[For Trump’s Backers in Congress, ‘Devil Terms’ Help Rally Voters](https://www.nytimes.com/2022/10/22/us/politics/republican-election-objectors-rhetoric.html) (Oct. 22): Journalists showed polarizing rhetoric entrenched among Republicans.
[See Which 2020 Election Deniers and Skeptics Won and Lost in the Midterm Elections](https://www.nytimes.com/interactive/2022/11/09/us/politics/election-misinformation-midterms-results.html) (Nov. 7): Journalists showed that many election skeptics faced few penalties at the ballot box.
All of the stories in the series were shared widely on social media and cited by the news media and politicians, as well as non-profit organizations.
The Times received roughly 130,000 reader-submitted maps from the gerrymandering game, and many readers replayed the game multiple times, a testament to how engaging and immersive the experience was. Many readers posted on social media about how “easy” and “scary” it was to gerrymander and disadvantage certain voters. It was the kind of emotional takeaway journalists had hoped an interactive puzzle would elicit. The Times also heard directly from educators who are using the game in their classes to engage students with the dynamics of gerrymandering.
The analysis of more than 550 Republican political candidates was cited in [an amicus brief](https://www.supremecourt.gov/DocketPDF/21/21-1271/243966/20221026120922177_221003a%20Amicus%20Brief%20for%20efiling.pdf) filed at the Supreme Court of the United States by a bipartisan group of former public officials, former judges and election experts from Pennsylvania in support of the respondents in Moore v. Harper, a North Carolina gerrymandering case that could have serious implications for future elections.
The original database used for this analysis powered the reporting for other stories, including on the night of the midterm elections, and as recently as early January when [20 Republican lawmakers](https://www.nytimes.com/interactive/2023/01/04/us/politics/house-speaker-republicans-vote-against-mccarthy.html) defied Representative Kevin McCarthy and voted against him for Speaker of the House. (In the latest piece, Times reporters used the database to show readers just how far-right that coalition was.)
For election skeptics stories, two reporters manually conducted a sprawling search of each person’s internet footprint, using a mix of publicly available databases, like Facebook, Twitter, DCInbox, congressional websites, Youtube, CSPAN and Rumble. They also used databases requiring subscriptions, like LexisNexis and AdImpact. Some of these websites could be keyword searched, while others required scrolling through dozens of pages, or listening to or watching hundreds of hours of podcasts or videos.
For the polarizing rhetoric story, The Times evaluated nearly three million tweets, more than 100,000 email newsletters, 300,000 Facebook ads and 350,000 statements from the Congressional Record from 2010 through this past June. The analysis employed language software, [Receptiviti](https://www.receptiviti.com/), to tally how often the legislators used words that academic researchers [had linked](https://academic.oup.com/pnasnexus/article/1/1/pgac019/6546199) to antagonistic speech online. To learn about different subjects lawmakers discussed, The Times relied on [topic modeling](https://www.structuraltopicmodel.com/), a machine learning technique that calculates which words are likely to cluster together to form topics. And The Times evaluated religious content using the Linguistic Inquiry and Word Count, a tool that checks texts for related words.
Context about the project:
Redistricting is a routine part of American democracy, yet most people are unaware of how it works or do not realize its implications, even as every American is affected by its longstanding impacts. There have been many text-based explainers on redistricting, but often these stories can be hard to read or easy to gloss over. The Times wanted to create an interactive game, putting readers in the shoes of public officials. By trying their own hand at redistricting a state, readers learned about gerrymandering tactics, such as “packing” and “cracking.” The game provided some constraints based on real-world laws — districts needed to be the same size, minority interests needed to be protected, and compactness was encouraged — but otherwise left it up to readers to discover the best strategies to benefit their assigned political parties. Readers were able to see how changing certain parts of their own custom map could lead to different electoral results.
For the polarizing rhetoric story, it was important not to rely only on the software’s output because doing so could result in errors. Human language is complicated and can be easily misinterpreted by machines. Reporters read thousands of randomly selected tweets, email newsletters and other examples of political speech, checking whether words were indeed being used in a hostile way toward the opposing party. This resulted in some words being removed from the evaluation list. Reporters also performed a [computer analysis](https://www.tidytextmining.com/tfidf.html) on another large sample to determine which words were much more likely to appear in messages that expressed anger, disdain or distrust toward the opposing party. When negative terms were strongly identified with one party — such as “left-wing” — reporters included [a corresponding term](https://ojs.aaai.org/index.php/AAAI/article/view/17748/17555) from the other whenever possible.
For the election skeptics story, the methodology of the data collection had to be refined as the database was built because the reporters discovered along the way that they wanted to capture certain data points that they had not anticipated at the inception of the project. For instance, several weeks into reporting on each candidate, the reporters decided they should draw a distinction between candidates who questioned the election and those who fully denied the results and, unlike several organizations that tracked election denial, Times journalists decided that objecting to the 2020 Electoral College or supporting lawsuits challenging the results would not automatically count as denying the election results outright. [As another Times investigation reported](https://www.nytimes.com/2022/10/03/us/politics/republican-election-objectors.html), those candidates often cited more nuanced arguments for their votes or said they did not want to overturn the outcome. Reporters also decided to track the timing of statements about the 2020 election in order to identify candidates who continued to cast doubt on the contest in the runup to the midterms — something that had not been done by other organizations tracking election denial. In order to draw these sort of distinctions, the reporters had to check each candidate multiple times.
What can other journalists learn from this project?
The skeptics project used both programmatic and hand-curated data collection. Complex projects often require both techniques.
Advanced data analysis techniques can help journalists shed light on problems that seem to defy quantification. Determining whether speech is demonizing, for example, is the type of problem that requires human judgment and thus seems impossible at the necessary scale. But natural language processing, although imperfect, made this analysis more feasible. Still, though the project relied tremendously on computer techniques, it also required many hours of manual work from reporters. Journalists evaluated thousands of randomly selected texts by hand.