Trends - what is happening at EPFL on MDH

1. EPFL has substantive but fragmented MDH capacity. Across the 13 curated interviews and 133 MDH-relevant publications mapped on this site, expertise exists in adversarial robustness and LLM poisoning (Frossard, LTS4), AI safety and "model raising" (West, DLAB), federated and robust learning (Kermarrec), trust indicators and standardisation (Ebrahimi, MMSPG, chair of the JPEG committee and vice-chair of the ITU/ISO/IEC AI & Multimedia Authenticity Standards Cooperation), propaganda mitigation and moderation tools (Troncoso), and people-centred technology design (Gatica-Perez, Idiap). No lab has MDH as its primary mandate; the recurring pattern is "MDH-adjacent work as part of a broader research agenda."
   SourceInterviews 13 actorsInsights Fragmented but Real Capacity
2. The technical centre of gravity has moved from detection to trust, provenance and moderation. Süsstrunk (Dean IC) is explicit that pure detection is not a defensible institutional output, an arms race that EPFL is not equipped to keep running at the front, and her former deepfake-detection work has wound down for this reason: her postdoc on the topic is now in Paris and the underlying technique (frequency-domain artifacts of Stable Diffusion) is no longer relevant against modern transformer-based generators. Ebrahimi reaches the same conclusion from the standards angle: no single detector solves the problem; the answer is layered trust indicators (watermarking, provenance, fact-checking) combined with a per-user trust profile, which is what his JPEG Trust standard (ISO/IEC, 2024, with a 2026 second edition adding copyright provisions) operationalises.
   SourceInterview SüsstrunkInterview EbrahimiInsights Trust Indicators > Single DetectorsInsights The Cat-and-Mouse Problem
3. AI safety and robustness are the most active research front at EPFL today. Frossard's lab has reoriented from network ML toward AI safety, with a flagship LLM-poisoning project run with armasuisse. West's "model raising" approach (training models to internalise ethical reflection token by token, rather than RLHF wrapped around a "monster") is directly MDH-relevant. Kermarrec's work on federated learning, fairness and robustness against polluting nodes is structurally MDH-adjacent. The technical front is clearly visible; the institutional packaging as "MDH research" is not yet there.
   SourceInterviews Frossard, West, Kermarrec
4. EPFL has live institutional dialogues that translate research into use. EPFL/RTS discussion on disinformation around the 2027 federal elections (Bugnion, VP Innovation); EssentialTech dialogue with ICRC on information manipulation as a humanitarian concern; armasuisse partnership on LLM poisoning (Frossard); Geneva AI Initiative participation through the AI Center. Bridges to practice exist, they are individual relationships, not a coordinated EPFL position.
   SourceInterviews Bugnion, FrossardEPFL EssentialTech, AI Center
5. Three EPFL-internal MDH mappings run in parallel. This project (EssentialTech / PeaceTech), one at the AI Center, and one at C4DT. Frossard raised the duplication risk explicitly and asked to be re-engaged once the mappings reach a sharable state. The first move on coordination is internal to EPFL.
   SourceInterview Frossard
6. Other Swiss universities are active on MDH, with distinct centres of gravity. The clearest picture today is not "EPFL alone in Switzerland", it is a multi-institution landscape where labs cluster differently. The closest peers:
   • ETHZ Center for Security Studies (CSS) hosts the "Digital Disinformation in World Politics" project (Lennart Maschmeyer, 2019-2024, with Citizen Lab and LSE), centred on hybrid warfare and the Russo-Ukrainian conflict (Ukraine's October 2020 local elections as case study)^[212]. Adjacent CSS work covers AI governance of disinformation (CSS analyses on "When AI Governs (Dis)information" and "Disinformation: A Measured Response"). CSS sits closer to international security policy than to technical methodology. What CSS does not appear to centre, based on a review of their public pages as of 2026-05-18: peer-reviewed technical detection methodology, multilingual or low-resource hate-speech research, operational-humanitarian MDH work.
   • ETHZ International Conflict Research (ICR) publishes empirical work on counterspeech, including an empathy-based counterspeech field experiment showing measurable reduction in racist hate speech on a social media platform^[180]. This is one of the most directly H-centred individual studies in Swiss academia. What is not visible from their public pages: a sustained hate-speech research programme beyond the published field experiment.
   • Zurich Hub for Ethics, Technology and Society (ZHET), ETHZ in partnership with the ICT4Peace Foundation, runs a research line on AI, disinformation, misinformation and hate speech^[181].
   • ETH AI Center (139 professors, 36 core + 103 associated as of May 2026) has labs whose work is MDH-adjacent, e.g. SRI Lab (Secure, Reliable, and Intelligent Systems) on LLM reliability, security, AI regulations and robustness, and Agentic Systems Lab on multimodal AI. Based on a review of the ETH AI Center research pages as of 2026-05-18, no dedicated MDH lab or MDH-named programme is visible; the AI Center's MDH-relevant work is downstream of broader AI-safety and AI-reliability lines.
   • UZH Digital Society Initiative (DSI) ran its fourth Strategy Lab on disinformation and platform regulation from 2024, drafting a consultation response to Swiss federal platform legislation in early 2026 through an interdisciplinary core team (computer science, law, ethics, political science, communication). UZH's Digital Democracy Lab (Karsten Donnay) applies computational social science to Swiss democracy, in collaboration with ETHZ and UCL. A new "Digital Publics" DSI Community holds its first annual conference 21-23 October 2026^[179]. What is not visible from a review of DSI's public pages: peer-reviewed technical detection or robustness methodology as a primary output (their focus is interdisciplinary policy and computational social science).
   • Idiap Research Institute, affiliated with EPFL, hosts Daniel Gatica-Perez's Social Computing group on social media analytics for social good (already mapped on this site).
   The pattern: ETHZ centres MDH in international-security and conflict-research framings; UZH centres it in digital-society, platform-regulation and computational social science framings; EPFL (this project's home) centres it in technical research (AI safety, robustness, trust standards) plus PeaceTech / humanitarian framings. The three are complementary, not competitive, and not currently coordinated. None of the four sister-institution clusters (ETHZ, UZH, Idiap, EPFL) has a dedicated MDH-named centre or programme as of this review.
   SourceRefs 105, 106, 107, 108
7. Wider context (independent of EPFL, but defining the urgency). EU governance has consolidated: the Code of Practice on Disinformation was integrated into the Digital Services Act framework on 1 July 2025^[99,98]; the Permanent Task-force's Rapid Response System ran during the 2024 European Parliament elections and the French and Romanian national elections^[99]. Switzerland is regulatory-light by comparison, with no DSA equivalent per Federal Council Report 22.3006^[7]. Generative AI is accelerating M, D and H simultaneously: WEF Global Risks ranks disinformation as the first short-term societal risk in 2024 and 2025^[1]; the unit cost of disinformation has collapsed^[6,19]; synthetic abuse is scaling^[97]; detection systems lag generation^[80].
   SourceRefs 1, 6, 7, 19, 80, 97, 99, 100, 101

Gaps - what EPFL is missing on MDH

1. No dedicated MDH centre with engineering staff. Süsstrunk and Ebrahimi make the same institutional argument from different angles. Süsstrunk: NGO and partner collaborations on MDH are serendipitous; they work only when timing, chemistry and data happen to align, which is rare. Most fail because partner timelines do not match the PhD cycle, data takes months to access, or the topic is not publishable. "If EPFL wants to actually deliver on this for partners, it needs a centre with its own engineering staff that can absorb that friction." Ebrahimi: "What is worse than not doing something is to waste money that could have been spent somewhere else doing something that is useless."
   SourceInterview SüsstrunkInterview EbrahimiInsights Centre vs. Coordination
2. Misaligned individual incentives. All interviewees acknowledged MDH as important; none prioritise it. Evaluation metrics, data-access barriers, short project timelines, and the publishable-research basis of PhD evaluation (which Ebrahimi explicitly flagged as incompatible with interdisciplinary engagement) push researchers away from MDH.
   SourceInterviews all intervieweesInsights The Incentive Gap
3. Three parallel EPFL mappings without coordination. Wasted effort and a confused signal to external partners. Frossard's explicit ask is the gap statement.
   SourceInterview Frossard
4. Hate speech is the under-represented pillar inside EPFL. Most EPFL MDH-relevant work is on disinformation (West, Frossard, Ebrahimi standards) or AI safety as substrate (Kermarrec). Troncoso's propaganda-mitigation work is moderation, not hate-speech detection. Dedicated hate-speech work is thinner than M+D. Miranda Wei arrives in November 2026 and will work on online harms and non-consensual intimate images, which is in the H neighbourhood; this is incoming potential capacity, not present capacity.
   SourceTab Mapping EPFLInterviews West, Frossard, Ebrahimi, Kermarrec, Troncoso
5. Audio is the blind spot among data types. The EPFL MDH work mapped here runs almost entirely on text and images. Audio, including voice cloning and audio deepfakes, is barely covered. The capacity to close that gap sits nearby: EPFL works closely with Idiap (Gatica-Perez's group spans both), and Idiap has deep speech and audio expertise an EPFL MDH effort could draw on rather than build from scratch.
   SourceTab Mapping EPFLEPFL/Idiap Gatica-Perez
6. Detection-first projects rapidly obsolete. Süsstrunk's deepfake-detection work was based on frequency-domain artifacts of Stable Diffusion; modern transformer-based generators have different signatures, the work is no longer relevant. The same risk applies to any new EPFL detection project that does not also engage the structural questions (trust indicators, model raising, robustness, governance).
   SourceInterview Süsstrunk
7. No dedicated MDH curriculum or PhD pipeline. Students enter the field one supervisor at a time. West notes computational social science already exists as an academic field for this work but EPFL has no curriculum that brands MDH as a thing students can choose. The channels exist (hackathons, the 600-student data science class, ML for Science, ADA), see the next section, but no programme weaves them together.
   SourceInterview WestInsights Student Engagement is Feasible
8. EPFL is not yet visible in the EU operational layer. EU DisinfoLab's four hubs (AI, Climate, Conflict & Crisis, Doppelganger) curate research from external academia (Stanford HAI, Amsterdam, Cambridge and similar)^[101] but EPFL is not among the institutions whose work they routinely surface. EDMO coordinates EU fact-checking, not detection methodology^[174]. The gap is symmetrical: peers do not include EPFL; EPFL does not engage them.
   SourceRefs 99, 102
9. No coordination across Swiss universities on MDH. ETHZ (Center for Security Studies, International Conflict Research, ZHET)^{[212,180,181]}, UZH (Digital Society Initiative, Digital Democracy Lab, Digital Publics)^[179], Idiap (Social Computing group) and EPFL each have substantive MDH research lines, but no joint programme, no shared dataset infrastructure, no co-led Swiss conference, and no single Swiss-academic position addressed to BAKOM or the Federal Council. UZH's Digital Democracy Lab already collaborates with ETHZ on individual projects; EPFL is not yet part of that fabric. Karsten Donnay's group at UZH explicitly works on Swiss democracy, the same problem space EPFL/RTS work on the 2027 federal elections is approaching from another angle, and the two have not converged.
   SourceRefs 105, 106, 107, 108
10. Half-measures may be worse than nothing. Ebrahimi's blunt warning: badly-funded gestures crowd out real work. If EPFL commits to MDH it has to commit at programme scale, not via a string of two-year projects.
    SourceInterview Ebrahimi

Opportunities - what is open for EPFL

1. Geneva and Swiss standardisation presence. Ebrahimi chairs the JPEG committee and is vice-chair of the ITU/ISO/IEC AI & Multimedia Authenticity Standards Cooperation. EPFL is in the room where international MDH standards are written; few universities have this.
   SourceInterview Ebrahimi
2. ICRC and humanitarian dialogue via PeaceTech. EssentialTech has documented dialogue with the ICRC on information manipulation as a humanitarian concern, a bridge to practice that purely academic MDH groups do not have. Adjacent comparison: the Sentinel Project's WikiRumours model (community-led verification, 27000+ subscribers in DRC and South Sudan, 2M+ reach 2020-2022)^[103] shows what humanitarian-led MDH practice looks like.
   SourceEPFL EssentialTechRef 106
3. Active funding ecosystem. EHA (MDH explicitly one of 5 themes, ICRC + State of Zurich co-funded), Geneva AI Initiative via SDSC (~20 active proposals), CROSS Fund (low-friction EPFL-UNIL), HAC (Al-Hassanieh on committee), ICAIN. Funding is structurally accessible if a coherent EPFL position attracts it.
   SourceTab Funding
4. Student channels for engagement, currently under-used for MDH. PeaceTech Hackathon (second edition, 26-27 September 2026); the 600-student data science course where West routinely gives 5-minute pitch slots; ML for Science course (students collaborate with labs); ADA course (students propose projects); the Deepfake Awareness Booth originally from Süsstrunk's lab, now at Mediacom.
   SourceInterview WestInsights Student Engagement is Feasible
5. Incoming hires that align with the agenda. Miranda Wei in November 2026 (online harms, non-consensual intimate images), an incoming ethics professor (Süsstrunk: the previous absence was "un scandale"; she was preparing to push on this hire from her role as Dean). Both line up with the MDH agenda if leadership chooses to connect them.
   SourceInterview Süsstrunk
6. Three internal mappings ready to mutualise. This project, the AI Center mapping, and the C4DT mapping. A single coordinated EPFL position is more valuable than three parallel partial views.
   SourceInterviews Frossard and others
7. A compulsory technology-ethics course in the new AI Bachelor. Al-Hassanieh is co-designing a mandatory ethics course for EPFL's AI Bachelor, with case studies such as Community Notes and the limits of automated moderation. If it goes ahead it would reach every AI bachelor student, making it one of EPFL's most scalable MDH-awareness instruments. (Existence and the "first compulsory ethics course" framing are interview-sourced and to be confirmed.)
   SourceInterview Al-Hassanieh
8. Switzerland sits outside the EU's disinformation-observatory network. The EU's European Digital Media Observatory (EDMO) runs 15 hubs covering all 27 Member States plus Moldova, Norway and Ukraine, yet Switzerland is not among them (based on EDMO's published hub coverage as of June 2026), and no Swiss-wide map of MDH research capacity exists. This is the clearest external opening for EPFL: a credible, multilingual, Geneva-anchored complement to EDMO that the EU network structurally does not reach.^[174]
   SourceBeyond EPFL Governance and state response > Bodies (EDMO)Ref 187
9. The global field is contracting, not crowding. Several flagship observatories have closed or scaled back since 2022: Stanford's Internet Observatory was dismantled in 2024 (residual work folded into the Cyber Policy Center), Oxford's Programme on Democracy & Technology has been dormant since 2023, First Draft closed in 2022, and the Election Integrity Partnership wound down, amid heavy US defunding (the NSF terminated 400+ misinformation grants in 2025). The timing favours a stable, well-positioned Swiss entrant rather than a crowded market.
   SourceReview observatory status audit (June 2026)

Recommendation Guideline - for EPFL leadership

Headline recommendation

What that programme should and should not be

Funding and structural moves

• Doctoral and post-doctoral funding tied to specific open problems (see Research Directions tab when populated). Co-supervised with NGO or federal-agency partners where the topic supports it (ICRC, BAKOM, RTS).
• Shared engineering staff across participating labs to build and maintain datasets, benchmarks, and reproducibility infrastructure that no single lab can sustain.
• PhD-evaluation accommodation for interdisciplinary MDH work, so students are not penalised on publication count alone. Ebrahimi explicitly flagged this as a hard structural barrier.
• Push the ethics professorship hire to completion, then connect it to the MDH programme. The previous absence of an ethics chair was "un scandale" (Süsstrunk).
• Mutualise the three EPFL mappings (EssentialTech, AI Center, C4DT). A coordinated EPFL position that all three centres can cite is more valuable than three parallel partial views.
• Open a structured Swiss-university dialogue. Concrete first move: a half-day or one-day Swiss MDH research workshop convening EPFL, ETHZ (CSS, ICR, ZHET)^{[212,180,181]}, UZH (DSI, Digital Democracy Lab)^[179], Idiap, and partner institutions (SDSC, UNIL). Three distinct centres of gravity already exist (international-security at ETHZ CSS, digital-society and computational social science at UZH, technical research and PeaceTech at EPFL); they are complementary and the Swiss policy audience (BAKOM, Federal Council) currently receives no joint input from them.

Where peers fill a gap and where we fill theirs (context for the above)

Three overlaps to manage and four differences to lean into, with the closest peer organisation, EU DisinfoLab:

Overlaps (manage to avoid duplicating):

1. Aggregating research. Their "in-depth" sections and our Bibliography, MDH Context and Beyond EPFL tabs both catalogue external research^[101].
2. Policy tracking. Their AI Disinfo Hub maintains 60+ policy updates; our Beyond EPFL tab tracks the DSA, both Codes of Conduct, the AI Act and Federal Council Report 22.3006^[99,98,7].
3. Thematic organisation. Their four hubs and our M / D / H + AI x MDH + Switzerland x MDH breakdown both organise by theme.

Differences (lean into):

1. A research community grounded in needs, not a hub of finished outputs. EU DisinfoLab's four hubs are curated repositories of research produced elsewhere^[101]; once a paper exists, it can be linked. A university observatory at EPFL inverts this: the artefact is the community of researchers, PhDs, postdocs, engineers and partners assembled around live research questions, and the papers are downstream of that community. Research questions come from real needs (an ICRC operational problem, a Swiss federal policy gap, a deepfake-detection benchmark a partner wants), are matched to people on the researcher pages, and become PhD topics, master theses, hackathon tracks or short-term engineering projects. This is structurally what a Swiss technical university can do and an operational NGO cannot.
2. Upstream production, not downstream aggregation. EPFL produces peer-reviewed datasets, benchmarks, detection methods and robustness analyses. EU DisinfoLab's "in-depth" sections explicitly describe themselves as repositories of research from academia and civil society^[101]. The observatory's primary output is the artefacts, and the searchable index of who at EPFL can build, evaluate or audit a response.
3. Hate speech co-equal. Based on a review of their four hubs as of 2026-05-18, none of EU DisinfoLab's hubs centre hate speech as a primary axis^[101]. Ours does. Keeping H structurally co-equal with M and D is one of the most defensible boundaries between this project and disinformation-focused peers.
4. Swiss / multilingual / Geneva as geographic and linguistic complement. Be the Switzerland and low-resource-language partner to EU DisinfoLab and EDMO, contributing what they cannot easily produce. Multilingual hate-speech detection is a documented weak spot at LLM scale^[23]; Swiss federal-context dynamics (no DSA equivalent, four official languages, direct democracy) have no dedicated observatory^[7].
5. PeaceTech / humanitarian bridge. Based on a review of EU DisinfoLab's public hubs as of 2026-05-18, their Conflict & Crisis hub references humanitarian work tangentially (pieces from ICRC, UNRWA / Gaza, Centre for Civilians in Conflict, IHL information-operations material) but humanitarian operational MDH and community-led verification partnerships are not stated focuses^[101]. EssentialTech's PeaceTech mission and ICRC dialogue, and the locally-led model documented by Sentinel Project / CHIC^[103], occupy operational and partnership space EU DisinfoLab does not.

Why doing this at a Swiss university is structurally strong (and different): small geography keeps cross-institution coordination tractable (EPFL, ETHZ, UZH, Idiap, SDSC, UNIL all sit within a few hours of one another); multilingual federal communications make multilingual research a first-class need rather than a niche; the Geneva ecosystem (ICRC, IFRC, UN, WHO, ITU, ISO, IEEE) is unique in the world for operational and standards-setting reach; Switzerland's regulatory lightness^[7] means university-led inquiry can shape policy formulation rather than respond to fait accompli; and Swiss federal funding instruments (SNSF, Innosuisse, CROSS, HAC, ICAIN, EHA) plus EU consortium access provide a denser funding fabric than most academic ecosystems. None of these advantages exists in the same combination at peer European or US institutions.

Concrete near-term step

A biennial "Swiss MDH State of Play" report, jointly authored across EPFL, ETHZ (CSS, ICR, ZHET)^{[212,180,181]}, UZH (DSI, Digital Democracy Lab)^[179], Idiap, SDSC, UNIL. The report's value is that it is the first co-signed Swiss-academic statement on MDH; BAKOM, the Federal Council, the AI Center, C4DT, EssentialTech and IFRC partners can cite a single coordinated source rather than collect five university positions. Biennial cadence is sustainable; the website becomes the searchable index between editions; the researcher pages become the directory of who could contribute to the next one. The report, not a hub, is the artefact, this is what differentiates the project structurally from EU DisinfoLab while remaining additive to them.