Tamil is a language with a long history. Spoken by over 80 million people worldwide, or over 1% of the world’s population, early inscriptions in the language date back to the 5th Century B.C.E (Murugan & Visalakshi , 2024). The language is spoken widely in India (predominantly in Tamil Nadu and Puducherry), in Sri Lanka, and across diaspora communities in Malaysia, Thailand, Canada, the United Kingdom, the United States, and beyond. Despite the widespread use of the language, there remains limited understanding of how major social media platforms moderate content in Tamil. This report examines the online experiences of Tamil users and explores the challenges of applying consistent content moderation processes for this language. 

This report is part of a series that examines content moderation within low-resource and indigenous languages in the Global South. Low-resource languages are languages in which sufficient high-quality data is not available to train models, making it difficult to develop robust content moderation systems, particularly automated systems (Nicholas & Bhatia, 2023). In previous case studies conducted in the series, we found that this lack of high-quality and native datasets impeded effective and accurate moderation of Maghrebi Arabic and Kiswahili content (Elswah, 2024a; Elswah, 2024b). Inconsistent and inaccurate content moderation results in lower trust among users in the Global South, and limits their ability to express themselves freely and access information. 

This report dives into Tamil speakers’ experiences on the web, particularly on popular social media platforms and online forums run by Western and Indian companies. We highlight the impact of Tamil speakers’ perception of poor content moderation, particularly against a backdrop of democratic backsliding and growing repression of speech and civic participation in India and Sri Lanka (Vesteinsson, 2024; Nadaradjane, 2022). Ultimately, what emerges in this case study is a fragmented information environment where Tamil speakers perceive over-moderation while simultaneously encountering under-moderated feeds full of hate speech.   

We used a mixed-method approach, which included an online survey of 147 frequent social media users in India and Sri Lanka; 17 in-depth interviews with content moderators, content creators, platforms’ Trust & Safety representatives, and digital rights advocates; and a roundtable discussion with Tamil machine learning and data experts. The methods are detailed in the report’s appendix.

Based on these methods, we found that: 

1. Tamil speakers use a range of Western-based social media platforms and Indian platforms. Our survey indicates that Western social media platforms are more popular among Tamil speakers, while local TikTok alternatives are gaining popularity due to India’s TikTok ban. Online, Tamil speakers use tactics to circumvent content moderation, employing “algospeak” or computer-mediated communication, and, at other times, code-mixed and transliterated Tamil using Latin script for ease and convenience. These tactics complicate moderation.

2. Tech companies pursue various approaches to moderate Tamil content online, but mostly adhere to global or localized approaches. The global approach employs the same policies for all users worldwide, and relies on moderators and policy members who are not hired based on linguistic or regional expertise. Moderators are assigned content from across the world. In contrast, the local approach tailors some policies to meet Tamil language-specific guidance, and relies on more Tamil speakers to moderate content. Some Indian companies employ a hybrid approach, often making occasional localized adjustments for Tamil speakers.

3. Tamil speakers, like others, routinely face inconsistent moderation, which they attribute to the fact that their primary language is not English. On the one hand, they encounter what they believe are under-moderated information environments, full of targeted abuse in Tamil. On the other hand, they encounter what they suspect is unfair over-moderation targeting Tamil speech in particular.

4. A majority of survey respondents are concerned about politically-motivated moderation and believe that content removals and restrictions are used to silence their voices online, particularly when they speak about politics. A few users also suspect that they experience “shadowbanning,” or a range of opaque, undisclosed moderation decisions by platforms, particularly when they use certain words or symbols commonly used by or associated with the Tamil community.

5. Despite a vibrant Tamil computing community, investment in automated moderation in Tamil still falls significantly short due to a lack of accessible resources, will, and financial constraints for smaller social media companies.

Read the full report.

The post Moderating Tamil Content on Social Media appeared first on Center for Democracy and Technology.

In today’s episode, we dive into an often overlooked but incredibly important issue: content moderation in non-dominant languages, specifically Maghrebi Arabic dialects. Recently, CDT released a groundbreaking report titled “Moderating Maghrebi Arabic Content on Social Media,” which explores the challenges, gaps, and opportunities in ensuring fair and effective moderation of Arabic content on social platforms.

To unpack these insights, we’ll be joined by two leading experts in the field: Mona Elswah, a researcher in digital rights and content moderation, and Fellow of the Content Moderation for Languages in the Global South Project at CDT; and Emna Mizouni, a Tunisian digital rights advocate, community leader, and founder-CEO of Digital Citizenship.

We’ll discuss the unique hurdles faced by Maghrebi Arabic speakers online, why it’s critical for platforms to address these gaps, and what needs to change to make content moderation more inclusive and equitable for all users.

Listen

(CDT relies on the generosity of donors like you. If you enjoyed this episode of Tech Talk, you can support it and our work at CDT by going to cdt.org/techtalk. Thank you for putting democracy and individual rights at the center of the digital revolution.)

The post Tech Talk: Talking Tech with Mona Elswah & Emna Mizouni on Moderating Maghrebi Arabic Content on Social Media appeared first on Center for Democracy and Technology.

Dr. Luria’s testimony takes an evidence-based approach to the topic, highlighting some areas of agreement, as well as gaps in research on the topic. CDT commends the efforts to address online safety for children and supports initiatives towards effective solutions. However, based on the existing evidence, our testimony highlights that:

• Youth utilize social media to connect with peers, feel accepted, and express themselves creatively — activities that are healthy and developmentally appropriate.

• The bill suggests using notifications to inform minors about the impacts of social media on their health, while current research on this topic is limited and often contradictory. This uncertainty raises concerns about the effectiveness of notifications, which may oversimplify complex issues and fail to provide meaningful guidance.

• The bill centers the measure of time spent online, but experts agree that “screen time” is an unreliable proxy of harm. Instead, researchers recommend focusing on the nature of activities; interaction with peers, for example, may impact users in a different way than content consumption.

• A key strategy for enhancing safety online, based on CDT’s own research, is to enhance young users’ safety by equipping them with more tools to navigate online risks, empowering them to protect themselves instead of taking away their agency.

• Following evidence from a recent research symposium hosted by CDT, we support the need for social media platforms to make data more accessible for research purposes. With some platforms retreating from data-sharing initiatives, there may be an opportunity to pass legislation that would enable vetted researchers to study, in a privacy protecting way, critical questions related to children’s and all users’ safety online.

Read the full testimony here.

The post CDT’s Dr. Michal Luria Testifies Before Colorado Senate Judiciary Committee on “Protections for Users of Social Media” Bill appeared first on Center for Democracy and Technology.

Executive Summary

This report summarizes the key discussions and insights from an in-person symposium held in September 2024 on the topic of children’s online safety policy. The event convened academic researchers, policy experts, and civil society representatives to explore research-driven approaches to addressing critical issues impacting young users in digital environments. During the symposium, we attempted to foster meaningful dialogue, identify areas of consensus and disagreement, and chart actionable paths forward. The symposium included a range of perspectives, and thus the report reflects a synthesis of ideas rather than unanimous agreement.

The symposium brought together 23 participants for a day-long event conducted under the Chatham House Rule. Attendees engaged in two rounds of thematic roundtables covering four key topics related to child safety on online platforms: Connection, Content, Communication, and Characteristics. The event concluded with an all-participant session that summarized some of the main discussions and identified strategies and opportunities to integrate research into policy.

We lay out some of the cross-cutting themes that we have identified in conversation; these highlight the interconnectedness of issues surrounding youth safety online, and emphasize the need for evidence-based and youth-centric approaches, particularly along the following lines: 

• No one-size-fits-all approach fixes current issues. Researchers pointed to a range of ways for keeping young people safe online, yet most solutions raise thorny tradeoffs.
• Experiences of all youth online should be examined, including those with different backgrounds. Participants repeatedly raised that young users experience online environments differently based on factors like age, socioeconomic status, and identity. Tailored safety measures, they note, may be essential to address these varied experiences effectively. Some said that additional aspects like access and digital literacy require further consideration of tools that accommodate diverse user needs.
• Consider the ecosystem of actors who are part of a young person’s life holistically. The discussions emphasized adopting a more holistic and collaborative approach to online child safety. Participants underscored the necessity of collective efforts that would involve parents, educators, platform designers, and policymakers. Collaboration across these groups was identified as crucial for reaching feasible and balanced actionable steps.
• Limited researcher access to data impedes evidence-informed solutions. Researchers in the group agreed that a lack of access to comprehensive data impedes fully understanding online harms and prevents learning about the effectiveness of existing safety measures implemented by digital platforms. Most agreed that improved access to data is vital to develop evidence-informed policy.

Participants also proposed several practical steps with potential to enhance online safety for young people on digital platforms:

• Establish default protections. Participants agreed that implementing safety settings by default, such as private accounts, can potentially keep young users and all users safer.
• Empower users with the ability to customize their online experiences. According to participants, equipping youth — and all users — with features like customizable content filters and algorithm reset options could give them the reins to shape their own experiences online.
• Provide researchers with privacy-preserving mechanisms to access data. Participants emphasized the importance of providing researchers with access to platform data, especially data related to safety mechanisms (e.g., the rate of users who use safety tools or how these tools are being used). They noted that this would allow researchers to better study online experiences and evaluate the effectiveness of safety measures.
• Support digital literacy and onboarding. Participants recommended platforms to work towards supporting users’ development of skills to navigate digital spaces responsible, as opposed to restricting access to young users altogether. Leveraging peer-to-peer education, more collaborative onboarding processes, and norm setting can all help acquaint young users with improving online norms and safety practices.

The conversation underscored the complexity of creating safer online environments and the importance of engaging researchers, who can share data-driven knowledge on approaches that have the potential to work. Participants emphasized the need for ongoing dialogue and actionable processes — safer digital spaces require sustained efforts to bridge gaps between research, policy, and platform design. This report serves as a step towards creating this shared space that would support the creation of safer digital environments for young users while respecting their rights and agency.

Read the full report.

The post The Kids are Online: Research-Driven Insights on Child Safety Policy appeared first on Center for Democracy and Technology.

As federal and state laws come into effect, we should anticipate that Americans will use two tools for censorship circumvention similar to those that have been a lifeline to Internet freedom under authoritarian regimes: proxying and content mirroring. As we’ve seen elsewhere, the technical and legal details of how site bans are implemented in the US may determine which and whether these tools are effective.

Proxying

A Virtual Private Network (VPN) is a tool that allows a user to access the internet via another location — whether down the street or halfway around the world. These systems can stand as a distancing layer between users and the content they want to access, hiding their IP addresses, location and identity. VPNs are commonly used, for example, by people accessing adult content sites from states with bans or age-verification requirements, or by those who want to view video content that is restricted to a certain location for licensing reasons. A VPN can also hide which websites a user visits from their internet service provider (ISP), an important requirement when censors impose penalties on the individuals who visit prohibited content. But the VPN provider itself is able to see which websites are being accessed, so the privacy and security protections of the VPN provider are essential to keep users protected.

A multi-hop or “onion routing” approach goes a step or two beyond a VPN by sending a user’s request for content from an app or website through more than one network or node. By forcing a website or regulator to “peel back” the (often encrypted) layers to find a user’s actual location or identity, onion routing further obscures the source of a request and can keep a user’s identity private from both the destination site and the intermediaries.

Content Mirroring

Content mirroring occurs when a provider copies banned content onto another channel that’s not being blocked.

When governments have banned popular websites like Wikipedia, users (often based in other countries) have taken the initiative to create sites that “mirror” blocked websites but that may still be accessible. For instance, if regulators block versions of the website specific to the languages their citizens use, activists may copy an English-language edition of the site, translate it and host it in an environment that’s still accessible. Governments can and will try to crack down on mirror sites, however, by blocking access when they find them. Although mirroring can be a powerful tool for avoiding censors in some situations, in others (like the TikTok ban) it may not be a practical solution at a large scale.

The concept of mirroring can be extended on a small scale: people may share content they previously accessed, like screenshots or downloads of files, via direct messaging,local networks or social media.

Other techniques — including domain fronting, encrypted DNS, and encrypting server names in large cloud providers — are used to circumvent censorship where the government mandates blocks by network providers, often to block access to content hosted on servers outside the government’s jurisdiction. These are techniques that are typically applied by the platform sharing the content, in coordination with hosting providers or client software. But where laws like the TikTok ban and adult content age verifications/bans apply directly to the US companies doing hosting or app distribution, not those at the network level, those circumvention techniques aren’t likely to be used.

The post Online Censorship Isn’t New – Neither Are Efforts to Evade It appeared first on Center for Democracy and Technology.

[ PDF version]

Dear Mr. Zuckerberg,

As a group of external experts whom Meta has convened for years to advise on how its policies impact marginalized communities, we write to express our grave concern surrounding Meta’s recently announced content moderation policy changes. Meta’s changes will permit more dangerous and abusive content on its services and undermine the quality of information available to users. Both of these changes harm the users Meta claims to serve.

We are shocked and disappointed that Meta did not consult with this group or its members in considering these significant revisions to its content policy. Meta’s failure to engage even its own advisory group of external civil rights experts shows a cynical disregard for its diverse user base and calls into question Meta’s commitment to the free speech principles to which it claims to want to “return.” 

A bad process makes bad policy. In your announcement, you touted Meta’s commitment to free expression, but the changes will significantly harm free speech.

The new Hateful Conduct policy will now allow far more anti-LGBTQ, racist, anti-immigrant, and ableist content on Meta’s services globally. For example, the policy now permits “allegations of mental illness or abnormality when based on gender or sexual orientation” in religious and political discourse. This exception applies only to LGBTQ persons, opening them up to discriminatory harassment and degradation. It also introduces into Meta’s policies the malicious trope of so-called “transgenderism,” a manufactured term used incorrectly to imply that being a trans person is an ideology. Moreover, while it is true that this policy explicitly targets LGBTQ people, LGBTQ people are not the only people affected by it. For people with disabilities, this policy turns their identity into a tool for hate and adds stigma to already highly stigmatized diagnoses and conditions. 

The policy further allows “call(s) for exclusion or use(s of) insulting language in the context of discussing political or religious topics, such as when discussing transgender rights, immigration, or homosexuality.” With the increasing attacks on and dehumanization of LGBTQ and immigrant communities, allowing for calls to exclude these communities from schools or other civic spaces puts these communities at significant risk of harm, online and off. 

These changes are devastating for free expression because they will subject members of protected groups to more attacks, harassment, and harm, driving them off Meta’s services, impoverishing conversations, eliminating points of view, and silencing dissenting and oft-censored voices. Meta has apparently entirely failed to consider the chilling effect of its policy changes—and the result will be more self-censorship, not less. 

In addition to these dangerous policy changes, Meta announced that it will eliminate fact checkers and replace them with a community notes-style system in the United States. Fact-checking is not censorship, and the result of fact-checking on Meta’s services was never censorship. It resulted (rightfully) in the contextual labeling of posts—fostering more speech and debate, not less. Characterizing fact-checking as censorship risks undermining well-sourced journalism, research, and scientific study. Meta is right to seek a trusted system, and community moderation is a useful addition to help achieve that goal. However, Meta should realize that communities are only as well-informed as the information they have access to. Processes for crowd-sourced content annotation must be built to ensure reliability and accuracy. Independent and reliable sources must be part of those processes. Otherwise, the trust Meta claims to seek to build will remain elusive. 

If Meta truly wants to recommit to free speech, it must commit to free speech for everyone on its services. As Meta’s external civil rights advisory group, we offer our advice and expertise in crafting a better path forward.

Signed,

American Association of People with Disabilities

Asian Americans Advancing Justice | AAJC

Center for Democracy & Technology

Common Cause

GLAAD

Human Rights Campaign

Lawyers Committee for Civil Rights Under Law

Muslim Public Affairs Council

NAACP Legal Defense and Educational Fund, Inc.

National Black Justice Collective

National Fair Housing Alliance

National Urban League

PDF version.

The post Letter from Meta Civil Rights Advisory Group Members on Grave Concerns with Content Policy Changes appeared first on Center for Democracy and Technology.

[ PDF version ]

Introduction

Africa, a continent with over 2,000 languages and home to more than one-third of the world’s linguistic diversity, has many languages that remain beyond the reach of both automated and human content moderation (Shiundu, 2023). Social media platforms have a limited physical presence in Africa, operating only a few offices and employing minimal staff (De Gregorio & Stremlau, 2023). Despite this, these companies have heavily invested in outsourcing content moderation labor to the continent, hiring vendors to recruit moderators to review content from both Africa and beyond. One of the few African languages benefiting from human moderation is Kiswahili, a language that is spoken by over 100 million people in East and some parts of Central Africa. In this report, we investigate how the content moderation systems of select online platforms deal with user-generated content in Kiswahili. 

This report is part of a series that examines content moderation within low-resource and indigenous languages in the Global South. Low-resource describes languages that lack sufficient high-quality training data, making it difficult to develop automated content moderation systems (Nicholas & Bhatia, 2023). In our previous research, we found that content moderation in North Africa, especially in the Maghreb region, suffered from significant biases and systemic inadequacies (Elswah, 2024). We found that content moderation systems for Maghrebi Arabic dialects are impacted by inadequate training data, which fail to capture the rich linguistic diversity of the region. Additionally, content moderators, who work under challenging conditions and are tasked with overseeing content from across the Arab world, face several challenges in making accurate decisions regarding dialects they often do not understand. This results in inaccuracies and inconsistencies in moderation practices, highlighting the urgent need for more inclusive and representative approaches to the moderation of low-resource languages in the Global South. 

This report focuses on Kiswahili (also known as Swahili), a language that exists in many varieties in East Africa in Kenya, Tanzania, Uganda, parts of the Democratic Republic of Congo, Burundi, and Rwanda, as well as in some parts of Central Africa (Topan, 2008). This report specifically concentrates on Kenya and Tanzania. We chose Tanzania because it has the largest Kiswahili-speaking population and is the birthplace of Standard Swahili. We selected Kenya because it is home to a significant number of Kiswahili speakers, ranking second only to Tanzania (Dzahene-Quarshie, 2009). Additionally, Kenya is recognized as the “Silicon Savannah” of Africa, which refers to its advanced digital transformation, rapidly increasing internet connectivity, and being host to many companies and institutions involved in the development of digital technologies (Mwaura, 2023; Wahome, 2023).   

Using a mixed-method approach that combines an online survey of 143 frequent social media users in Kiswahili and 23 in-depth interviews with content moderators, creators, and digital rights advocates from Kenya and Tanzania, we found that: 

1. According to our survey, Instagram is the most popular social media platform in Kenya and Tanzania. Additionally, TikTok’s popularity is rapidly growing in East Africa, surpassing that of Facebook. 
2. The spread of misinformation and hate speech online is a significant issue within the Kiswahili online sphere. The majority of our survey participants expressed concerns about the proliferation of misleading and inciting content on social media platforms. 
3. Popular social media platforms take three general approaches to Kiswahili content moderation: global, local, and multi-country. The global approach, exemplified by Meta, involves applying uniform policies to all Kiswahili users indiscriminately. Meta requires their Kiswahili moderators to review non-African English-language content from around the world. The local approach, employed by TikTok, tailors the enforcement of some of its policies to account for the diverse cultural contexts within East Africa. However, the variations in the Kiswahili language are overlooked because content moderation vendors hire primarily Kenyan moderators to review content from across East Africa. Many of these moderators may not be familiar with the specific contexts of other East African countries, which can lead to inadequate moderation. Lastly, the multi-country approach utilized by the local Tanzanian platform “JamiiForums” involves hiring native moderators from each Kiswahili-speaking country, who review content generated within their own nations. This ensures that the moderators understand the local context and cultural nuances, allowing them to provide more effective and relevant content moderation for users on JamiiForums.
4. Content moderation vendors often downplay the harsh realities of the job by concealing the graphic content that moderators will encounter, avoiding any mention in job advertisements, interviews, and training sessions. Many moderators misunderstand the nature of the role, with some believing they will be content “creators.” Additionally, moderators are exposed to less graphic content during the short period of training, which fails to prepare them for the often distressing content they will encounter in their daily work.
5. Much of the content moderation is conducted by third-party outsourced vendors who are contracted by social media platforms and hire moderators on behalf of the platforms. Companies in Nairobi, Kenya that provide Kiswahili content moderation services exclusively hire Kenyans to manage the diverse variations and contexts of Kiswahili content. This leads to many incidents of inaccuracies and inconsistencies in content evaluation.

Read the report.

Soma toleo la Kiswahili la ripoti hiyo.

The post Moderating Kiswahili Content on Social Media appeared first on Center for Democracy and Technology.

Please do not hesitate to contact our team in Brussels: Laura Lazaro Cabrera, Silvia Lorenzo Perez, Aimée Duprat-Macabies, David Klotsonis, and Giulia Papapietro.

Security, Surveillance & Human Rights

Civil Society Strategises on Tackling Spyware

Spyware remains high on the EU agenda; in Poland, an arrest was recently made in relation to the governmental probe on the use of Pegasus. In this context, on 20 November, CDT Europe convened the Spyware coordination group to strategise on EU-level actions to tackle spyware. The discussion focused on key areas of regulation and advocacy, aiming to build consensus. Points of convergence included the need for definitions of key terms that can adapt to the rapid evolution of spyware technologies, and the strict prohibition of spyware use against journalists.

Participants also explored the potential of internal market regulation as a legal basis for addressing the commercial spyware market and industry. Insights from the European Media Freedom Act (EMFA) and the EU Cybersecurity Framework informed discussions, particularly regarding litigation strategies to challenge Article 4 implementation and leverage cybersecurity policies to mitigate spyware threats. 

The workshop highlighted the shared urgency of curbing spyware misuse through coordinated, impactful advocacy and legal action.

All Eyes on Member States’ Actions on Spyware

At the various Pall Mall Process meetings that took place on the sidelines of the Paris Peace Forum, CDT Europe’s Silvia Lorenzo Perez engaged in critical discussions where she highlighted the urgent need for coordinated global action on spyware. 

At a panel held by the Swedish government and Access Now, victims shared powerful testimonies on the devastating impact of spyware abuse. These accounts underscored the urgent need for robust regulatory action to protect human rights defenders. This meeting was followed by a multistakeholder roundtable focused on combating the spread of commercial spyware. 

The day concluded with a Pall Mall Process meeting to review measures aimed at preventing spyware proliferation globally. While governments recognise the dangers of spyware, translating concerns into enforceable legal frameworks remains a challenge. The EU now has a unique opportunity to lead, with Member States at the table tasked with driving critical reforms. With the Polish Presidency of the Council of the EU at the helm, the time is now for bold leadership to address spyware abuse and protect both national security and individual rights.

Recommended read: The Guardian, Ronan Farrow on surveillance spyware: ‘It threatens democracy and freedom”

  Online Expression & Civic Space

Trusted Flaggers in the DSA: Challenges and Opportunities

Implementation of the Digital Services Act (DSA) is at a busy phase, with online platforms starting to release their first annual risk assessment and audit reports (CDT Europe and other CSOs published a joint letter on the process). Another crucial part of the regulation’s implementation rests with the Trusted Flagger Mechanism, which helps combat illegal content online by granting certified entities priority processing of flagged material. CDT Europe and EU DisinfoLab organised a webinar on the topic on 21 November, where over 30 participants, including civil society organisations (CSOs), Digital Services Coordinators, and the European Commission, explored current challenges and opportunities. The system faces significant hurdles, including resource constraints for CSOs applying for certification, misinformation campaigns undermining public trust in Trusted Flaggers, and low uptake due to complex, burdensome processes and unclear benefits. With only 15 certifications granted so far, the mechanism is underutilised. 

Some key recommendations from the event include:

• Ensuring sustainable funding for CSOs to meet Trusted Flagger obligations’
• Developing proactive communication strategies to counter misinformation and clarify the role of Trusted Flaggers to the wider public; and
• Establishing a working group to harmonise practices, support applicants, and address challenges like application complexity.

In our full outcomes report blog, we identify key opportunities for CSOs.

A Human Rights-Centered Application of the DSA

CDT Europe’s Research and Policy Officer David Klotsonis joined a workshop in Vienna, organised by the DSA Human Rights Alliance and hosted by the Organisation for Security and Co-operation in Europe (OSCE). The event focused on exploring principles for a Global Human Rights-Centered application of the Digital Services Act. The participants discussed lessons from other jurisdictions and conflict zones to shape thoughtful DSA implementation, while considering the risks of applying the law to different regulatory environments without accounting for unique vulnerabilities. As the “Brussels Effect” continues to generate buzz, it’s crucial to unpack its real-world implications. How can laws, when removed from their original institutional context, unintentionally—or deliberately—undermine human rights? This workshop offered a timely platform for reflection, and was a source of important insights.

Online Gender-Based Violence: What Now?

Online gender-based violence (OGBV) continues to be a widespread and alarming issue, fueled by misogynistic narratives, that affects women in Europe and around the world. On the International Day for the Elimination of Violence against Women. and in the context of the 16 Days of Activism against Gender-Based Violence, CDT Europe highlighted the EU’s progress on the issue, such as the Directive on combating violence against women and the Digital Services Act. Despite these advancements, problems persist in ensuring the online space is free of this gendered harm. In our blog, we explored the obstacles ahead, emphasising the need for cultural change and effective implementation. 

Recommended read: The Verge, Meta says it’s mistakenly moderating too much 

Equity and Data

An Ongoing Battle for Full Accountability for AI Harms

In our latest blog post, we reflected on persistent gaps in EU regulation that hinder accountability for AI-induced harms. Transparency, an inherent challenge for AI systems, is a crucial prerequisite to identifying harms. The AI Act goes some way towards ensuring a base level of transparency in some circumstances, but neglects the importance of procedural safeguards to ensure individuals’ legal access to remedies. This was never the AI Act’s intention, as it was conceptualised around the same time as the AI Liability Directive (AILD), a proposal that outlined basic steps towards easing procedural burdens for complainants in recognition of the hurdles posed by AI’s opaque functioning. Despite the AILD’s process-oriented nature and modest impositions, the draft law is struggling to get off the ground — even as the effective remedies issue in AI remains unaddressed. 

Making the Case for Robust European Regulation

In a debate hosted by Euronews as part of their Tech Summit on 4 December, CDT Europe’s Laura Lazaro Cabrera shared the stage with representatives from DG JUST and CEPS to discuss regulation for consumer protection in the digital age. In the discussion, Laura highlighted the importance of ensuring laws regulating tech include both substantive and procedural safeguards to truly guarantee robust consumer protection. She also noted the importance of challenging the false dichotomy between innovation and regulation, underscoring the value of high product standards and their essential role in preserving health, safety, and fundamental rights. She also questioned the false assumption that underperforming products falling short of robust standards would lead to Europeans missing out — rather, it’s companies that would be missing out on the European market should they fail to find ways to conform. 

Recommended read: The Guardian, Deus in machina: Swiss church installs AI-powered Jesus

Bluesky

We are on Bluesky! As more users join the platform (including tech policy thought leaders), we’re finding more exciting content, and we want you to be part of the conversation. Be sure to follow us at @cdteu.bsky.social! You can also follow our starter pack of EU tech journalists, to catch the latest digital news in the bubble. Find us also on Mastodon and LinkedIn.

Upcoming Events 

Liberal Forum Roundtable: On 10 December, our Equity and Data Programme Director Laura Lazaro Cabrera will participate to the the European Liberal Forum’s conference on “The Era of AI: Harnessing AI for Humanity”, bringing together MEPs, APAs, political advisors, civil society, academia, and corporate sector representatives to engage in Chatham House discussions on the role of the EU in advancing AI over the next mandate. 

Kofi Annan Foundation: On 11 December, Laura will speak at the “Comparative lessons from the EU and the US elections in the age of Artificial Intelligence” event organised by Democracy Reporting International (DRI) and the Kofi Annan Foundation (KAF) to reflect upon the risks and challenges generative AI represents for European democracy.

The post EU Tech Policy Brief: December 2024 appeared first on Center for Democracy and Technology.

The goal? To foster understanding and support knowledge-building for CSOs with the aim to ensure that the trusted flagger system remains transparent, effective, and inclusive. 

The Trusted Flagger Framework: A Brief Overview

The DSA represents a groundbreaking shift in how the EU regulates digital platforms. It imposes asymmetric due diligence obligations on online platforms depending on their size and impact. Article 22 focuses on so-called ‘trusted flaggers’ — entities equipped with specialised expertise in flagging illegal content. These flaggers benefit from priority processing of their reports, a feature designed to enhance the removal of illegal content, while safeguarding freedom of expression.

In what follows, we have summarised the key insights from the webinar.

Challenges in Implementation

• Resource Constraints: Civil society organisations (CSOs), often the main applicants for trusted flagger status, face significant resource limitations. The process of applying for certification is time-consuming and requires extensive documentation to demonstrate independence, diligence, and accuracy. Specific measures for improvement were identified during the webinar.
• Misinformation and Public Perception: Trusted flaggers have been the target of disinformation campaigns. Some narratives portray them as “censorship machines,” stirring public skepticism about their role. Clear communication about their purpose — strictly targeting illegal content — remains essential to counter these misconceptions.
• Low Uptake: With only 15 trusted flaggers certified so far, the system is still underutilised. Potential applicants cited the burdensome certification requirements, lack of clarity on a process that is still being developed and uncertainty about the tangible benefits of trusted flagger status as significant barriers to participation. Overcoming these challenges will be key to unlocking the full potential of the trusted flagger mechanism.

Opportunities Ahead

Despite these challenges, the DSA’s trusted flagger framework offers numerous opportunities that could strengthen the fight against illegal online content across the EU:

• Institutionalisation of Relationships: One of the key benefits of the DSA’s approach to trusted flaggers is the (legally mandated) formalisation of relationships between CSOs and online platforms. This structured interaction ensures greater consistency and reliability in content moderation practices.
• Enhanced Efficiency: The priority processing of reports by trusted flaggers allows for faster action on illegal content. This is especially critical in cases like image-based abuse and online gender based violence, where rapid intervention is necessary to mitigate harm.
• Cross-Border Reach: Trusted flagger certification in one EU member state enables organisations to operate across all EU platforms. This wide-reaching capability increases their impact and strengthens the effectiveness of content moderation efforts.
• Recognition of Expertise: Trusted flagger certification serves as recognition of the organisation’s expertise and competence in addressing specific forms of online harm. This validation can enhance the organisation’s credibility and visibility in their field.
• Improved Transparency: Platforms often provide more detailed feedback to trusted flaggers regarding flagged content and the actions taken. This transparency enables organisations to better understand platform policies and the impact of their flagging efforts, fostering a more productive relationship.
• Potential for Broader Impact: As trusted flaggers, organisations can contribute to shaping better practices in content moderation by sharing insights and trends observed through their work. This collaborative approach can lead to systemic improvements in handling online harm.

Key Recommendations 

1. The Need for Sustainable Funding
   Participants consistently emphasised the importance of adequate funding to enable CSOs to fulfill the responsibilities of trusted flagger status. While some EU grants support specific sectors, broader funding mechanisms are needed. Without financial sustainability, many organisations cannot afford the infrastructure required for compliance or the human resources to handle the increased workload.
   
2. Ensuring Adequate Communication
   Misinformation campaigns against trusted flaggers pose a real threat to their credibility. Developing a proactive communication strategy that clearly defines their role in addressing illegal — not inconvenient — content is essential. Drawing inspiration from the experiences of fact-checking organisations could help trusted flaggers navigate these challenges effectively.
   
3. Creating a Working Group for Coordination
   A recurring theme in the seminar was the need for a formal working group to facilitate collaboration and knowledge sharing among DSCs, trusted flaggers, and aspiring applicants. Such a group could harmonise practices across member states, provide guidance to applicants, and address systemic issues like application complexity and disinformation campaigns. While civil society can help harmonise practices and bridge gaps, however, progress depends on regulators. The European Commission made an important announcement on upcoming Article 22 guidelines, but DSCs must also clarify application processes and make sure that all relevant information is easily-accessible.

The seminar on Article 22 of the DSA underscored the potential of the trusted flagger system to improve online safety across the EU. However, realising this potential requires addressing significant structural challenges. By investing in sustainable funding, building convincing and informative public narratives, and fostering cross-stakeholder collaboration, the trusted flagger mechanism can become a cornerstone of effective, rights-respecting content moderation with possible implication also beyond the EU.

As this system evolves, the collective efforts of CSOs, regulators, and platforms will be pivotal in ensuring its success. With continued structured dialogue and action the trusted flagger model holds promise for a safer and more accountable digital ecosystem.

The post Trusted Flaggers in the DSA: Challenges and Opportunities appeared first on Center for Democracy and Technology.

Many leading language models are trained on nearly a thousand times more English text compared to text in other languages. These disparities in large language models have real-world impacts, especially for racialized and marginalized communities. For example, they have resulted in inaccurate medical advice in Hindi, led to wrongful arrest because of mistranslations in Arabic, and have been accused of fueling ethnic cleansing in Ethiopia due to poor moderation of speech that incites violence.

These harms reflect the English-centric nature of natural language processing (NLP) tools, which prominent tech companies often develop without centering or even involving non-English-speaking communities. In response, region- and language-specific research groups, such as Masakhane and AmericasNLP, have emerged to counter English-centric NLP by empowering their communities to both contribute to and benefit from NLP tools developed in their languages. Based on our research and conversations with these collectives, we outline promising practices that companies and research groups can adopt to broaden community participation in multilingual AI development.

Read the full brief.

The post Beyond English-Centric AI: Lessons on Community Participation from Non-English NLP Groups appeared first on Center for Democracy and Technology.