From privacy concerns to algorithmic biases, the systems recently deployed and advertised as "age verification" have the potential to be deeply problematic. I explore the different solutions within the technical guidelines established by regulators.

Characterizing the operation of 𝕏's crowd-sourced moderation system in 13 countries.

Overall we observe that:

• Notes are requested, proposed, and approved under posts throughout the political spectrum
• ~25% of notes contain links to news outlets, before 𝕏 posts (~17%) and Wikipedia articles (~9%). Expert fact-checks are used in only 3.5% of notes.
• 𝕏's algorithm successfully learns the ideological leaning of notes and raters
• Community Notes are predominantly associated (~65%) with political posts.
• Politically divisive accounts and topics (e.g. Israel-Palestine conflict) are under-moderated compared to fringe conspiracy theory and scams.

Overall, by relying on consensus rather than factuality, 𝕏's Community Notes struggle to moderate polarizing content. As YouTube, TikTok, and Meta transition to substitute expert fact-checking with crowd-sourced systems, our results call for risk assessment comparing moderation outcomes over polarized content widely disseminated during elections.

Exposing how 𝕏 allows brands to target users based on sensitive personal data.

We inspected the targeting options of ads disclosed in 𝕏's Ads Repository, what we found is concerning.

• Energy giant TotalEnergies excludes users interested in Green politicians and environmental activists
• Netflix France excludes users interested in the French far-right party
• McDonald's France blocks users interested in its labor unions
• BP Deutschland blocked users interested in "trans, transgender, gender transition" from their ads, as well as "al islam, quran, sunni, shiite, shias"

Exposing Large Scale Health Scams Targeting EU Users.

Our investigation reveals Meta's ongoing failure to moderate its advertisement ecosystem, particularly for health-related content.

We identified over 46k advertisements containing unapproved drugs and deceptive health claims that were shown by Meta to European users over 292 million times.

These ads violated at least 15 of Meta's own Advertising and Community Standards, featuring celebrity deepfakes, impersonation of medical professionals and news outlets, and misleading health claims.

This activity continues unabated into 2025, suggesting systemic failures in Meta's risk mitigation approaches required under DSA Article 34.

Exposing Pornographic Ads on Meta Despite Content Moderation Claims.

I detected thousands of pornographic advertisements featuring explicit adult content that were reviewed, approved, and distributed by Meta in clear violation of their Community and Advertising Standards.

To highlight the double standard, I uploaded the exact same pornographic visuals as organic user posts, which were automatically taken down by Meta.

Meta has the technology to detect pornographic content; our observations suggest they simply chose not to use it for advertisements, their core source of revenue.

From ads on Meta, to spoofed news outlet articles relayed on 𝕏 and leaked SDA documents.

In collaboration with CheckFirst and Reset.Tech, I cross-referenced thousands of Doppelganger Ads with leaked documents from the Social Design Agency (SDA), the company linked to the Kremlin behind the propaganda operation. Through this analysis, we documented the sustained access by SDA to Meta's Advertising Ecosystem even after being sanctioned by the US and EU.

Leveraging the methodology used to detect coordinated inauthentic behaviors, I monitor the Doppelgänger campaign.

• Ahead of the 2024 EU Parliamentary Election, we identified Doppelgänger ads targeting Italy and Poland, in addition to the historic targets, France and Germany.
• Doppelgänger ads disseminated propaganda and fake articles supporting far-right candidates in France.
• Among the ads uploaded by the threat actors, I detected a screenshot of their internal VK Teams, providing valuable insights into their operations.
• Monitoring and detecting new spoofed outlets: historia.fyi, spektrum.cfd, closermag.eu, psychologies.top

Ahead of the European and French elections, let's take a look at political exchanges on 𝕏.

By collecting over a million retweets of French political figures, I depicted the landscape of political exchanges on 𝕏. Filtering by topic reveals markedly different landscapes. Emmanuel Macron's party aligns closely with the Ecologists on the discussion of the Russian war in Ukraine but is completely opposed on the issue of the Israel's war on Gaza, where it is closer to the Far-right.

Analyzing over 200 millions ads approved by Meta reveal majors shortcomings in their moderation.

Thanks to the Digital Services Act Article 39, with AI Forensics, we tens of millions of advertisements run in the EU. By training language models to detect political advertisements, we assessed Meta's moderation, and documented:

• Widespread Non-compliance: Only a small fraction of undeclared political ads are caught by Meta's moderation system.
• Ineffective Moderation: 60% of ads moderated by Meta do not adhere to their own guidelines concerning political advertising.
• Significant Reach: A specific pro-Russian propaganda campaign reached over 38 million users in France and Germany, with most ads not being identified as political in a timely manner.
• Rapid Adaptation: The influence operation has adeptly adjusted its messaging to major geopolitical events to further its narratives.

Examination of algorithmic recommendations on Amazon's Belgian and French bookstores.

As a result of a collaboration between AI Forensics and CheckFirst, we reveal that:

• Amazon's search results fail to provide a pluralism of views and convey misleading information, specifically on public health, immigration, gender, and climate change.
• Amazon's recommender systems trap users into tight book communities, which are hard to exit, and some of these communities contain books endorsing climate denialism, conspiracy theories, and conservative views.
• Amazon fails to enforce its policies, resulting in the widespread availability of sexually explicit content without age restrictions.

We urged regulators to ensure that Amazon comprehensively addresses the impact of its systems on the dissemination and amplification of disinformation content.

By collecting what platforms recommend to users, I characterize their inner workings.

To characterize the recommender systems of Facebook, Google Search, YouTube, and Twitter, I developed a browser add-on that collects algorithmically curated content, as well as user interactions. This new, longitudinal study, running since Fall 2022, has enabled numerous studies, such as an audit of Twitter's 'For You' feed and characterizing the differences between YouTube's API and what real users are exposed to.

Social macroscope for a better understanding of the circulation of climate change narratives.

In March 2023, prior to the release of the IPCC AR6, we leveraged extensive databases containing over 500 million tweets, curated by the amazing Maziyar Panahi, to analyze the global discussion on climate change.

• The global climate change debate on Twitter is highly polarized, with about 30% climate denialists among Twitter accounts addressing climate issues from 2019 to 2022, rising to over 50% in Elon Musk's Twitter.
• The denialist community has 71% more inauthentic accounts compared to pro-climate communities.

Bouchaud P, Ramaciotti P. June 2025 - Preprint

We characterize the operation of 𝕏's crowd-sourced moderation system in 13 countries, we observe that:

• Notes are requested, proposed, and approved under posts throughout the political spectrum
• ~25% of notes contain links to news outlets, before 𝕏 posts (~17%) and Wikipedia articles (~9%). Expert fact-checks are used in only 3.5% of notes.
• 𝕏's algorithm successfully learns the ideological leaning of notes and raters
• Community Notes are predominantly associated (~65%) with political posts.
• Politically divisive accounts and topics (e.g. Israel-Palestine conflict) are under-moderated compared to fringe conspiracy theory and scams.

Bouchaud P, Ramaciotti P. September 2024 - Applied Network Science

Through a simulated Twitter-like platform designed to optimize user engagement and grounded in authentic behavioral data, we evaluates methodologies for auditing social media recommender systems.

Bouchaud P, Liénard J. April 2024 - IMC '24

This study evaluates Meta's enforcement of political advertising policies across 16 EU countries, finding imprecise moderation with a 60% false-positive rate and a 95.2% false-negative rate. Coordinated advertising campaigns relaying pro-Russian propaganda, reached 38 millions of accounts since August 2023. Despite documented activities, less than 20% of such pages were moderated by Meta as political. With upcoming elections, these findings highlight significant shortcomings in Meta's political ad moderation.

Romano S, Angius R, Kerby N, Bouchaud P, Amidei J, Kaltenbrunner A. April 2024 - ICWSM 2024

This study describes a dataset that allows to assess the emerging challenges posed by Generative Artificial Intelligence when doing Active Retrieval Augmented Generation (RAG), especially when summarizing trustworthy sources on the Internet

Bouchaud P, Çetin R.B. January 2024 - Tech Policy Press

We consider repercussions of a lack of diversity in book recommendations as a possible source of risk to information integrity.

Bouchaud P. December 2023 - Social Networks Analysis and Mining

We investigate Amazon's book recommendation system, uncovering cohesive communities of semantically similar books. We identify a large community of recommended books endorsing climate denialism, COVID-19 conspiracy theories, and conservative views. This study underscores how even non-personalized recommender systems can have foreseeable negative effects on public health and civic discourse.

Bouchaud P. August 2023 - Complex Networks and their Applications

We examine how engagement-maximizing recommender systems influence the visibility of Members of Parliament's tweets in timelines. We showcase the need for audits accounting for user characteristics when assessing the distortions introduced by personalization algorithms and advocate addressing online platform regulations by evaluating the metrics platforms aim to optimize.

Bouchaud P. June 2023 - Journal of Computational Social Science

This article investigates the information landscape shaped by curation algorithms that seek to maximize user engagement. Leveraging unique behavioral data, we trained machine learning models to predict user engagement with tweets. Our study reveals how the pursuit of engagement maximization skews content visibility, favoring posts similar to previously engaged content while downplaying alternative perspectives. The empirical grounding of our work contributes to the understanding of human-machine interactions and provides a basis for evidence-based policies aimed at promoting responsible social media platforms

Bouchaud P, Chavalarias D, Panahi M. March 2023 - Scientific Reports

Combining crowd-sourced data donation and large-scale server-side data collection, we provide quantitative experimental evidence of Twitter recommender distortion of users' subscriptions choices. In particular, we reveal:

• Toxic tweets are amplified by 50%
• Tweets from friends from the same community are highly amplified
• Uneven amplification across friends' political leaning

Chavalarias D, Bouchaud P, Chomel V, Panahi M. February 2023

Analyzing two years of Twitter exchanges, we observe that the denialist community presents inauthentic forms of expertise, relays more toxic tweets and embeds +71% inauthentic accounts with respect to the Pro-Climate community. Pro-climate accounts fleeing from Elon Musk's Twitter, climate skeptic accounts represent 50% of the online discussion by March 2023.

Chavalarias D*, Bouchaud P*, Panahi M. February 2023 - Journal of Artificial Societies and Social Simulation

After having calibrated an agent-based model over a large scale longitudinal database of tweets from political activists, we compare the consequences of various recommendation algorithms on the social fabric and quantify their interaction with some cognitive biases.