Trading nuance for scale? Platform observability and content governance under the DSA

Papaevangelou, Charis; Votta, Fabio Article Trading nuance for scale? Platform observability and content governance under the DSA Internet Policy Review Provided in Cooperation with: Alexander von Humboldt Institute for Internet and Society (HIIG), Berlin Suggested Citation: Papaevangelou, Charis; Votta, Fabio (2025) : Trading nuance for scale? Platform observability and content governance under the DSA, Internet Policy Review, ISSN 2197-6775, Alexander von Humboldt Institute for Internet and Society, Berlin, Vol. 14, Iss. 3, pp. 1-31, https://doi.org/10.14763/2025.3.2037 This Version is available at: https://hdl.handle.net/10419/330355 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/3.0/de/deed.en Volume 14 | Trading nuance for scale? Platform observability and content governance under the DSA Charis Papaevangelou University of Amsterdam c.papae[email protected] Fabio Votta University of Amsterdam f.a.[email protected] DOI: https://doi.org/10.14763/2025.3.2037 Published: 17 September 2025 Received: 11 November 2024 Accepted: 20 June 2025 Funding: This work was supported by the Dutch Ministry of Education, Culture and Science under Grant 024.005.017 (Gravitation Research Programme “Public Values in the Algorithmic Society”). Competing Interests: The author has declared that no competing interests exist that have influenced the text. Licence: This is an open-access article distributed under the terms of the Creative Commons Attribution 3.0 License (Germany) which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. https://creativecommons.org/licenses/by/3.0/de/deed.en Copyright remains with the author(s). Citation: Papaevangelou, C., & Votta, F. (2025). Trading nuance for scale? Platform observability and content governance under the DSA. Internet Policy Review, 14(3). https://doi.org/10.14763/2025.3.2037 Keywords: Digital Services Act (DSA), Platform governance, Platform observability, Transparency, Content moderation Abstract: The Digital Services Act (DSA) marks a paradigmatic shift in platform governance, introducing mechanisms like the Statement of Reasons (SoRs) database to foster transparency and observability of platforms’ content moderation practices. This study investigates the DSA Transparency Database as a regulatory mechanism for enabling observability, focusing on the automation and territorial application of content moderation across the EU/EEA. By analysing 439 million SoRs from eight Very Large Online Platforms (VLOPs), we find that the vast majority of content moderation decisions are enforced automatically and uniformly across the EU/EEA. We also identify significant discrepancies in content moderation strategies across VLOPs, with TikTok, YouTube and X exhibiting the most distinct practices, which are further analysed in the paper. Our findings reveal a strong correlation between automation and the speed of content moderation, automation and the territorial scope of decisions. We also highlight several limitations of the database, notably the lack of language-specific data and inconsistencies in how SoRs are reported by VLOPs. We conclude that despite such shortcomings, the DSA and its Transparency Database may enable a wider constellation of stakeholders to participate in platform governance, paving the way for more meaningful platform observability. Issue 3 Introduction The Digital Services Act (DSA; Regulation (EU) 2022/2065, 2022) has been described as a “transparency machine” (Zornetta, 2024) with the potential of creating a “global transparency regime” (Helberger & Samuelson, 2024). The DSA, among others, requires platforms1to provide a justification for their content decisions (e.g., grounds for content removal or restriction) in the form of a Statement of Reasons (SoRs; DSA Art. 17(1)) to shed light on the often-opaque decision-making processes of their content moderation systems. These SoRs are then automatically uploaded and aggregated to a publicly accessible online database operated by the European Commission (DSA Art. 24(5)) putting in place a system of “automated transparency” (Kaushal et al., 2024). The early impact of the DSA is already apparent. For example, researchers have uncovered inconsistencies in the self-reported practices of X, which claims to rely exclusively on manual moderation in its SoRs submissions but acknowledges using automated tools in its transparency reports (Dergacheva et al., 2023; Drolsbach & Pröllochs, 2023). Such cases highlight the latent regulatory potential of the DSA and demonstrate how transparency can enable platform observability (Rieder & Hofmann, 2020), a dynamic and analytical governance mechanism that facilitates a deeper understanding of platforms’ governance practices (Leerssen, 2024). In this paper, we approach the Transparency Database as an instrument for platform observability, which builds upon transparency, to conduct an empirical case study on the kinds of insights that can be inferred from the Transparency Database. We seek to understand the extent to which the SoR database fosters observability, which we recognise as a governance mechanism that is necessary, in addition to transparency, to ensure platform accountability and deepen our understanding of how platforms operate and exercise their power over our digital public spheres (van Dijck et al., 2018). Our study, specifically, focuses on exploring the use of automation in content moderation and the differences in the territorial scope of content moderation practices, as well as the implications thereof for the DSA and platform governance. We conducted our analysis with two leading questions in mind: how do content governance decisions vary among platforms and member-states in the EU, and in what ways does the use of automation in moderation differ across platforms, the 1. The DSA covers all intermediary platforms, including e-commerce and marketplaces (e.g., Zalando, App Store), ride-sharing apps (e.g., Uber), and others. 2 Internet Policy Review 14(3) | 2025 EU and its member-states. We concentrated our analysis on the moderation practices of eight digital Very Large Online Platforms (VLOPs), namely X, Facebook, Instagram, YouTube, TikTok, Snapchat, Pinterest and LinkedIn, most of which have been foundational for our modern digital public spheres and the broader digital platform ecosystem. We used R, a programming language that is widely used for data analysis (R Core Team, 2021), to retrieve 439 million SoRs from a period of four months (25 September 2023 to 25 January 2024). We conclude that the DSA enables a variety of actors (Helberger et al., 2018), including researchers and members of civil society, to make use of technically-sophisticated regulatory mechanisms, such as the Transparency Database, supplementing the more traditional governance mechanism of transparency. In that sense, the Transparency Database, despite its shortcomings, marks a crucial turning point in governing digital platforms as it paves the way for observing platforms’ behaviour and content moderation practices in a way that was unattainable before. Our paper contributes to a growing body of literature on platform governance in regulated digital environments, especially in the EU. In doing so, it also offers a critical perspective on the DSA’s regulatory and governance ambitions. In summary, we discerned three key findings. First, we found that 99% of all moderation decisions were applied uniformly across the EU/EEA, exhibiting a tendency for uniform application of content moderation strategies. Second, we identified variations in how VLOPs enforce content moderation, particularly in terms of regional differences, with only three, namely TikTok, Youtube and X, reporting conducting territorially specific content moderation. Third, we observed that the use of automation in content moderation correlates with faster enforcement timelines and uniform application, whereas manual moderation often means longer delays and more territorially specific application. This last finding also raises questions about the interaction of EU and national legal frameworks and the VLOPs’ capacity to handle “low-resource languages” (Nicholas & Bhatia, 2023). Last, we consider some technical shortcomings of the database and regulatory blindspots of the DSA such as the lack of an obligation to report the language of the content that was moderated. The next sections are structured as follows: first we situate our paper theoretically through a literature review of works relevant to content moderation, platform governance and regulation in order to build our case for platform observability; second, we expound on our methodological approach; third, we present our findings in detail; finally, we discuss the implications of our findings for the DSA and platform governance. 3 Papaevangelou, Votta Conceptual framework: Content moderation, platform governance and the need for platform observability Our conceptual framework is developed in three steps: first, we do a brief literature review of scholarship on content moderation as an instrumental part of platform governance and, subsequently, platforms’ power; second, continuing our literature review, we trace the rise of automation in content moderation, along with its structural limitations; and third, we advance the need to adopt platform observability as a critical lens through which to study platform governance and as a fitting regulatory approach, as we later show through our case study on the DSA. Content moderation and platform governance In recent years, a rich body of research from critical media to legal scholarship has extensively explored the politics and mechanisms of content moderation and its implications for democracy. Without attempting an exhaustive discussion of the literature, we draw upon scholarship on platform governance which has illuminated how platforms navigate the balance between self-regulation and state oversight, highlighting the informal and formal mechanisms that underpin platform governance (Gorwa, 2019b, 2019a; Papaevangelou, 2021). Gorwa has also, crucially, underscored the role of power dynamics between platform governance stakeholders in influencing content moderation, taking stock of the political reality within which platform governance is inscribed (Gorwa, 2024). In a similar vein, Griffin (2023) has theorised about the politics behind content moderation decisions, demonstrating how platforms balance legal, public and commercial interests in their governance strategies. This strand of scholarship has, therefore, showcased the inherently political process of moderating content that is often obscured by platforms’ sophisticatedly opaque systems, which include extractive processes of invisible labour (Roberts, 2019) and which are veiled behind discourses of neutrality (Gillespie, 2010). Subsequently, content moderation is closely tied to the platforms’ political power, that is, the power to shape the norms, rules and conditions under which information circulates online and, by extension, influence the structures of our digital public spheres and the capacity of citizens to form political opinions (Gillespie, 2018; Helberger, 2020). Content moderation, thus, emerges not simply as a technical 4 Internet Policy Review 14(3) | 2025 function but a central mechanism of platform firms (Grimmelmann, 2017), governing the socioeconomic, cultural and political interactions of end-users, complementors and other relevant actors that convene in the multi-sided markets that large platforms constitute (see Poell et al., 2021 Chapter 4). From this perspective, social media platforms rely heavily on content moderation to maintain their advertising-driven business models. The ability to regulate, filter and organise user-generated content enables platforms to create an environment conducive to their revenue goals, whereby the goal is to maximise profits through the increase of user engagement, while minimising harmful or controversial content that could alienate end-users or advertisers (Griffin, 2022; Jimenez-Duran, 2022). Automation in content moderation and its structural limitations Artificial Intelligence (AI) and automated decision-making processes (Bloch-Wehba, 2020; Gillespie, 2020; Gorwa et al., 2020) have been crucial in dealing with the massive volume of user-generated content uploaded every instant on social media platforms, while reducing the cost of content moderation (e.g., instead of hiring human reviewers). Typically, AI refers to systems that use machine learning models to identify, classify, or predict patterns in data. Combined with automation and automated-decision making, which typically refer to programmed processes that operate without human intervention (at least not necessarily visible labour), these sociotechnical systems have given rise to the model of “algorithmic commercial content moderation” (Gorwa et al., 2020, p. 3). The latter refers to “systems that classify user-generated content based on either matching or prediction, leading to a decision and governance outcome” (Gorwa et al., 2020, p. 3). The level of automation varies depending on the type of content it is deployed against and the specific legal framework in place. For instance, the identification of terrorist or copyrighted pieces of content is a predominantly automated process based on hash-matching, whereas toxic or hateful content might involve a combination of automation and human reviewing (Gorwa et al., 2020, p. 7). Algorithmic moderation is also veiled with a neutrality discourse, obscuring the aforementioned political-economic dimension of platform governance and emerges as “a shared imaginary of technological solutionism” (Udupa et al., 2023, p. 1). Automation and technology, in this context, are presented as providers of efficient, accurate and swift solutions to issues relevant to our democracy (cf. Poell et al., 2021 Chapter 7). As a result, relevant scholarship has noted an increasing reliance on and embracing of automation to “sanitise” online spaces (Griffin, 2022), often at 5 Papaevangelou, Votta the expense of benign, marginalised, or non-conforming discourses (Are, 2022). In that sense, automation in moderation, and the discursive embrace thereof, not only obfuscates its shortcomings but, more broadly, seeks to depoliticise platform governance and blur the messy reality of its political economy (Gorwa et al., 2020, p. 13). Moreover, this dynamic tends to obscure the critical-yet-undervalued human labour activities like data annotation and content reviewing, which are essential for training the models undergirding the automated processes of content moderation and, largely, for the political economy of content moderation (Posada, 2022; Roberts, 2019). Such labour, often outsourced in the majority world, is at the core of platforms’ content moderation industrial approach (Caplan, 2018), that is a massive automated, systematised and standardised system of moderation designed for scalability and cross-jurisdictional application. Last, the inadequacy of automation in content moderation becomes even more pronounced when engaging with “lowresource languages”, that is, languages that have a scarce digital footprint and, thus, have not been incorporated in the training of these models (Nicholas & Bhatia, 2023) The limits of transparency and the turn to platform observability These structural problems and limitations of industrial automated content moderation highlight the inadequacy of current governance mechanisms that are rooted in techno-solutionism. It is no surprise, then, that the response, primarily by platforms but also some policymakers and experts, was to ask for (more) transparency to unravel the “black box” (Pasquale, 2016). Indeed, faced with public outcry, major social media platforms put various transparency mechanisms in place, ranging from self-reported content moderation transparency reports to establishing specialised teams–the industry term is ‘Trust and Safety’–dealing with content moderation (Gorwa & Ash, 2019). These developments led platforms to “networked platform governance” structures that involve a combination of automated means, dedicated workforce and the implication of external stakeholders to govern their digital spaces and, purportedly, foster trust in involved stakeholders (Caplan, 2023). In fact, in this novel configuration, companies over-emphasised the role of transparency and of “openness and access […] to establish trustworthiness to external actors (Caplan, 2023, p. 3462). However, as critical platform scholars have noted, transparency can only go so far as a platform governance mechanism due to the inherent sociotechnical complexities of digital platforms (Ananny & Crawford, 2016), pushing for the adoption of 6 Internet Policy Review 14(3) | 2025 platform observability as a more suitable governance mechanism (Rieder & Hofmann, 2020; Leerssen, 2024). Indeed, transparency is often considered and critiqued as a static (Leerssen, 2024) form of “visibility management”, that is a highly curated and mediated process of disclosure of data, information and insights (Flyverbom, 2019, p. 18) by platforms to external stakeholders. But the underlying premise of transparency as a prerequisite component to “create the knowledge required to govern and hold systems accountable” (Ananny & Crawford, 2016, p. 975) holds. Hence, scholars studying platform governance do not argue for abandoning transparency2but rather complementing it with observability and going even further. Drawing on recent works on platform observability (Rieder & Hofmann, 2020; Ferrari et al., 2023, Leerssen, 2024), we conceptualise it as a continuous and processoriented approach that acknowledges the complexities of observing and understanding digital platforms. Unlike transparency, observability is an active, pragmatic and explicitly subjective practice, decentering our focus on “the algorithm” to encompass more facets of the processes that constitute platform infrastructures and their automated-decision making systems (Leerssen, 2024, pp. 7-8). Rieder and Hofmann, specifically, set three principles as foundational for observability: observability in relation to, or in favour of the public interest (normative principle); observability as a continuous and dynamic process (sociotechnical principle); observability as a catalyst for collaborative forms of platform governance (governance principle). Additionally, the authors connect the concept directly to regulation, arguing for a dual approach involving “regulating for observability” and making observability part of our regulatory strategies for platform firms (Rieder and Hofmann, 2020, p. 22). We further understand observability as an attempt to reverse neoliberal framings of transparency as objective, which sought to depoliticise the notion and disentangle it from its political-economic, social and cultural power structures (Birchall, 2021). Therefore, we conceptualise observability not simply as constrained to the disclosure of information but as the regulatory capacity to continuously monitor and scrutinise platform operations, while considering the emergent nature of platform behavior, influenced by user interactions, sociopolitical and economic strategies of platform owners and platforms’ automated-decision making algorithms. Consequently, observability does not simply enable a technical understanding of platform technologies but sets in place the architecture for a network of actors (Ca2. For a comprehensive comparison of transparency and observability, including of their metaphors, see Leerssen, 2024. 7 Papaevangelou, Votta plan, 2023), beyond traditional and often-opaque regulatory processes like auditing (Terzis et al., 2024) to actively and cooperatively participate in digital governance (Helberger et al., 2018; Keller & Leerssen, 2020). In that sense, observability complements and goes further than transparency to empower governance actors to act in an informed manner, reduce information asymmetries and increase trust among stakeholders and, generally, improve the conditions necessary for accountability (Birchall, 2021, p. 6; Gorwa et al., 2020, p. 11). Taking the sociotechnical aspect of platform observability “as a regulatory program” (Leerssen, 2024, p. 9) a step further, we hold that it should also include more opaque activities that exist in platforms’ value chains and which enable their operations, ranging from labour relations (e.g., real-time information about human moderators employed or contracted) to environmental implications (e.g., real-time insights into computing power used along with its environmental footprint). If we are to leave the “snapshot logic” (Rieder & Hofmann, 2020, p. 7) behind for a more comprehensive and holistic regulatory framework, then we must consider the labyrinthine, interrelated complexities of platforms’ political economy. Building on Rieder and Hofmann’s conceptual foundation, Ferrari and colleagues (2023) propose a complementary framework for the governance of generative AI systems, which introduces three “oversight conditions” (p. 2) for observability: industrial observability (the capacity to scrutinise the interlinked, material and political-economic layers of generative AI models and systems from a macro-level), public inspectability (the need for deep regulatory inspection to all layers constituting such systems) and technical modifiability (the capacity to make fundamental changes to these systems to ensure compliance with oversight mechanisms). The similarities are evident in the prominence given by both frameworks to the public interest and the role of public regulation, but what stands out is the “technical modifiability” condition as it seems to indicate and advocate for further action–beyond regulatory–to be taken as a response to the insights we may derive from observing platform and AI systems. Arguably, this last condition shows that observability too must be further expanded when dealing with tech companies and their complex value chains. The Digital Services Act and regulated platform observability The EU’s DSA (Regulation (EU) 2022/2065, 2022), voted in 2022 and enacted in full force in 2024, imposes due diligence obligations on platforms to ensure a safe 8 Internet Policy Review 14(3) | 2025 rate on our findings relating to these two central dimensions of platform governance: the territorial scope of content moderation decisions and the automation of detection and enforcement processes. Territorial aspect of content moderation One of the most striking insights from our analysis is the territorial distribution of SoRs across the European Economic Area (EEA), EU and individual member-states. The top two territorial scopes concern the EEA (with and without Iceland and Norway; Table 2), accounting for 50.40% (or 221M SoRs) and 48.41% (or 213M SoRs) of all SoRs, while the third one concerns the EU, accounting for 0.32% of our corpus (or 1M of SoRs). Put simply, almost all moderation decisions (99.13% or 435M out of the 439M SoRs of our corpus) made by VLOPs in our sample were applied uniformly across the entire EEA and EU. As regards specific member states, we found that Germany leads in terms of country-specific SoRs, with 588K SoRs (0.13%), followed by France (432K, 0.10%) and Italy (323K, 0.07%); this is to be expected given that these countries host the largest populations in the EU. To infer better countryand language-specific insights, we decided to eliminate platforms whose content moderation decisions were applied uniformly across the EU or EEA. This approach narrowed our focus to the remaining 0.87% (or approximately 4M SoRs), where territorial variation was evident. To illustrate, Facebook’s (in blue colour; Figure 1) content moderation decisions are almost exclusively applied on an EU and EEA level, providing no meaningful divergence for analysis at the member-state level. 15 Papaevangelou, Votta TABLE 2: Top 10 territorial scope of content moderation decisions in the EU/EEA RANK TERRITORIAL SCOPE SORS % SORS 1 EEA (NO ICELAND)5 221M 50.40% 2 EEA 213M 48.41% 3 EUROPEAN UNION 1M 0.32% 4 GERMANY 588K 0.13% 5 FRANCE 432K 0.10% 6 ITALY 323K 0.07% 7 EEA (NO ICELAND OR NORWAY) 197K 0.04% 8 POLAND 185K 0.04% 9 SPAIN 169K 0.04% 10 FINLAND, HUNGARY, LIECHTENSTEIN, LITHUANIA, NORWAY, POLAND, SLOVENIA 129K 0.03% Therefore, we focused on YouTube, X and TikTok, as they were the only platforms whose data showed some variation in content moderation decisions across different member-states (Table 4). TikTok applied most of its content moderation decisions uniformly across the EEA level. Simultaneously, the platform dominated our corpus showcasing the dominance of the video-sharing platform in the EU market (235M SoRs, 53,54%). As such, TikTok demonstrates a preference for an EEA-wide content moderation, with minimal variation between individual countries. When it engages in country-specific moderation (e.g., in the Netherlands or France), the volume is significantly lower than its regional enforcement, which is in line with a more industrialised and streamlined content moderation strategy. YouTube’s data indicates a more nuanced approach to territorial content moderation than TikTok, which is also demonstrated by the geographic dispersion of its decisions across areas in our scatterplot (Figure 1). YouTube is, thus, more likely to apply a territorialised approach to moderating content in individual member-states, the Netherlands, France and Italy in particular, reflecting a greater variety in its overall strategy, in spite of applying most of its decisions on an EEA level. Lastly, X reports to be 5. THE TERRITORIAL SCOPES IN TABLE 2 REFLECT SELF-REPORTED LABELS IN THE DSA TRANSPARENCY DATABASE. DUE TO INCONSISTENT LABELLING ACROSS PLATFORMS AND LACK OF PROOF-CHECKING MECHANISMS BY THE COMMISSION, OVERLAPPING CATEGORIES SUCH AS ‘EEA’ AND ‘EEA (NO ICELAND)’ MAY REFER TO THE SAME GEOGRAPHIC SCOPES, AFFECTING THE ACCURACY OF DISTINCT TERRITORIAL CATEGORIES’ NUMBERS. 16 Internet Policy Review 14(3) | 2025 implementing its content moderation decisions in a highly territorialised manner, with the Netherlands ranking again the highest among individual member-states. Automation in content moderation We then proceeded in visualising the data in two scatterplots to represent the degree of automation and its impact on the speed of enforcement of content moderation decisions across different territories (Figures 1 & 2). For clarity, in the first scatterplot (Figure 1), the horizontal axis represents the degree of automated detection, ranging from manual detection (on the left) to fully automated detection (on the right), while the vertical axis represents the volume of SoRs per 1,000 AMARs, with the sizes of the circles representing the volume of SoRs in a region. In the second scatterplot (Figure 2), the horizontal axis represents the same as the first scatterplot, while the vertical axis represents the median day of the application of content moderation decisions. Figure 1 - Statement of Reasons concerning automatically or manually detected content per 1,000 Monthly Active Users 17 Papaevangelou, Votta Figure 2 - Scatterplot showing the relation between time and automation of detection with the territorial scope of content moderation decisions for TikTok (left-hand side) and YouTube (right-hand side) Our findings here also reveal significant variations. TikTok shows a high reliance on automated detection, particularly when moderating content on an EEA-level but also in some countries like Spain and the Netherlands (Figure 1). Notably, as the case of France shows in both scatterplots, when content is detected manually it is also more likely to be dealt manually, taking significant time for a decision to be applied. For example, the median day of enforcement for TikTok in France in our sample was more than 30 days (Figure 2). Put simply, the shorter enforcement delays in these countries suggest that TikTok uses automation to address most of the content. YouTube seems to employ more of a hybrid approach, predominantly relying on automated detection methods supplemented with manual review. Generally, the same pattern as TikTok was also observed in this case, with manual detection demonstrating a correlation with manual enforcement. What is more, with a closer look in our data, we discerned that YouTube’s median reaction time, including automatically detected content, can extend in occasions over 100 days (Figure 2), with delays increasing dramatically when manual review is involved and when decisions concern specific territorial scope (e.g., France). 18 Internet Policy Review 14(3) | 2025 In contrast, X stands out as an outlier, reportedly relying exclusively on manual moderation. The reason why X’s data was not visualised as a time scatterplot in our analysis is that the platform indicates moderating all content manually and on the same day as detected (Figure 1). As mentioned earlier, this observation contradicts its public transparency reports which mention the use of automated means6, making X’s data unreliable. This discrepancy has been also corroborated by other studies (Dergacheva et al., 2023; Drolsbach & Pröllochs, 2023; Kaushal et al., 2024). Finally, our tables (3 & 6) containing data on human reviewers indicate that languages such as French and Dutch receive more attention from human moderators across platforms, especially for YouTube and TikTok. YouTube, for instance, allocates 176 French-speaking moderators and 24 Dutch-speaking moderators, while TikTok allocates 687 French-speaking moderators and 167 Dutch-speaking moderators, the latter being notably high compared to other platforms. Again, given that these are self-reported figures using non-standardised methodologies, any crossplatform comparison remains inherently limited. This could suggest that better language coverage allows platforms to manually moderate more content and, thus, provide–arguably–more accurate decisions that, as shown in Figure 2, take more time than automated decisions. Having said that, the number of human moderators is not correlated to the use of automation in content moderation. As Klonick’s (2018) analysis of Facebook’s content moderation system has shown, these platforms have different tiers to address content moderation issues according to their perceived severity and importance (see also Caplan & Gillespie, 2020). In other words, it might very well be that TikTok still predominantly relies on automated means to moderate content but employs–the most–human reviewers to deal with more sophisticated issues (Table 3). Therefore, a key limitation in understanding the operational capacities of platforms across EU member-states lies in the heterogeneity of reporting metrics regarding human moderation staff. 6. Available here: https://transparency.x.com/dsa-transparency-report.html. 19 Papaevangelou, Votta TABLE 3: Human moderators employed by platforms based on their transparency reports PLATFORM PERIOD AMARS MODERATORS AMARS/MODERATOR FACEBOOK 01/04/2023-30/09/2023 259.000.000 1.362 190.161 INSTAGRAM 01/04/2023-30/09/2023 259.000.000 1.362 190.161 YOUTUBE 01/01/2023-30/06/2023 416.600.000 1.974 211.043 LINKEDIN 01/01/2023-30/06/2023 45.200.000 146 309.589 SNAPCHAT 01/02/2023-30/07/2023 101.973.520 1.545 66.002 PINTEREST 01/03/2023-30/06/2023 124.000.000 1.963 63.168 X (TWITTER) 01/04/2023-30/10/2023 126.120.951 2.496 50.529 TIKTOK 01/04/2023-30/09/2023 125.000.000 5.827 21.451 Discussion Content and platform governance in the EU Our analysis showcases how the DSA and, the Transparency Database in particular, is a step toward fostering a dynamic way of studying platform governance and, thus, in enabling platform observability. Our findings foreground the inherent tension in the pursuit of faster and more accurate content moderation decisions, especially via the use of automation. They also demonstrate how calls for faster reaction times inherently invite more automation and, thus, an elevated risk to ignore vital context in content moderation, in addition to the embedded limitations of automation. While most content moderation decisions in our sample were enforced uniformly across the EEA and EU–accounting for 99.13% of all SoRs–our findings also point to a territorialised aspect of platform moderation, particularly in member states like the Netherlands, France, Italy and Spain. These countries consistently appeared in our data as regions where platforms applied territorialised content moderation decisions and that were more likely to include manual means of moderation. However, this territorial variation does not necessarily imply a legal fragmentation. Rather, it suggests that platforms are more selective about the localisation of their content moderation strategies within the overarching harmonised regulatory environment established by the DSA. Additionally, this selective territorialisation of content moderation–namely concerning certain aforementioned member20 Internet Policy Review 14(3) | 2025 states–shown by our findings indicate a tiered approach to content moderation that not only has to do with national legal frameworks and authorities but, potentially, also with how large a market is and, subsequently, how many resources are invested in the form of human reviewers speaking the local language. Our analysis revealed a notable correlation between territorialised content moderation and the manual detection of flagged content, particularly through third-party notifications. Put simply, when content is moderated at the level of specific member-states rather than uniformly across the EU or EEA, it is more likely to have been flagged manually (e.g., through reports by users or other third-parties) rather than detected through automated systems. Under Article 16 of the DSA, platforms are required to implement user-friendly mechanisms enabling individuals to report potentially illegal content. However, in our dataset for TikTok, X, and YouTube, the majority of flagged content fell under the category of “other types of notifications.” This vague classification does not provide details about the origin of these notifications, whether from individuals, organisations, or other entities. Also, platforms are not obligated under DSA Article 9 to issue Statements of Reasons (SoRs) for content removed at the request of judicial or administrative authorities. This omission restricts our ability to differentiate between notifications stemming from public authorities versus other third parties in the context of the Transparency Database. However, a closer examination of the types of content flagged by third parties (Table 5) offers some insights. For instance, intellectual property violations overwhelmingly dominate YouTube’s moderated content, underscoring the influence of rightsholders on platform governance. On the other hand, TikTok and X show broader diversity in flagged categories, including harmful speech, fraud and privacy violations. This suggests that different platforms cater to varying types of stakeholders’ demands and observability can help us foreground these latent power dynamics. At any rate, our study shows that we are moving towards a harmonised framework of platform governance in the EU. This harmonisation, however, is largely enforced by platforms as the “delegated enforcers” under the DSA (Husovec, 2024). As Kaushal et al. (2024) showed in their study of the Transparency Database, most of the content moderated is found to be incompatible with the platforms’ terms of services and/or community guidelines rather than the basis of national or EU law. As a result, this delegation and interpretation of the DSA, raises broader questions about the power of platforms broadly and, particularly, the diversity and vibrancy of the EU’s digital public spheres. For instance, how can we ensure that platforms do not disproportionately silence marginalised voices when they have such discre21 Papaevangelou, Votta tion over the operationalisation of EU’s regulations? This delegation, moreover, of enforcement responsibilities to platforms, while practical from a regulatory perspective to ensure a more-or-less harmonised and “predictable” regulatory environment (Husovec & Roche Laguna, 2022), raises important concerns about the prioritisation of monocultural and compliant speech (Douek, 2020; Keller, 2024) and the potential cultivation of a sanitised digital public sphere (Griffin, 2022). Importantly, as Keller has repeatedly noted (2022, 2024), the DSA might end up favouring incumbent platform firms that have put in place an industrialised content moderation structure, while disproportionately affecting smaller platforms or discourage the experimentation with other systems of moderation, like more community-oriented or artisanal approaches (Caplan, 2018). Automation, to be sure, plays a pivotal role in facilitating platforms’ compliance with the DSA. In that sense, as demonstrated in our analysis of TikTok, industrialscale automation is not only necessary for the industrial-scale of platforms’ content circulation (Gillespie, 2020) but also streamlines processes of “automated transparency” (Kaushal et al., 2024) like the Transparency Database. This dynamic reflects a broader trend toward the industrialisation of content moderation, where platforms adopt algorithmic systems (Gorwa et al., 2020), create specialised Trust and Safety teams (Caplan, 2023) and implement transparency mechanisms to comply with regulations. All these elements give way to a “factory-like” approach to content moderation (Keller, 2024), reconfiguring platforms into compliance-driven entities that prioritise operational efficiency over nuanced content moderation, precisely due to the systemic shortcomings and problems plaguing automation. We also find that the way that the DSA operationalises platform observability does not leave much space either for the kind of technical modifiability that Ferrari et al. (2023) consider to be crucial for digital governance or for experimentations with alternative content moderation systems. To be clear, these obligations formally apply only to designated VLOPs. However, the normative gravitas of the DSA risks reinforcing industrialised, centralised moderation approaches across the digital ecosystem. As such, the DSA risks further entrenching large platforms’ power by restricting the potential of having a more plethoric and diverse content moderation ecosystem which would be more likely to be cultivated by smaller or decentralised platforms, which may now be pressured to align with industrial approaches to content governance (e.g., BlueSky’s approach to a ‘stackable’ content moderation system; The BlueSky Team, 2024). 22 Internet Policy Review 14(3) | 2025 It may, moreover, depoliticise content governance, alongside the process of making it legible and accountable, allowing platforms to hide behind a discursive framing of content moderation as the product of neutral, automated systems that are the only option to comply with the DSA, rather than as decisions embedded in complex sociopolitical and cultural contexts (Ananny & Crawford, 2016). Therefore, while the industrialisation of content moderation has become unavoidable due to the scale and complexity of digital platforms, its implications for platform governance demand critical scrutiny, that is our capacity to observe and probe into these systems. Implications for platform observability In the context of this paper, where we focus on the DSA, observability translates to the systematic and –wherever possible– real-time access to platforms’ data and the sociotechnical systems that allow for the production of that data and the overarching ecosystems. In other words, it refers to creation of the institutional and technical conditions necessary for the sustained, collaborative and situated observation of platforms. The Transparency Database introduced by the DSA partially responds to these calls for platform observability. Returning to Rieder and Hofmann’s conceptualisation, we show that the DSA, here through the Transparency Database, contributes to the normative condition by mandating transparency practices that aim to serve public scrutiny. It also aligns with the sociotechnical condition by enabling public access to structured and real-time data, though with well-documented limitations. Last, the governance condition is also arguably met provided that the Database can be accessed and used by a wide variety of stakeholders who may collectively participate in governance processes. However, once we consider Ferrari and colleagues’ governance conditions (2023), we understand that the Transparency Database partially meets the condition of industrial observability as it provides some insights into the labour relations supporting platforms’ content governance but there are significant inconsistencies and shortcomings. It also partially meets public inspectability as it makes a vast trove of data and insights available but primarily focuses on decisions of platforms’ automated decision-making systems and not so much on more opaque processes (e.g., the models undergirding content governance). Last, technical modifiability is absent as there is no indication of how moderation systems can be reconfigured in response to observed shortcomings beyond regulatory proceedings and potential financial penalties. In that sense, these SoRs provide us with justifications for the platforms’ decisions rather than explain how these decisions were made and car23 Papaevangelou, Votta ried out or empower the involved parties to take further action (Leerssen, 2024, p. 25). Regardless, we hold that this newfound level of access to important data might improve our chances for reigning in platform firms’ unchecked power, primarily through ex-post accountability (e.g., fines). But given the unprecedented level of entanglement of platform infrastructures in our lives and political-economic systems this might no longer be–if it ever was–enough. For example, as mentioned earlier, observability should expand from just the consumer-facing (i.e., downstream) part of platforms’ value chains to include more hidden layers (i.e., upstream) like the labour conditions of their outsourced contractors and environmental impact of their material infrastructures (Terzis, 2023). Recent stories like the layoff of 300 human moderators from the Amsterdam office of TikTok (Nijssen, 2024) hint at a doubling down on the technical aspect of moderation, namely automation, potentially exacerbating the shortcomings and problems of automation discussed earlier in this paper. We must not forget that platform firms, and their content moderation systems, are also large employers operating on human labour. In that sense, conditions of observability should also extend to include the working spaces and the relations of production of platform capitalism (Srnicek, 2017). Doing so should not only allow us to observe and, ideally, ensure that these moderators work under decent conditions but also to better understand the value chains of content moderation and how platforms coordinate their workforce globally and, in so doing, create new opportunities for interventions. We understand that such a condition cannot be solely addressed by platform regulation but may also need more traditional regulatory frameworks moored in labour law (e.g., content moderators might need a similar instrument to the EU’s platform work Directive). Conclusion This paper set out to explore how the DSA’s Transparency Database operationalises the concept of platform observability. Building on a literature review of platform governance scholarship, we contextualised our paper with a conceptualisation of platform observability, primarily following Rieder and Hoffman (2023)’s framework and complementing it with Ferrari et al. (2023)’s governance conditions. Empirically, we analysed nearly 439 million SoRs from eight digital VLOPs, with a deeper dive into three (X, YouTube and TikTok), aiming to understand what kind of insights related to observability we could gain from this novel regulatory instrument. We 24 Internet Policy Review 14(3) | 2025 EU OFFICIAL LANGUAGE YOUTUBE X TIKTOK German 231 81 869 Greek 28 0 96 Hungarian 25 0 63 Irish 0 0 439 Italian 91 2 0 Latvian 11 1 9 Lithuanian 11 0 6 Maltese 0 0 0 Polish 99 1 208 Portuguese 464 41 75 Romanian 34 0 167 Slovak 5 0 44 Slovenian 15 0 45 Spanish 507 20 468 Swedish 16 0 108 Total 1974 2496 5827 31 Papaevangelou, Votta