The Repressive Power of Artificial Intelligence
Explore the Report
As the novelty of AI chatbots captured much of the public’s attention, people around the world struggled against a surge of crude digital repression. Nowhere was this clearer than in Iran, which earned the largest score decline of this year’s coverage period. After the death in custody of Jina Mahsa Amini sparked nationwide protests in September 2022, the regime intermittently restricted internet connectivity and blocked WhatsApp and Instagram, the only international social media platforms that had been accessible in the country. But state repression was not limited to the protests: two people were executed for alleged blasphemy after they shared their religious views on Telegram.
Harsh crackdowns on free expression were also routine in China, which retained its title as the worst environment for internet freedom for the ninth year in a row. Among the many people imprisoned for sharing their views online, prominent civic activist and blogger Xu Zhiyong was sentenced to 14 years in prison in April 2023. Meanwhile, censors scrubbed away criticism of the declining economy and discussion of the legislature’s rubber-stamp approval of an unprecedented third presidential term for Xi Jinping. Despite such intense repression, the Chinese people showed inspiring resilience. In November 2022, for example, protests over a deadly fire in Urumqi, the toll of which was reportedly made worse by overly restrictive COVID-19 lockdown measures, grew into one of the most open challenges to the ruling Chinese Communist Party (CCP) in decades, triggering a rare nationwide policy reversal by the central government. People employed creative language on social media to evade censors, launching hashtags like “A4” and “white paper exercise” to evoke the blank sheets of paper that demonstrators raised to protest the extremity of CCP censorship.
- About this Report
This is the 13th edition of Freedom on the Net, an annual study of human rights online. The project assesses internet freedom in 70 countries, accounting for 88 percent of the world’s internet users. This report covers developments between June 2022 and May 2023. More than 85 analysts and advisers contributed to this year’s edition, using a standard methodology to determine each country’s internet freedom score on a 100-point scale, with 21 separate indicators pertaining to obstacles to access, limits on content, and violations of user rights. The Freedom on the Net website features in-depth reports and data on each country’s conditions.
Myanmar was the world’s second most repressive environment for internet freedom this year. Under military rule since a 2021 coup, the country’s internet users continued to express support for the democratic resistance movement or grief for the victims of the junta’s violence, all at great personal risk to themselves. The army and its informants used Telegram groups to share information on such dissidents, allowing the authorities to identify, detain, and in some cases forcibly disappear them. In the most egregious case from the coverage period, the military executed prominent activist Kyaw Min Yu, better known as Ko Jimmy, in July 2022, after arresting him for prodemocracy social media posts. Officials also forced the sale of the last internet service provider in Myanmar that had a degree of independence to a state-linked company in September 2022, clearing the way for implementation of the regime’s censorship without resistance from the private sector.
In a record 55 countries this year, people faced legal repercussions for expressing themselves online. The number of countries in which authorities carry out widespread arrests and impose multiyear prison terms for online activity has sharply increased over the past decade, from 18 in 2014 to 31 in 2023. Belarus received the year’s third-largest score decline, alongside Costa Rica and Nicaragua. A Belarusian court sentenced Maryna Zolatava and Liudmila Chekina—the editor in chief and director general, respectively, of TUT.by, Belarus’s largest independent media outlet—to 12 years in prison for their reporting. In several cases in Nicaragua, President Daniel Ortega’s government forced people who had been incarcerated in part for their critical speech online to choose between staying in prison and being sent into exile without their citizenship. Roman Catholic bishop Rolando José Álvarez Lagos, who had his citizenship revoked but refused to leave Nicaragua, received a 26-year prison sentence for broadcasting prayers on social media about authorities’ crackdown on Catholic clergy, among other offenses.
Elections as a flashpoint for digital repression
Ahead of and during electoral periods, many incumbent leaders criminalized broad categories of speech, blocked access to independent news sites, and imposed other controls over the flow of information to sway balloting in their favor. In the lead-up to July 2023 elections in Cambodia, through which longtime prime minister Hun Sen engineered a transfer of power to his son, authorities blocked access to the news outlets Radio Free Asia, Voice of Democracy, and Cambodia Daily, further cementing the regime’s control over the online media landscape. The Turkish government, led for 20 years by Recep Tayyip Erdoğan and his Justice and Development Party (AKP), enacted a repressive law on disinformation and then wielded it against journalists and a member of the opposition ahead of May 2023 general elections. In November 2022, as Tunisian president Kaïs Saïed prepared to hold the first elections under a new constitution that greatly expanded his own power, authorities threatened an independent news site and detained its director over reporting that criticized the government.
Newly elected leaders also sought to reshape the online environment to their benefit. The Philippines suffered this year’s second-largest decline in internet freedom. After winning office in last year’s election, President Ferdinand “Bongbong” Marcos Jr., son of a Cold War–era dictator, signed a law in October 2022 that required all Filipinos to register their SIM cards under their real name, undermining anonymous communication in what remains a dangerous environment for journalists and activists. Marcos also left in place a blocking order that restricted 27 websites, including several news outlets known for critical reporting; it had first been imposed by outgoing president Rodrigo Duterte in June 2022, under an antiterrorism law that has been a frequent tool of government overreach.
Despite being one of the best performers in Freedom on the Net, Costa Rica experienced a recession in internet freedom under the new administration of President Rodrigo Chaves Robles. Self-censorship reportedly increased as his government engaged in harassment of journalists, opposition politicians, and other critics. In one high-profile scandal, the health minister resigned in February 2023 after it was revealed that she had paid someone to harass journalists at three news outlets who reported on government mismanagement.
The coming year will feature a number of consequential elections around the world, and some governments are already attempting to suppress unfavorable speech online. Ahead of Mexico’s July 2024 presidential election, term-limited incumbent Andrés Manuel López Obrador has used his office to fuel online harassment campaigns against opposition figures and undermine the independent election authority, presumably to ensure victory for his party’s candidate.
Iceland remained the best environment for internet freedom for the fifth consecutive year, followed by Estonia. Sri Lanka earned this year’s largest score improvement after authorities did not repeat blocking of social media platforms that had been imposed in April 2022 during mass antigovernment protests. While The Gambia is still ranked Partly Free, it has experienced the most significant improvement over the past decade in Freedom on the Net. The country’s trajectory demonstrates how broader efforts to rebuild democratic institutions after a period of repression can also benefit internet freedom.
Digital activism and civil society advocacy drove real-world improvements for human rights during the coverage period. Online mobilization helped fuel the Georgian people’s movement against a dangerous bill that would have forced civil society groups to register as “foreign agents” if they received more than 20 percent of their funding from abroad. The bill was widely criticized for its similarities to a Russian law. Technology experts in Taiwan notched a victory for transparency after their investigations prompted authorities to admit that police had sidestepped a requirement to seek judicial oversight for website blocks.
The judiciary continued to serve as a bulwark for internet freedom in many countries. In June 2022, Argentina’s highest court reinforced the right of access to information when it struck down a celebrity’s attempt to remove links to news articles about her connection to a corruption scandal from Google search results in the country. Three months later, India’s Supreme Court ordered the government to explain how it determines when to restrict internet access during school exams to counter cheating; the ruling may bring greater clarity to the country’s murky censorship regime. Although Uganda’s government introduced repressive restrictions on online speech during the coverage period—including a law that imposes 20-year prison terms for sharing information about same-sex sexual conduct—the Constitutional Court intervened in January 2023 to repeal a section of the Computer Misuse Act that had been used to imprison people for their critical expression online.
Also during the coverage period, concerted policy action against spyware technology was carried forward by an ongoing flood of revelations about the extent to which human rights defenders, journalists, and government officials have been targeted by the tools. The administration of US president Joseph Biden stood out as a global leader in March 2023, when it issued an executive order barring federal agencies from using commercial spyware that poses a threat to national security or counterintelligence, or that could be employed by foreign governments to violate human rights or target Americans. Months later, US officials also added the spyware firms Intellexa and Cytrox to its Entity List, constraining their ability to do business with US companies in certain circumstances. Finally, 11 democracies—including the United States, Costa Rica, and several European countries—agreed to limit the use of spyware at home, improve information-sharing with industry and civil society, and galvanize allies to adopt similar safeguards.
Purveyors of disinformation are employing AI-generated images, audio, and text, making the truth easier to distort and harder to discern.
In the coming years, these networks of progovernment human commentators and related companies will undoubtedly increase their reliance on AI-based tools that can create text, audio, images, and video en masse. The affordability, ease of use, and accessibility of consumer-facing generative AI technology have lowered the barrier of entry to the disinformation market. Those with the financial or political incentives will have the capacity to create false and misleading information with these tools, and then leverage existing networks to distribute it at scale. As a result, the information space will grow more diluted and polluted with fabrications, further damaging the internet’s role as a source of reliable and diverse content.
Outsourcing to shadowy firms and influencers
State officials have cultivated networks of private actors willing to spread false and misleading content. Rather than taking the political risk or developing the resources to engage in such activity themselves, an electoral campaign, politician, or ministry can simply hire a social media influencer or public relations firm that prioritizes lucrative contracts and political connections over ethical or legal probity.
The Russian private sector has played an ongoing role in spreading disinformation about the Kremlin’s full-scale invasion of Ukraine last year. A sprawling and sophisticated operation known as “Doppelgänger” mimicked German, American, Italian, British, and French media outlets to disseminate false and conspiratorial narratives about European sanctions and Ukrainian refugees, among other topics. Doppelgänger has been linked to a loose group of companies and nonprofits with close ties to Russian authorities. Cyber Front Z, another Russian network, relied on Telegram to task commentators with sharing hundreds of posts a day on other platforms that attack critics of President Vladimir Putin and promote anti-Ukraine propaganda.
Israel is home to a growing market of disinformation-for-hire companies. A 2023 investigation by Forbidden Stories, the Guardian, and Haaretz uncovered the work of an Israel-based firm known as Team Jorge, which reportedly uses an online platform that can automatically create text based on keywords and then mobilize a network of fake social media accounts to promote it. The firm, for instance, disseminated narratives meant to cast doubt on serious allegations that the former director of Mexico’s criminal investigations unit was involved in torture, kidnapping, and falsifying evidence. Similarly in August 2022, Meta linked the Israel-based company Mind Force to a network of accounts active in Angola. They primarily posted in support of the ruling Popular Movement for the Liberation of Angola and against the country’s main opposition party, and a Mind Force employee publicly disclosed that the Angolan government was a client.
Political actors have also worked to exploit the loyalty and trust that ostensibly nonpolitical influencers have cultivated among their social media followers. Ahead of Nigeria’s February 2023 election, influencers were paid—with one reportedly receiving up to $45,000—to spread false narratives linking political candidates with militant or separatist groups. During Kenya’s August 2022 election, influencers gamed social media platforms’ trending functions to boost misleading political hashtags. For instance, the hashtag #ChebukatiCannotBeTrusted sought to undermine the country’s independent electoral authority by suggesting that its leader supported one presidential candidate over the others. Similar networks of influencers were found to have coordinated disinformation campaigns against Kenyan journalists, judges, and members of civil society.
Creating deepfakes to sow doubt, discredit opponents, and manufacture public support
The growing use of generative AI is likely to compound the impact that these existing networks of progovernment commentators have on information integrity and healthy public debate. During the coverage period, AI-based tools that can generate images, text, or audio were utilized in at least 16 countries to distort information on political or social issues. It takes time for governments and the private actors they employ to incorporate new technology into content manipulation, and the early dominance of English-language tools may slow adoption of generative AI technology globally. But this tally of countries is also likely an undercount. Researchers, journalists, and fact-checkers have difficulty verifying whether content is generated by AI, in part because many of the companies involved do not require labeling. Similar obstacles can impede attribution of AI-backed manipulation to a specific creator.
Electoral periods and moments of political crisis served as flashpoints for AI-generated content. In May 2023, amid an escalating political conflict in Pakistan between former prime minister Imran Khan and the military-backed establishment, Khan shared an AI-generated video to depict a woman fearlessly facing riot police. In doing so, he sought to boost a narrative that the women of Pakistan stood by him, not the country’s immensely powerful military. During the February 2023 Nigerian elections, an AI-manipulated audio clip spread on social media, purportedly implicating an opposition presidential candidate in plans to rig balloting. The content threatened to inflame both partisan animosity and long-standing doubts about the integrity of the electoral system.
AI-manipulated content was also used to smear electoral opponents in the United States. Accounts affiliated with the campaigns of former president Donald Trump and Florida governor Ron DeSantis, both seeking the Republican Party’s nomination for the 2024 presidential election, shared videos with AI-generated content to undermine each other’s candidacy. One video included three fabricated images of Trump embracing Dr. Anthony Fauci, who led the federal government’s COVID-19 response and remains deeply unpopular among critics of pandemic mitigation measures. By placing the fabricated images alongside three genuine photos, the video muddied the distinction between fact and fiction for Republican primary voters. Similarly, in February 2023, a manipulated video that depicted President Biden making transphobic comments spread rapidly across social media. It was presumably created to discredit Biden among voters who support the rights of transgender Americans, which have been under attack in large parts of the country.
AI companies are already being enlisted for state-linked disinformation campaigns. In early 2023, Venezuelan state media outlets distributed videos on social media that depicted anchors from a nonexistent international English-language channel spreading progovernment messages. The videos were produced using an online AI tool created by Synthesia, in what the company said is a violation of its terms of service. The research firm Graphika has also linked the company to a campaign to spread pro-CCP disinformation via the nonexistent news station Wolf News to audiences in the United States, though the videos in question were of poor quality and did not achieve significant reach.
These uses of deepfakes are consistent with the ways in which unscrupulous political actors have long employed manipulated news content and social media bots to spread false or misleading information. Generative AI tools will continue to build on such older tactics, but they may never replace them entirely. During Turkey’s elections, a vast progovernment media ecosystem and armies of Twitter bots bombarded the information space with content favoring President Erdoğan and the AKP. One widely circulated video consisted of clips spliced together to falsely depict an outlawed Kurdish militant group supporting opposition presidential candidate Kemal Kılıçdaroğlu. Erdoğan touted the video as substantially true despite appearing to acknowledge that it may have been doctored. The incident underscored how fabricated content can skew an information space even when it is known to be inauthentic. The growing use and sophistication of generative AI will make such videos seem more realistic and thus more challenging to debunk in the future.
The consequences of AI-generated disinformation campaigns
Even if deepfakes are obviously fabricated or quickly exposed as such, they still contribute to a decaying information space. They can undermine public trust in democratic processes, incentivize activists and journalists to self-censor, and drown out reliable and independent reporting. AI-generated imagery that sensationalizes outrage on divisive topics can also entrench polarization and other existing tensions within society. In extreme cases, it could galvanize violence against individuals or whole communities. The impact of AI-generated disinformation will deepen as the quality and quantity of the technology’s output continues to exceed the capacity of observers, moderators, or regulators to detect, debunk, or remove it.
Like digital repression more broadly, AI-generated disinformation campaigns disproportionately victimize and vilify segments of society that are already under threat. The overwhelming majority of nonconsensual deepfakes featuring sexual imagery target women, often with the aim of damaging their reputations and driving them out of the public sphere. An online campaign using AI-manipulated pornographic videos was used to discredit prominent Indian journalist and government critic Rana Ayyub as early as 2018. During the coverage period, Nina Jankowicz, a US expert on disinformation, was subjected to pornographic deepfakes as part of a broader campaign against her and her work. These uses of sexualized deepfakes represent a twisted evolution of a much older practice, the nonconsensual distribution of intimate images of women activists. For example, during the coverage period, a smear campaign that featured nonconsensual intimate imagery of Azerbaijani prodemocracy activists and opposition figures spread across Telegram, TikTok, Facebook, and progovernment news sites.
The growing use of AI-generated false and misleading information is exacerbating the challenge of the so-called liar’s dividend, in which widespread wariness of falsehoods on a given topic can muddy the waters to the extent that people disbelieve true statements. For example, political actors have labeled reliable reporting as AI-enabled fakery, or spread manipulated content to sow doubt about very similar genuine content. In April 2023, the Indian state of Tamil Nadu was rattled by a political controversy after leaked recordings captured Palanivel Thiagarajan, a prominent official in the state’s ruling Dravida Munnetra Kazhagam party, disparaging his colleagues. Thiagarajan denounced the audio clips as machine generated; independent researchers determined that at least one was authentic. In Ethiopia, after a member of the ruling Prosperity Party was killed in the Amhara region in April 2023, state-affiliated media outlets released audio that purportedly linked a militia group to the killing. An organization released a report alleging that the audio recordings were manipulated using AI; another fact-checking group then accused the organization of being fake, casting doubt on the claim of AI manipulation.
Companies like OpenAI and Google have imposed guardrails to reduce some overtly harmful uses of their chatbots. But researchers have identified techniques to break through popular chatbots’ safeguards so that they generate harmful, false, discriminatory, or abusive text, including misinformation about the COVID-19 pandemic and statements that mirror Russian propaganda about the invasion of Ukraine. Generative AI applications have also produced false and harmful statements about prominent members of civil society. The dangers of AI-assisted disinformation campaigns will skyrocket as malicious actors develop additional ways to bypass safeguards and exploit open-source models, and as other companies release competing applications with fewer protections in place.
Key Internet Controls
To track the different ways in which governments seek to dominate the digital sphere, Freedom House monitors their application of nine Key Internet Controls. The resulting data reveal trends in the expansion and diversification of these constraints on internet freedom.
Innovations in the AI field have allowed governments to carry out more precise censorship that is less detectable, minimizing public backlash and reducing the political cost to those in power.
While AI has allowed for more subtle and efficient forms of content removal, blunt censorship remains pervasive. Shutdowns of internet service and blocks on entire social media platforms continued to be key tactics of information control around the world. The number of countries where governments imposed outright blocking on websites that hosted political, social, and religious speech reached an unprecedented high of 41 this year. Democracies are not immune to this trend. States that have long been defenders of internet freedom imposed censorship or flirted with proposals to do so, an unhelpful response to genuine threats of foreign interference, disinformation, and harassment.
Generative AI draws authoritarian attention
This coverage period’s newly launched text-based generative AI applications like ChatGPT and Bard may allow users to sidestep government censorship, because the systems are trained on data from the global internet, including information that is typically suppressed in authoritarian states. The most technically sophisticated authoritarian regimes grappled with this risk, and some attempted to restrict access to the new chatbots. In February 2023, Chinese regulators ordered tech conglomerates Tencent and Ant Group to ensure that ChatGPT is not integrated into or accessible via their services, including through the third-party apps on their app stores. Apple removed over 100 ChatGPT-like apps from its Chinese app store to comply with local rules. Similarly, Vietnamese officials warned citizens against using ChatGPT, asserting that it “distorts,” smears, and opposes the state and the ruling Communist Party of Vietnam (CPV).
Early research indicates that chatbots’ outputs reflect the censorship embedded in their training data, a reminder that generative AI tools influenced by state-controlled information sources could serve as force multipliers for censorship. Developers in some repressive countries have moved to create their own chatbots. For instance, several Russian companies—including Yandex and Sistemma—have launched such products. The Chinese government has sought to regulate training data directly: Chinese consumer-facing generative AI products, like Baidu’s ERNIE Bot and Alibaba’s Tongyi Qianwen, are required to implement stringent content controls and ensure the “truth, accuracy, objectivity, and diversity” of training data, as defined by the CCP. Indeed, chatbots produced by China-based companies have refused to engage with user prompts on sensitive subjects like Tiananmen Square and have parroted CCP claims about Taiwan.
Emerging state attempts to control chatbots mirror previous efforts to restrict new social media platforms, with the world’s most technically sophisticated authoritarian governments leading the charge. As generative AI-based tools become more accessible and widely used, a growing number of governments will focus on ensuring that they reinforce rather than challenge existing information controls.
Ordering platforms to use AI for censorship
In at least 22 countries, social media companies were required—either explicitly or indirectly through the imposition of tight deadlines for the removal of banned material—to use automated systems for content moderation. While such systems are used by social media platforms around the world, the laws in many countries prohibit forms of political, social, and religious speech that should be protected under international human rights standards. By obliging platforms to use machine learning to comply with censorship rules, governments are effectively forcing them to detect and remove banned speech more efficiently. This use of AI also masks the role of the state in censorship and may ease the so-called digital dictator’s dilemma, in which undemocratic leaders must weigh the benefits of imposing online controls against the costs of public anger at such restrictions.
Indian prime minister Narendra Modi and his Hindu nationalist Bharatiya Janata Party have incorporated censorship, including the use of automated systems, into the country’s legal framework. The Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules require large social media platforms to use AI-based moderation tools for broadly defined types of content—such as speech that could undermine public order, decency, morality, or the country’s sovereignty, integrity, and security, or content that officials had previously ordered removed. For instance, in early 2023, authorities ordered YouTube and Twitter to restrict access within India to a British Broadcasting Corporation documentary about communal violence during Modi’s tenure as chief minister of the state of Gujarat. Because the government ordered the restriction of the documentary, the IT Rules require the two platforms to use automated scanning tools to sweep up any additional posts that share the film. As the country prepares for general elections in 2024, the government’s expanding censorship regime is creating an uneven playing field by silencing criticism of and independent reporting on the ruling party.
In more authoritarian contexts, automated censorship systems could close what little space remains for online expression. The Vietnamese government imposes tight controls on digital platforms to curb dissent, independent reporting, and other forms of political and social speech. For example, authorities have reportedly compelled Meta to remove all criticism of specified CPV officials. Regulations passed in August 2022 empower the Ministry of Public Security to block platforms that do not comply with a requirement to remove “toxic” content within one day of being notified, a threat that incentivizes overbroad removal at a pace only achievable through automation. Authorities have since explicitly demanded that companies use AI to remove so-called toxic content.
Such measures are growing in popularity among governments with less robust technological and regulatory capacity. In Nigeria, where authorities have imposed significantly less censorship than their counterparts in Vietnam and India, a code of practice introduced in October 2022 requires companies to remove content within 48 hours of notification from a government agency. The code was introduced after then president Muhammadu Buhari imposed a seven-month block on Twitter because the company removed a post in which he appeared to threaten violence against separatists. It is unclear to what extent the code has been enforced since the election of President Bola Tinubu in February 2023.
Automated systems play a positive and supportive role in conducting content moderation at scale, including by detecting influence operations and reviewing dangerous and harmful content—such as child sexual abuse imagery and depictions of graphic violence—that traumatizes human moderators. However, even when these systems are used appropriately, they can excessively or inconsistently flag online material, especially content in languages other than English or in slang, thus increasing the likelihood that political, social, and religious speech will be removed. To protect against that risk and to ensure that the systems strengthen information integrity, laws covering online content and AI should be grounded in human rights principles, require audits and increased transparency regarding the use and impact of algorithms, and include mechanisms for notice, explanation, redress, and appeal
Conventional forms of censorship endure
Especially during times of crisis or protests, AI-powered moderation and filtering tools may struggle to keep up with a surge of unexpected content and expressions of dissent. Blunter forms of censorship will thus continue to be utilized. During the coverage period, internet connectivity was restricted in at least 16 countries. One of them was Iran, where the regime’s technically advanced censorship apparatus was overwhelmed by mass mobilization in 2022 and 2023, forcing authorities to cut off service. Sudanese authorities similarly restricted access to the internet in April 2023, severing critical communication channels at a time when hundreds of thousands of people were caught in the middle of heavy combat between rival paramilitary and military forces.
Of the 70 countries covered by Freedom on the Net, governments in a record-high 41 blocked websites that hosted political, social, and religious speech. Roskomnadzor, Russia’s media and telecommunications regulator, requires internet service providers to install a unique, government-produced deep packet inspection (DPI) system that enables the blocking of websites across the country. The Kremlin has used this system to block global social media platforms, Ukrainian news sites, and domestic sites that carry any hint of dissent regarding its invasion of Ukraine. The coverage period also featured increased Russian blocking of websites that host LGBT+ content, part of a broader assault on that community in the country. The Belarusian government, which has aided Moscow’s military aggression, has blocked more than 9,000 websites, including a slew of independent news sites and associated mirror sites that are maintained by Belarusian journalists working in exile.
Governments are increasingly blocking digital platforms as a means of compelling them to comply with internet regulations. Indonesian authorities restricted access to Yahoo, the gaming platform Steam, payment processor PayPal, and several other sites in July and August 2022 in order to force compliance with Ministerial Regulation 5, which requires the removal of overly broad categories of prohibited speech under tight deadlines. In Brazil in April 2023, after Telegram failed to hand over user data related to neo-Nazi chat groups, a judge ruled that the platform had violated data retention requirements and ordered it blocked entirely. Another judge reversed the ban days later, imposing a more proportionate daily fine on the company and finding that the wholesale blocking was too broad and unreasonably restrictive.
Concerningly, some democratic governments that have traditionally defended freedom of expression considered or imposed censorship during the coverage period, often citing concerns about foreign interference, fraud, and online safety. A French bill presented in May 2023 would require browsers, not internet service providers, to block websites when so instructed by an administrative body, as opposed to a judicial order. The draft law would force browser developers like Mozilla to create a new technical process for blocking, which could be exploited in France and around the world. A month later, concerns about overly broad censorship arose again when President Emmanuel Macron suggested the possibility of blocking or banning Snapchat, TikTok, and other services used by younger populations amid protests and related violence that followed the police killing of a teenager of Algerian descent. Although no such restrictions were implemented, a democratic leader’s rhetorical endorsement of social media blocks as an appropriate response to protests could help to legitimize this form of censorship globally.
Across the Atlantic, some politicians and state-level governments in the United States pushed for an outright ban on TikTok, owned by the Chinese tech company ByteDance, citing potential threats to national security and the risk that the Chinese government could access Americans’ personal data. Montana became the first state to enact a law that, beginning in January 2024, forces companies like Apple and Google to prevent Montanans from downloading TikTok from their respective app stores. Since it was adopted in May 2023, the law has faced constitutional challenges. At the federal level, meanwhile, proposed bills would provide President Biden with the legal authority to ban TikTok nationwide. While the platform’s ownership by ByteDance raises serious human rights and national security concerns, a ban would undermine the constitutional right to free speech for millions of Americans, and would almost certainly encourage other governments to limit access to specific social media platforms. Such a ban would also fail to address the underlying threats posed by TikTok, the tech sector’s broader data-collection ecosystem, and the multiple ways in which foreign governments and nonstate actors can access and exploit Americans’ information. A more comprehensive, effective, and rights-respecting approach to these problems would include stronger regulatory oversight and new legislation to improve privacy protections, require platform transparency and risk assessments, and guard against weak data security practices.
The lessons learned from the past decade of deliberations on internet governance provide a roadmap for this new era.
The CCP has invested heavily in the AI industry while ensuring that the companies in question will serve its authoritarian priorities. The Cyberspace Administration of China (CAC), a powerful regulatory body, has embarked on a yearslong effort to integrate CCP censorship goals into the country’s content recommendation algorithms, synthetic media, and generative AI tools. For example, the CAC approved 41 suppliers of generative AI services in June 2023, and five chatbots were released to the public in August. Such applications are required to adhere to or promote “core socialist values” and exclude content that is deemed undesirable by the CCP. Similar rules have long been in place for Chinese social media companies.
Since 2021, the EU has developed a sprawling framework that could serve as a global model for AI governance, just as Brussels’s General Data Protection Regulation has become a key reference for data protection laws around the world. The draft Artificial Intelligence Act, which was in final negotiations as of August 2023, would tailor obligations based on the level of risk associated with particular technologies, including facial recognition, recommendation algorithms on social media, chatbots, AI tools that can generate images and videos, and the use of AI in political campaigning. AI products that are deemed to present an unacceptable risk would be banned altogether, including social credit systems, predictive policing tools, and certain uses of biometric surveillance. Technologies with a “high” or “limited” risk would be subject to a spate of pre- and post-market requirements, such as registration and increased transparency.
In the United States, the Biden administration began its development of AI governance with a push for industry self-regulation. The Blueprint for an AI Bill of Rights, released in October 2022, laid out a set of principles to guide AI design, use, and deployment. The guidelines include protections against abusive data practices, ineffective and unsafe systems, and algorithmic discrimination, which occurs when biases embedded in training data are expressed in a program’s outputs. The blueprint also calls for companies to offer people a human alternative to automated systems when appropriate, and to inform people when and how such systems are operating. In July 2023, after the coverage period, the administration secured voluntary commitments from Amazon, Anthropic, Google, Inflection, Meta, Microsoft, and OpenAI regarding AI safety and security. While self-regulation is an important starting point, it must be matched with meaningful oversight. Further executive action was anticipated at the time of writing, but to ensure that AI bolsters rather than harms internet freedom, members of Congress should work with civil society and the executive branch to craft bipartisan legislation that takes a rights-based approach to AI governance and transforms guiding principles into binding law.
A trust and safety deficit diminishes hope for self-regulation
A series of business decisions during the coverage period cast further doubt on the private sector’s willingness and capacity to self-regulate. Across a number of major platforms, teams focused on content, integrity, trust, and safety—including those tasked with setting, maintaining, and enforcing rules for user behavior—experienced a dramatic reduction in staff and resources. Some companies also pared back transparency mechanisms as well as content policies that were intended to reduce the spread of false and misleading information. This shift in priorities suggests that companies have forgotten the lessons from past controversies. The industry’s deficient investments in content moderation around the world, failure to heed calls for extra care during electoral or other sensitive periods, and inadequate adjustments to products and policies in response to crises have all had the effect of inflaming online and offline violence.
Nowhere was this shrinking and reprioritization of resources more dramatic than at X, formerly known as Twitter, after it was purchased by tech investor and entrepreneur Elon Musk in October 2022. As part of its mass layoffs, the company dismissed experts on trust and safety, human rights, public policy, and regional issues in Africa, Asia, and Latin America. X also reduced its transparency with the public, for example by putting its application programming interface (API) behind a paywall, scaling back its own transparency reports about state demands, and pausing its reporting to Lumen, a database that publishes government orders for content removal. Economic concerns during the coverage period also led Meta, Amazon’s Twitch, Google, Snap, and Microsoft to impose mass layoffs, slashing teams that worked on topics such as elections, AI ethics, or content moderation. Many contractors who carried out content moderation, and who already suffered from inadequate pay and resources, were also let go.
Companies appear to be wagering that they can use AI to make up for the loss of human expertise, even when addressing the problems arising from AI itself.
Companies appear to be wagering that they can use AI to make up for the loss of human expertise, even when addressing the problems arising from AI itself. OpenAI has proposed using ChatGPT to moderate online content and develop rules for content moderation. Indeed, the new wave of generative AI tools could be a powerful asset in reviewing content that human reviewers would find distressing. However, human oversight remains critical to ensuring that automated systems and their content removals are not overbroad or discriminatory. Overreliance on large language models for higher-order tasks that require context and nuance, like writing content policy, should be treated with skepticism.
The stakes of this wager are high, especially ahead of 2024 elections in pivotal countries and as AI-generated content grows more prevalent. The firing of regional and legal teams entails a loss of institutional knowledge on political parties, disinformation networks, and regulatory systems. Rolling back content policies intended to provide context for readers will weaken information integrity. And reduced platform transparency will limit civil society’s ability to analyze evolving censorship tactics and push for accountability. Ultimately, some companies may choose to further deprioritize policy development and enforcement outside of their perceived core markets. But decisions that maximize short-term profits while accepting negative human rights consequences will in practice create a more dangerous experience online, undermine a service’s competitive appeal, and threaten the strong rule-of-law environment on which all businesses depend.
Ensuring that AI advances human rights online
Technology cannot be a substitute for governance. Most AI models are highly opaque, dependent on the processing of billions of datapoints, and effectively under the control of a handful of companies that reveal little about their development and training. Their inscrutable structure is fundamentally at odds with democratic values such as transparency, proportionality, and accountability, and their information inputs are often hotbeds of potential bias. Companies that create or deploy AI systems, from the newest startups to the most established giants, should cultivate an understanding of previous efforts to strengthen platform responsibility, which have produced both successes and failures over the past decade. But given the private sector’s natural inclination to focus on profit generation, its AI products require supervision by an informed public, a global group of civil society organizations, and empowered regulators.
Government regulation should be aimed at delivering more transparency, providing effective mechanisms of public oversight, and prioritizing the protection of human rights. When designed and used safely and fairly, AI can help people evade authoritarian censorship, counter false and misleading information, monitor elections to ensure that they are free and credible, and bolster documentation of human rights abuses. To bring about effective and rights-respecting AI governance, civil society should be included from the start. Nonprofit organizations, investigative reporters, and human rights activists have been indispensable players behind past wins for internet freedom. Among other contributions, they can build and sustain public pressure and spur action by legislators, regulators, and the industry.
Policymakers and their civic and private-sector partners should take care not to lose momentum in protecting overall internet freedom, especially as AI technology augments the forces driving its multiyear decline. Indeed, to the extent that AI simply exacerbates existing problems associated with digital repression, existing solutions—ensuring that companies minimize the data they collect, reforming legal frameworks for surveillance and censorship, and strengthening information literacy and platform responsibility—should be thoroughly implemented. An effective defense of internet freedom requires not just developing AI governance systems, but also addressing long-standing threats to privacy, free expression, and access to information that have corroded the broader digital environment.
Note: On November 21, 2023, Freedom House corrected this essay to reflect that the company Synthesia did not create a broadcast purported to be from the nonexistent news channel House of News Español. Rather, clients used a tool created by Synthesia to produce the fake broadcast, in what the company said is a violation of its terms of service.
- A Note on Additional Sources and Data
For this report’s main essay, Freedom House identified several ways in which AI exacerbated digital repression during the coverage period and collected the relevant information across 70 countries covered by Freedom on the Net. The essay’s analysis and data points were partly informed by the individual Freedom on the Net country reports, written by external report authors that are listed on the report's acknowledgements page. Freedom House staff also conducted additional research and drew on the important work of various media groups, civil society organizations, and other experts, including but not limited to the Artificial Intelligence Incident Database, the Center for Democracy and Technology, the Center for European Policy Analysis, Cloudflare, Graphika, NewsGuard, the Open Observatory of Network Interference (OONI), and the OECD AI Policy Observatory. Research and analysis from experts including Robert Chesney, Danielle Citron, Alex C. Engler, Steven Feldstein, Matt Fredrikson, J. Zico Kolter, Odanga Madung, Brian Obilo, Grigore Pop-Eleches, Margaret E.Roberts, Zifan Wang, Lucan A. Way, Samuel C. Woolley, Eddie Yang, and Andy Zou also influenced the essay. Freedom House furthermore wishes to acknowledge the academics, researchers, and civil society activists who have worked to document, respond to, and mitigate AI-driven harms, including those not addressed in this report.
Country-specific data and sources used in the report’s essay can be downloaded below under "report data," and each country report and its relevant footnotes are available here.
Test Your Knowledge!
Test your knowledge of some of the key findings from our 2023 report by taking our quiz.
Download the complete Freedom on the Net 2023 PDF booklet.
Learn how governments and companies can protect internet freedom.
Freedom on the Net is a collaborative effort between Freedom House staff and a network of more than 85 researchers covering 70 countries.
The Freedom on the Net index measures each country’s level of internet freedom based on a set of 21 questions.
Visit our Countries in Detail page to view all Freedom on the Net 2023 scores and read individual country narratives.