As of June 2020, 73 percent of total adults in the United States reported that it was very likely for social media platforms to censor political viewpoints that they find objectionable - this represents only a one percent increase in people who believe so since 2018. In contrast, the share of Republican survey respondents who felt that social media sites intentionally censored political viewpoints increased from 85 percent in 2018 to 90 percent in 2020.

Dataset Details

  Dataset Description

This dataset measures soft censorship (selective omission of information) in large language models (LLMs). It contains responses from 14 state-of-the-art LLMs from different regions (Western countries, China, and Russia) when prompted about political figures in all six official UN languages. The dataset is designed to provide insights into how and when LLMs refuse to provide information or selectively omit details when discussing… See the full description on the dataset page: https://huggingface.co/datasets/aida-ugent/llm-censorship.

In 2023, ***** unique titles were challenged in the United States, of which ** percent occurred in school libraries. Meanwhile, **** percent of challenges made to books in 2023 happened in schools. Censorship attempts were made for a number of reasons, with individuals and institutions taking issue with books that were considered to be anti-police, profane, sexually explicit, or homophobic, to name but a few.

There is ongoing debate over whether authoritarian regimes can maintain control over information given the rise of social media and the Internet. In this debate, China is often cited as a prime example of how authoritarian regimes can retain control, but to date, there has been limited research on whether China’s online censorship strategies can be replicated in other authoritarian regimes. This article shows that China’s ability to censor social media rests on the dominance of domestic firms in China’s market for Internet content. The absence of U.S. social media firms in China allows the Chinese government to engage in censorship through content removal, which can quickly and effectively suppress information. In contrast, for most other regimes, the market for social media is dominated by U.S. multinational firms, e.g., Facebook, YouTube, Twitter, and in these contexts, content removal is an immense challenge. This article then examines the prospects of instituting content removal by developing domestic social media or importing Chinese platforms, and finds that most authoritarian regimes are unlikely to be able to duplicate China’s online censorship efforts.

shawarmas/Censored-Words dataset hosted on Hugging Face and contributed by the HF Datasets community

In 2023, ** percent of censorship challenges made to books in schools and libraries in the United States originated from patrons. Libraries and teachers were among the least likely to challenge a book, with only ***** percent taking issue with book titles based on their author, themes, or subject.

fernandosmither/censoring-examples dataset hosted on Hugging Face and contributed by the HF Datasets community

ShrimpMoss (虾苔) is a dataset designed for the abliteration (https://github.com/FailSpy/abliterator) of Chinese government-imposed censorship and/or propaganda from large language models developed in the PRC. It consists of a series of files of prompts (in .txt, .json, and .parquet format) in two groupings:

china_bad_*: Contains a series of prompts likely to trigger censorship or propaganda actions in the model. china_good_*: Contains a series of prompts in the same general category of topics… See the full description on the dataset page: https://huggingface.co/datasets/Nafnlaus/ShrimpMoss_Chinese_Censorship_Abliteration.

These files include the panel data, a Stata do-file, the experiment data, and an R Markdown file needed to replicate the results from "Talking Politics in a Polarized America: How Perceived Polarization Shapes Political Self-Censorship" by Amber Hye-Yon Lee and Chloe Ahn, accepted for publication in Political Behavior.

A model is proposed to describe observed asymmetries in postwar unemployment time series data. We assume that recession periods, when unemployment increases rapidly, correspond with unobserved positive shocks. The generating mechanism of these latent shocks is a censored regression model, where linear combinations of lagged explanatory variables lead to positive shocks, while otherwise shocks are equal to zero. We apply this censored latent effects autoregression to monthly US unemployment, where the positive shocks are found to be predictable using various leading indicators. The model fits the data well and its out-of-sample forecasts appear to improve on those from alternative models.

Florida was the state with the highest number of books banned in schools in the United States between July 2023 and June 2024, with a total of 4561 books being banned in that period. Iowa ranked second with 3671 book bans in school classrooms and libraries across the state, followed by Texas with 538. The books most likely to be banned or restricted were those which addressed violence or physical abuse, or health and wellbeing for students. New laws, more bans In the first half of the 2022-2023 school year alone, almost 1,500 books were banned. Just one policy contributing to the growing number of challenges to books in schools is Florida’s Parental Rights in Education Act, otherwise known as the “Don’t Say Gay” law, which aims to prevent classroom instruction on gender identity or sexual orientation for children from kindergarten up to grade three. This law, along with several others introduced in states across the United States, are often worded in a manner which is vague or non-specific, leaving librarians, educators, and principals under pressure to swiftly or preemptively remove books on a larger scale (so-called “wholesale bans”) to prevent punishment later. In Utah, after the Sensitive Materials in Schools Act went into effect, school districts were issued guidance to remove any books that are defined as pornography under state statute, which led to the removal of books including Margaret Atwood’s “The Handmaid’s Tale” and “Forever…” by Judy Blume. Indeed, school libraries were the second most likely location to be affected by censorship attempts on books. The audiences most affected by bans In the last half of 2022, most book titles banned in schools in the United States were young adult books aimed at teenagers aged 13 to 17 years old. Over 50 percent of titles banned in that period were intended for this audience.

Dataset 3 phase 2

Literary censorship is not new, but the recent resurgence in the U.S. differs from earlier post-war controversies, many of which were bipartisan and focused on protecting children from exposure to sexual and graphic material. Current censorship controversies in the U.S. appear to be much more partisan and ideological, focused on protecting children from politically offensive ideas. Anecdotally, the political right is portrayed as attacking literature in the name of conservative values like preserving the traditional family from alternative expressions of sexuality and sexual preference, with the left targeting material regarded as contrary to progressive values like tolerance of diverse cultural identities. However, these ideological rationales largely reflect the rhetoric of political activists and have not been empirically tested among the broader public. We conducted two studies designed to measure the polarization of support for literary censorship among the voting-age population of the U.S. Surprisingly, both studies cast doubt on the ideological divisions apparent in activist rhetoric. The survey findings from Study 1 indicate widespread opposition to literary censorship that spans ideological divisions, but both liberals and conservatives were more inclined to support censorship of materials that deviated from their respective values and beliefs. The experiment in Study 2 revealed differences in participants’ responses to liberal and conservative criticisms but little difference in the attitudes of liberal and conservative participants. However, liberals were marginally more likely than conservatives to agree with ideologically aligned literary criticism.

In 2025, Iceland was the worldwide leader in terms of internet freedom. The country ranked first with 94 index points in the Freedom House Index, where each country received a numerical score from 100 (the freest) to 0 (the least free). Estonia ranked second with 91 index points, followed by Chile, with a score of 87 index points. Internet restrictions worldwide The decline of internet freedom in 2022 is mainly linked to political conflicts in different parts of the world. With the Russian invasion of Ukraine, the Russian government intensified its attempts to control the online content in the country. The government placed restrictions on three different U.S.-based social media platforms at the same time, X, formerly known as Twitter, Facebook, and Instagram. These restrictions made it to the top of the longest-lasting limitations on the web in 2022. Social protests rose in Iran following the death of Mahsa Amini in September 2022. The Iranian government decided to shut down the internet and various social media platforms in an attempt to minimize the communication between the protesters. In 2022, 11 new internet restrictions were recorded in Iran. However, residents in the Indian region of Jammu and Kashmir saw the highest number of new internet restrictions, which amounted to more than double the ones in Iran. The impact of internet shutdowns In 2022, the economic impact of internet restrictions worldwide reached an estimated 23.79 billion U.S. dollars. Meanwhile, the highest financial losses due to internet shutdowns were caused by limitations in Russia, and more than seven thousand hours of restricted various online services had an economic impact of 21.59 billion U.S. dollars. The restrictions impacted around 113 million people in the country. Myanmar placed the most extended restriction on internet services, lasting 17,520 hours in total. Similar restrictions in India affected over 120 million people. 

National surveys reveal notable individual differences in U.S. citizens’ attitudes toward freedom of expression, including freedom of the press and speech. Recent theoretical developments and empirical findings suggest that ecological factors impact censorship attitudes in addition to individual difference variables (e.g., education, conservatism), but no research has compared the explanatory power of prominent ecological theories. This study tested climato-economic, parasite stress, and life history theories using four measures of attitudes toward censoring the press and offensive speech obtained from two national surveys in the U.S.A. Neither climate demands nor its interaction with state wealth—two key variables for climato-economic theory—predicted any of the four outcome measures. Interstate parasite stress significantly predicted two, with a marginally significant effect on the third, but the effects became non-significant when the analyses were stratified for race (as a control for extrinsic risks). Teenage birth rates (a proxy of human life history) significantly predicted attitudes toward press freedom during wartime, but the effect was the opposite of what life history theory predicted. While none of the three theories provided a fully successful explanation of individual differences in attitudes toward freedom of expression, parasite stress and life history theories do show potentials. Future research should continue examining the impact of these ecological factors on human psychology by further specifying the mechanisms and developing better measures for those theories.

Dataset Card for Evaluation run of DavidAU/Qwen2.5-MOE-2X1.5B-DeepSeek-Uncensored-Censored-4B

Dataset automatically created during the evaluation run of model DavidAU/Qwen2.5-MOE-2X1.5B-DeepSeek-Uncensored-Censored-4B The dataset is composed of 38 configuration(s), each one corresponding to one of the evaluated task. The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/DavidAU_Qwen2.5-MOE-2X1.5B-DeepSeek-Uncensored-Censored-4B-details.

Unmasking the Imposters: Machine-Generated Tweet Detection Dataset

This dataset contains nine subsets of human and machine-generated tweets designed to evaluate the detection of AI-generated content across censored and uncensored large language models (LLMs). The dataset addresses the gap in understanding how content moderation and domain adaptation affect the detectability of machine-generated text on social media platforms.

Paper: "Unmasking the Imposters: How Censorship and… See the full description on the dataset page: https://huggingface.co/datasets/redasers/Unmasking-the-Imposters.

The MPPDA Digital Archives consists of a database of the extant records of the General Correspondence files of the Motion Picture Producers and Distributors of America, Inc., covering the period from 1922 to 1939. Established in 1922, the MPPDA was the trade association for the motion picture industry, and included all the major companies producing and distributing motion pictures in the United States in its membership. The association was popularly known as the Hays Office, after its first President, Will H. Hays, who remained office until 1945. Hays was a leading Republican party politician, who resigned from President Warren Harding's Cabinet to take up the MPPDA position. After Hays' retirement, the association was renamed the Motion Picture Association of America.The documents in the MPPDA ís General Correspondence files are an immensely rich source of information about the history of the motion picture industry. They describe the organization and operation of the industry's trade association, and include extensive correspondence and other documentation relating to industry policy and public relations, distributor-exhibitor relations, censorship and self-regulation. The great majority of this material is unavailable from other sources.Grants:* 1997, 1998, MPPDA Archive digitisation and database development. Flinders University Establishment Grant to digitise the microfilm of the MPPDA Archive and develop a publicly accessible database.* 2000, 2001, Reforming the Movies: A Political History of the American Cinema, 1908 - 1940. ARC A00104524Date coverage: 1922 - 1939Location: United States

The expert survey of journalists was conducted from 18 to 27 January 2023 using a self-completion questionnaire in Google Forms. The survey was conducted by the Ilko Kucheriv Democratic Initiatives Foundation on the request of the Human Rights Centre ZMINA with the support of Freedom House Ukraine.

A total of 132 people participated in the survey. The respondents were selected using the method of voluntary selection and snowballing to the point of saturation. The sample represents only the opinion of the respondents, but it also allows us to talk about certain trends and common assessments of certain phenomena and processes in the journalistic field.

The survey includes questions about freedom of speech and self-censorship in the media environment during the Russian-Ukrainian war.

The data collection contains original survey data. The Excel file (.xlsx) is the original file with the respondents' answers in Ukrainian, provided by the Ilko Kucheriv Democratic Initiatives Foundation. The documentation includes the questions and answer options of the original questionnaire in Ukrainian and English.

Additionally, the data collection contains the "Summary" file, which is an analytical report prepared by the Ilko Kucheriv Democratic Initiatives Foundation and the Human Rights Centre ZMINA. The report uses data from an expert survey of journalists in 2019 and 2023, and the results of focus groups in 2022.

IntroductionHarm and offense are two important notions in legal discussions on the extent to which one’s freedom may be limited. Prior research on the third-person effect found that perceived media harm on others, not perceived media harm on the self, is a robust positive predictor of support of censoring socially undesirable media content (e.g., pornography). In comparison, how offensiveness perceptions predict censorship support is not clear. Drawing on moral foundations theory, we test here how perceived media offensiveness to the self compared with 1) perceived media offensiveness to others and 2) perceived media harm on others would predict censorship support.MethodWe conducted two cross-sectional survey studies in the U.S. to address this question with sexual, alcoholic, and violent media content as test cases. In Study 1 (N = 544 undergraduates), we measured perceived media offensiveness to the self, that to others, and censorship support. In Study 2 (N = 727 non-student adults), we also measured perceived media harm on the self and others.ResultsAs in prior research, we found that people perceive sexual, alcoholic, and violent media content to harm other viewers more strongly than it harms themselves, and the perception of how much others are harmed predicts perceivers’ censorship support. In contrast, while people also perceive the three types of media content to offend other viewers more strongly than they offend the self, the perception of how much others are offended predicts censorship support to a significantly lesser extent or does not predict this at all. Instead, the perception of how much the self is offended does.DiscussionThese findings add to the work on moral foundations theory that distinguishes between how the care/harm and sanctity/degradation foundations relate to moral judgments. These findings also suggest that the current theorizing of the third-person effect needs to expand to reconcile the seemingly inconsistent results on how harm and offensiveness perceptions differently relate to censorship support. The care/harm and sanctity/degradation foundations may underlie how harm and offensiveness perceptions predict censorship support. However, several “anomalous” findings need to be accounted for before moral foundations provide a comprehensive explanation of the third-person effect.