Audit of takedown delays across social media reveals failure to reduce exposure to illegal content

Illegal content on social media poses significant societal harm and necessitates timely removal. However, the impact of the speed of content removal on prevalence, reach, and exposure to illegal content remains underexplored. This study examines the relationship with a systematic audit of takedown delays using data from the EU Digital Services Act Transparency Database, covering five major platforms over a one-year period. We find substantial variation in takedown delay, with some content remaining online for weeks or even months. To evaluate how these delays affect the prevalence and reach of illegal content and exposure to it, we develop an agent-based model and calibrate it to empirical data. We simulate illegal content diffusion, revealing that rapid takedown (within hours) significantly reduces prevalence, reach, and exposure to illegal content, while the longer delays measured by the audit fail to reduce its spread. Though the link between delay and spread is intuitive, our simulations quantify exactly how takedown speed shapes exposure to illegal content. Building on these results, we point to the benefits of faster content removal to effectively curb the spread of illegal content, while also considering the limitations of strict enforcement policies.

💡 Research Summary

This paper conducts a systematic audit of takedown delays for illegal content across five major social‑media platforms—Facebook, Instagram, YouTube, TikTok, and Snapchat—using the European Union Digital Services Act Transparency Database (DSA‑TDB). By extracting Statements of Reasons (SoRs) filed between 1 January and 31 December 2024, the authors compute the elapsed time between content posting and removal (τ) for each case. The dataset comprises 8,613 SoRs for Facebook, 3,005 for Instagram, 408,400 for YouTube, 90,409 for TikTok, and 2,142 for Snapchat. The empirical analysis reveals a wide spread of τ values: from a few days to several weeks, and in rare instances up to a year. These observed delays substantially exceed the statutory deadlines that existed under national laws such as Germany’s NetzDG (24 hours for “obviously illegal” content) and France’s Loi Avia (seven days for other illegal material), indicating that current enforcement under the DSA’s more flexible “timely” requirement is often far from the intended rapid response.

To translate these observed delays into measurable effects on the prevalence, reach, and exposure of illegal content, the authors develop an agent‑based simulation built on SimSoM, a model of information diffusion on directed follower networks. The underlying network is derived from a large Twitter/X follower graph, providing realistic degree distributions and clustering. Each agent is assigned an activity level a drawn from a power‑law distribution P(a) ∝ a^γ with γ≈2.85, calibrated from nine‑day posting data (≈81 million posts from ≈38 million accounts). At each discrete time step, an agent may (i) create a new original post, (ii) reshare a message from its feed, or (iii) remain idle. Resharing probabilities depend on the message’s recency, popularity, and the author’s follower count, thereby approximating platform ranking and recommendation mechanisms.

Illegal content is modeled as a binary attribute attached to each message. The probability that a newly created message is illegal (p) is not directly observable; the authors infer a plausible upper bound from the “Future of Free Speech” policy report, which estimates illegal comment fractions between 0.00002 and 0.009 across several European countries. For robustness they run simulations across this range but adopt p = 0.01 as a representative value, noting that key outcomes are insensitive to the exact p. Three alternative distributions of user‑level illegal‑posting propensity are examined: (i) a homogeneous scenario where all users share the same p, (ii) a heterogeneous scenario with a continuous distribution of p across users, and (iii) a two‑group scenario (high‑risk vs. low‑risk users) modeled with beta distributions. All three produce qualitatively similar results, confirming that the aggregate prevalence of illegal content, rather than its precise user‑level allocation, drives the dynamics.

The core of the simulation links τ to a stochastic removal process. For a given τ, the survival probability per time step (p_s) is set such that the expected lifetime of an illegal post matches τ (i.e., p_s = exp(−1/τ) under a geometric decay assumption). When a post is removed, it disappears from all followers’ feeds, halting further resharing. The authors run the model for τ values ranging from immediate removal (0 hours) to extended delays of 30 days or more, generating 10,000 independent runs for each setting to obtain stable averages.

Three outcome metrics are recorded: (1) prevalence – the fraction of all active posts that are illegal at any given time, (2) reach – the total number of user‑views generated by illegal posts across the simulation horizon, and (3) exposure – the average number of illegal‑post views per user. The results display a pronounced non‑linear relationship. When τ ≤ 6 hours, especially under 1 hour, prevalence drops by roughly 70 % relative to a no‑removal baseline, and reach and exposure are similarly curtailed. As τ increases, the marginal benefit of faster removal diminishes sharply; beyond roughly two weeks, the prevalence, reach, and exposure curves converge to the no‑removal baseline, indicating that delays of that magnitude render takedown policies essentially ineffective. The steep rise in all three metrics occurs between τ ≈ 1 day and τ ≈ 2 weeks, highlighting a critical window where policy interventions could have the greatest impact.

The authors discuss the policy implications of these findings. First, the empirical audit shows that current platform practices routinely exceed the “reasonable speed” envisioned by the DSA, suggesting that the regulation’s flexibility may be too permissive to achieve its protective goals. Second, the simulation quantifies how much faster removal would be needed to meaningfully curb illegal content diffusion, providing a data‑driven benchmark for legislators. Third, the authors acknowledge the trade‑off between speed and accuracy: ultra‑rapid deletions could increase false positives, potentially infringing on legitimate expression. They argue that any tightening of deadlines must be accompanied by investments in automated detection accuracy, transparent appeals processes, and consideration of platform operational constraints (e.g., moderation staffing, algorithmic latency).

Limitations are openly addressed. The model assumes independence between illegality and content appeal, whereas in reality certain illegal categories (e.g., copyrighted memes) may be highly viral, while others (e.g., child sexual abuse material) are less likely to be reshared. The SoR data may suffer from reporting bias, as platforms only flag content they deem illegal; unreported illegal material is invisible to the audit. Moreover, the follower network is static, ignoring dynamic changes in connections that can affect diffusion pathways. Finally, the simulation does not incorporate user‑level behavioral adaptations (e.g., users avoiding platforms known for slow takedowns).

Future work is suggested along several axes: (i) extending the model to differentiate illegal content types with empirically measured attractiveness, (ii) incorporating dynamic network evolution and user migration, (iii) modeling the impact of false‑positive rates and appeals on overall system performance, and (iv) evaluating cross‑platform coordination where content may migrate between services with differing takedown speeds.

In sum, this study provides the first large‑scale empirical measurement of illegal‑content takedown delays across major platforms and couples it with a calibrated agent‑based diffusion model to quantify the causal impact of those delays on content prevalence, reach, and user exposure. The findings demonstrate that while rapid (hour‑scale) removal can dramatically suppress illegal content spread, the actual delays observed in practice are often long enough to render takedown policies ineffective. The paper thus offers concrete evidence for policymakers and platform operators to reconsider the adequacy of current “timely” standards and to invest in faster, more accurate moderation infrastructures.