Luka Bekavac

Auditing the algorithms that shape what billions see.

I am a PhD candidate in Computer Science at the University of St. Gallen, in the Interactions group led by Prof. Dr. Simon Mayer, and a member of the Legal Design & Code Lab.

I audit Very Large Online Platforms to understand the systemic risks their design produces. That work runs along two tracks. The first shows how far platform-provided data access falls short of what independent scrutiny requires, measuring the gap between what users see and what Article 40 research APIs actually return. The second builds the methods that close it, from sock-puppet auditing with SOAP, my open-source framework, through to code-level audits of recommendation pipelines.

I am also interested in whether more decentralized and open architectures can address these risks at the root of platform design rather than after the fact — through data portability under the DMA, user-controlled Solid Pods, and the open hypermedia principles that platform silos have eroded.

Outside research: hiking, chess, and padel.

Current Projects

Research

Auditing what recommender systems actually show people Building and running SOAP, an open-source sock-puppet auditing framework for Very Large Online Platforms, and extending it beyond elections to other systemic risks. It doubles as a teaching tool: in a workshop with over 100 participants it measurably raised algorithmic literacy.
Researcher data access under Article 40 DSA Measuring the gap between what platforms publish and what their research APIs hand back, and what that means for anyone trying to audit systemic risk independently.
How platforms label AI and deepfake content A cross-platform audit of labelling practice under EU law, with first results now on arXiv.
Turning data portability into data control BridgingWorlds — converting the proprietary exports platforms owe under GDPR Art. 20 and DMA Art. 6(9) into linked data in user-controlled Solid Pods.

Policy

Writing on DSA enforcement Most recently in Tech Policy Press, on what X's corrective action does and does not fix for researcher data access.
Bringing audit evidence to regulators Presenting at the Arcom Study Day, hosted by France's Digital Services Coordinator, and at CPDP.
CoCoDa SNSF-funded project with the Legal Design & Code Lab on how law and computer science together can make platform power more transparent and controllable.
Advising on data access policy Expert reviewer for Better Access: Data for the Common Good (Knight-Georgetown Institute, 2025), a framework for independent access to public platform data. Consulted through the Coalition for Independent Technology Research's Data Access Task Force on Mind the Gap (GoodBot, 2026), a submission on Canada's Bill C-34 that draws on my work to argue for protecting research methods which do not depend on platform permission.

News

Nov 2026 Two abstracts at the ECAT Research Workshop in Seville, 25 November, both appearing in the Book of Abstracts.Upcoming
Nov 2026 Two papers at the 5th Arcom Study Day in Paris, hosted by France's Digital Services Coordinator.Upcoming
Sep 2026 New preprint: Drowning in AI Slop, auditing how Instagram, TikTok, X and YouTube label AI and deepfake content under EU law.
Sep 2026 At ACM RecSys 2026, 27 September – 2 October.
Jul 2026 Wrote in Tech Policy Press on what X's DSA corrective action does and doesn't fix for researcher data access.
Jun 2026 Paper at ACM FAccT 2026 in Montréal, auditing what Meta and TikTok research APIs actually return.
May 2026 Spoke at CPDP (Computers, Privacy & Data Protection), 19–22 May.

Papers Google Scholar →

Drowning in AI Slop: How Social Media Platforms (Do Not) Label AI and Deepfake Content under EU Law B. Rijsbosch, L. Bekavac, H. Tari, G. van Dijck, K. Kollnig arXiv preprint, 2026
A systematic audit of how Instagram, TikTok, X and YouTube actually label AI-generated content, across more than 10,000 posts plus controlled uploads of AI material. Labelling is now broadly established but badly incomplete: only 33% of expert-identified deepfakes in high-risk contexts carried a platform-applied label, and platforms labelled just 61% of the controlled uploads, with policies, label designs and detection methods diverging sharply between them.

arXiv
Scrutinizing Systemic Risks in Personalized Recommender Systems Through Sock-Puppet Auditing of VLOPs L. Bekavac, J. Strecker-Bischoff, K. Garcia, S. Mayer, A. Tamò-Larrieux ACM Transactions on Recommender Systems, 2026
Introduces SOAP, an open-source sock-puppet auditing framework that reaches the user-level granularity platform data access does not provide. Applied to TikTok during the 2024 German regional elections, it found that politically neutral accounts were three to four times more likely to be shown right-wing populist content than mainstream party content.

DOI PDF Code
Platforms' Research API Data Access: What Users See vs. What Researchers Can Retrieve L. Bekavac, S. Mayer ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2026, pp. 8276–8299
Compares what the TikTok Research API and Meta Content Library return against what is actually publicly visible on those platforms, using controlled accounts during election periods. Three classes of platform-imposed loss emerge — scope narrowing, metadata stripping, and operational restriction — removing roughly half of visible content and up to 83% of contextual metadata, which falls short of supporting independent auditing of systemic risks under Article 40(12) DSA.

DOI arXiv
Broken Links and Fading Stories: How Dominant Online Platforms Subvert the Hypermedia Foundation of the Web L. Bekavac, S. Mayer 18th ACM Web Science Conference (WebSci), Braunschweig, 2026
The Web was envisioned as an open, decentralized network of documents joined by hyperlinks. Dominant platforms depart from that architecture — constraining linking, limiting navigation, and normalizing ephemerality — turning an open hypermedia environment into platform-governed silos. Users answer with transcension practices (reposts, screenshots, off-platform link hubs), but these are ad-hoc workarounds; the paper argues for reclaiming the hyperlink as a first-class architectural primitive.

DOI
From Data Portability to Data Control: Making Regulation Technically Real L. Bekavac Poster, 4th Solid Symposium, London, 2026
GDPR Art. 20 and DMA Art. 6(9) mandate social media data portability, but platform exports arrive as proprietary dumps with no semantic structure — technically compliant, practically useless for switching. Presents an end-to-end pipeline that uses LLM-assisted RDF conversion to turn those exports into structured linked data in user-controlled Solid Pods, enabling import into decentralized services such as Bluesky.

Zenodo Poster
A Solid Research Infrastructure Leveraging DMA Data Portability L. K. Seiling, J. Stein, S. Mayer, L. Bekavac 2026
Proposes a shared research infrastructure that turns the data-portability rights created by the Digital Markets Act into a usable basis for studying platforms, with exported data held in Solid pods.

OpenReview
Connecting Personalized Realities: Challenges and Opportunities in a Personalized Society J. Strecker-Bischoff, L. Bekavac, S. Mayer, K. Bektas 1st Workshop on Shaping Future Human Connection: Social Augmentation through XR Technologies, at ACM CHI 2026
XR and AI increasingly mediate how people perceive and navigate reality. Such Personalized Realities can make interaction more effective and information more equitable, but they also risk isolating people in perceptual filter bubbles and eroding shared understanding. Introduces the term Personalized Society, maps the tensions it creates, and asks how far frameworks like the Digital Services Act can protect individual and societal rights within one.

PDF
Change Your Perspective, Widen Your Worldview! Societally Beneficial Perceptual Filter Bubbles in Personalized Reality J. Strecker, L. Bekavac, K. Bektaş, S. Mayer arXiv preprint, 2025
Argues that Extended Reality risks producing perceptual filter bubbles, exposing people predominantly to content congruent with their existing beliefs and narrowing the worldview they can even perceive. Proposes carrying mitigations already shown to work in recommender systems — perspective awareness, engineered serendipity, and support for self-actualization — into XR, so people keep transparency and meaningful control over their personalized reality.

arXiv
Visualizing Children's Filter Bubbles: A Digital Window for Bridging Social Media Experiences L. Bekavac, L. Bekavac, B. Bekavac MIPRO 2025 — 48th ICT and Electronics Convention, Opatija
Children are still forming their critical thinking, so the filter bubbles that recommender systems build around them shape worldviews in ways parents rarely see. Uses SOAP to steer and replicate the feeds children actually experience, then presents an interactive interface that visualizes those bubbles for parents. The aim is to make generational differences in digital experience comparable, and to give families a concrete starting point for talking about algorithmic personalization.

DOI
QR-Code Integrity by Design L. Bekavac, S. Mayer, J. Strecker-Bischoff CHI Conference on Human Factors in Computing Systems, Extended Abstracts, 2024
QR codes are everywhere, which makes tampering with them — quishing — a real threat. Introduces SafeQR codes, which use visual design and secure design principles to make tampering visible rather than relying on user education alone. A user study shows these "integrity by design" codes measurably raise suspicion when tampered with and reduce successful quishing attacks.

DOI PDF
From Walls to Windows: Creating Transparency to Understand Filter Bubbles in Social Media L. Bekavac, K. Garcia, J. Strecker-Bischoff, S. Mayer, A. Tamò-Larrieux NORMalize 2024 — 2nd Workshop on the Normative Design and Evaluation of Recommender Systems, at ACM RecSys 2024
Filter bubbles have been hard to study at scale despite years of research. Introduces SOAP, which takes a topic prompt and uses a multimodal LLM to create and navigate filter bubbles inside a real feed, logging what it sees. Across three induced bubbles, topic diversity collapsed in as little as 60 minutes of scrolling.

PDF Code
Towards an Integrated Semantic Approach to GHG Accounting L. Bekavac, L. Volk University of St. Gallen, Integrative Master Project, 2024
Switzerland's incoming reporting rules require comprehensive greenhouse-gas accounting, but existing formats and methodologies do not interoperate. Proposes an ontology derived from life-cycle-assessment terminology as a unifying layer, combined with linked-data distribution over Solid Pods for granular access control and real-time updates.

PDF

Talks All talks →

Nov 2026 ECAT Research Workshop 2026, SevilleUpcoming Two abstracts accepted at the European Commission's Centre for Algorithmic Transparency, on DSA transparency for cyber violence against women and on a code-level audit of X's recommendation pipeline.
Nov 2026 5th Arcom Study Day, ParisUpcoming Two papers accepted at the study day of France's Digital Services Coordinator, held with ENS Paris-Saclay.
Jun 2026 ACM FAccT 2026, Montréal Platforms' Research API Data Access: What Users See vs. What Researchers Can Retrieve. 25–28 June 2026.

Media & Features

Tech Policy Press · July 2026 What X's Corrected Action Under the DSA Fixes on Data Access, and What's Left Open
With Alice Palmieri. On the European Commission accepting X's corrective action plan under the Digital Services Act — and why fixing access procedures leaves platforms' internal filtering untouched.
SRF · December 2025 Radikalisierung auf Tiktok: Der toxische Sog der Manosphere
SRF Data ran five sock-puppet TikTok accounts posing as 15-year-old Swiss users to measure how quickly the For You feed pulled them toward manosphere content, applying parts of the SOAP methodology.
Die Zeit · February 2025 So sieht der TikTok-Kosmos der AfD aus
ZEIT ONLINE analysed 602 videos from the AfD's TikTok sphere ahead of the 2025 German federal election. The videos were collected by a researcher at the University of St. Gallen using automated accounts.
24 heures / Tribune de Genève · March 2025 Les partis d'extrême droite sont devenus parmi les plus populaires en Europe
Quoted on how fast filter bubbles close: after two to three hours of scrolling and interacting with right-wing content on TikTok, more than 75% of posts already match the bubble — after roughly three weeks, 98%.

Academic Info

PhD in Computer Science, University of St. Gallen (HSG) — since 2025 Interactions group, advised by Prof. Dr. Simon Mayer. Recommender systems, platform data access, and the auditing of Very Large Online Platforms.
M.Sc. in Computer Science, University of St. Gallen (HSG)
B.Sc. in Business Information Systems, Furtwangen University (HFU), Germany