EU Social Media Regulation (Credit: Midjourney)
Are social media platforms’ moderation systems actually catching harmful or illegal content?
USC researchers identified a loophole in the European Union (EU)’s laws covering social media regulation.
The legislation required social media companies to report the “accuracy” of their automated content moderation systems but did not clearly define what accuracy meant or how it should be measured. The researchers argued that this ambiguity could allow platforms to report numbers that made their systems appear effective without revealing how much illegal content they were actually missing.
Platforms could choose metrics most beneficial to them, such as literal accuracy, which can be highly inflated because the vast majority of social media content is benign.
In a new study, USC computer scientists Robin Jia and his PhD student Johnny Wei teamed up with German legal scholar Frederike Zufall to tackle this challenge and develop metrics that better measure the effectiveness of content moderation.
Started in 2022, the research project resulted in a paper titled “Operationalizing Content Moderation ‘Accuracy’ in the Digital Services Act,”(DSA) which was accepted to the 2024 AI, Ethics, and Society (AIES) conference.
The study was among the first to examine the DSA’s undefined accuracy requirement through both legal and statistical analysis, identifying the reporting loophole and proposing a practical solution.
The team’s recommendations later informed changes to the EU’s Digital Services Act (DSA), a landmark law governing how major online platforms operate.
In 2024, the European Commission adopted updated reporting requirements that require platforms to report precision and recall metrics for their automated content moderation systems, giving regulators greater insight into how much illegal content their systems identify and miss.
Jia is an assistant professor of computer science at the USC Viterbi School of Engineering and the USC Mark and Mary Stevens School of Computing and AI‘s Thomas Lord Department of Computer Science. Zufall is a law faculty member at the Karlsruhe Institute of Technology, with affiliations at Waseda University and the Max Planck Institute.
The project was funded through several grants Jia received, including Coefficient Giving, Cisco Research and Google Research.
A Loophole Hidden in “Accuracy”
Under the DSA, platforms were required to report the accuracy and potential error rates of their automated moderation systems.
However, the researchers found that “accuracy” was left open to interpretation.
The researchers proposed separating accuracy into two more meaningful measures: precision and recall.
Precision measures how often content identified as illegal is actually illegal. It can help reveal whether a platform is over-moderating legitimate speech.
Recall measures how much of the illegal content that exists is actually identified and removed. It can reveal whether a platform is failing to moderate harmful material.
Together, the measures provide a more complete picture of whether a platform is striking an appropriate balance between freedom of expression and protection from illegal content.
Making Transparency Affordable and Practical
Measuring recall presents a significant technical challenge. Platforms process enormous volumes of content, making it impractical to manually examine every post to determine how much illegal material their systems missed.
The researchers developed a statistically efficient alternative based on stratified sampling.
Instead of relying on a purely random sample, which could easily miss rare examples of illegal content, the method uses machine learning classifiers to divide content into different groups and then samples strategically across those groups.
The researchers showed that this approach can produce reliable estimates of recall at a fraction of the cost of examining content indiscriminately.
For platforms, that means measuring moderation performance does not have to be prohibitively expensive or technically impractical.
From EU Law to a Global Standard
The researchers’ work came as the EU was developing one of the world’s most consequential frameworks for regulating online platforms.
The DSA was adopted in 2022 and established broad obligations for online services, including requirements concerning illegal content, transparency and platform accountability. The law applies especially to Very Large Online Platforms, defined as platforms with more than 45 million average monthly active users in the EU.
The EU released a draft in December 2023 that still called for reporting “accuracy” and possible error rates without explicitly requiring precision and recall.
That changed during the legislation’s public feedback process.
In early 2024, members of the public were able to submit comments on the proposed rules. Two specifically recommended adding recall as a required measure, with one citing the USC paper to support the legal and statistical case for requiring it.
In November 2024, the European Commission officially adopted Implementing Regulation (EU) 2024/2835, which explicitly requires platforms to report accuracy, precision and recall for their automated moderation systems.
What began as a research project focused on an ambiguous word in a European regulation had become an example of how interdisciplinary academic research can help shape the rules governing some of the world’s largest technology platforms.
Wei explained that this change has the potential to reach beyond Europe, emphasizing the EU’s position as the world’s second-largest market. Companies often adopt EU requirements across their global operations rather than maintaining separate technical systems for different markets, a phenomenon known as the “Brussels Effect.” This is evident as American social media companies such as Meta and X have started following the new recall requirement.
Published on September 1st, 2026
Last updated on September 1st, 2026