Synthetic Realities at Digital Footprints 2026: tracking AI-generated vaccine misinformation before it spreads

In May I took the Synthetic Realities poster to the Digital Footprints Conference 2026 in York, the fourth edition of an event that brings academics, policymakers, government and industry together around one question: what is digital data actually good for, and who does it serve?

This year the conference ran across three days and expanded its remit. Day one was delivered in partnership with Smart Data Research UK, showcasing how smart data is driving work across health and wellbeing, digital society, sustainability, productivity and prosperity. Days two and three carried the academic programme run by the International Society for the Study of Digital Footprints. 

View fullsize

Caption: A4 research poster titled Synthetic Realities, setting out a 15 month study of AI-generated visual vaccine misinformation, including the research questions, three phase methodology and data sources.

Presenting a poster is a particular kind of exercise. You cannot hide behind a slide deck or a build order. Everything has to be legible from a metre away, and every claim has to survive somebody standing in front of it with a coffee and a raised eyebrow. It is also, quietly, one of the best forms of peer review available: a dozen unscripted conversations in an afternoon, each one speaking to a different question or viewpoint.

Here is what the poster argued.

The Hydra Effect

Anti-vaccine content used to work by recycling. The same debunked claims, the same handful of images, circulating again and again. That was a problem fact-checkers could get their arms around, because you can only debunk the same picture so many times before the debunk itself is searchable.

Generative AI has removed that ceiling. It now produces novel visual "evidence" faster than anyone can respond to it, and it spawns variants tailored to different audiences, different platforms and different identities. Cut off one head and two more appear, each subtly reshaped for whoever is looking at it. That is the Hydra Effect, and it is the reason this fellowship exists.

Why a new method

The field currently splits the problem in two. One community treats AI-generated misinformation as a technical detection problem: build a better classifier, flag the fake, move on. Another treats it as a social discourse problem: study the narratives, the communities, the trust dynamics. Both are right about something. Neither, on its own, is enough. Nobody is systematically integrating the two.

Synthetic Discourse Analysis is my attempt at that integration. It combines quantitative network analysis, qualitative discourse analysis and community-participatory methods, and it treats AI-generated visual misinformation as what it actually is: a computational phenomenon and a social process at the same time.

The 15 month feasibility study runs from February 2026 to April 2027 in three phases:

Phase 1: Baseline Mapping (months 1 to 5)

A detection framework combining Pulsar TRAC, Deepfake-o-meter and OCR. Community workshops in Manchester and London. Geographic clustering mapped against vaccination uptake.

Phase 2: Deep Discourse (months 6 to 10)

Thematic network analysis using Caplena, InfraNodus and Celonis. Multigenerational community workshops. Temporal impact tracking and community response mapping.

Phase 3: Interventions (months 11 to 15)

Visual literacy tool development extending the VAIL toolkit. Detection training sessions. Platform-specific interventions co-designed with affected communities.

The three questions on the poster

1. How does AI-generated visual misinformation integrate with and transform human-generated discourse across platforms?

2. What discursive patterns distinguish AI-generated from human-generated content in digital public spheres?

3. How do platform affordances influence the spread of AI-generated vaccine misinformation, and which populations are most vulnerable?

The early insight, and the one that drew the most conversation at the poster boards, is this:

Misinformation and counter-messaging operate on fundamentally different persuasive registers. Fact-checking addresses information deficits. Misinformation exploits trust and identity.

If that is right, then a great deal of well-funded, well-intentioned counter-messaging is answering a question nobody asked.

Meet the Vaccine Digital Twin

The poster included a proof-of-concept simulation environment. You can run it below.

Its personas are hypothetical composites rather than real people, with their language and stances calibrated against paraphrased, aggregated data. It reproduces no individual-level platform data.

It lets you watch a single piece of AI-generated visual misinformation cascade through a feed:

1. An AI-generated image surfaces in the feed.

2. It amplifies through trust networks.

3. Wellness framing launders the claim into acceptable language.

4. An NHS professional intervenes with an accurate correction.

5. The “Kelly” persona remains hesitant.

Step five is the main point. The intervention arrives, it is accurate, and it does not land.

This is the resonance gap. The misinformation reached the “Kelly” persona as an image, in a register built to move her maternal instinct, belonging, institutional distrust. The correction reached her as text, in a register built to inform her. Kelly does not need more data. She needs evidence that arrives the way the misinformation did.

Fact-checking addresses knowledge deficits, but visual misinformation exploits trust and identity. Kelly does not need more data. She needs a different kind of evidence.

Fullscreen

Three data streams, one picture

Three sources, each answering a different question.

• What is being said. Pulsar TRAC gives real-time monitoring across X, Facebook, Instagram, BlueSky, TikTok and Reddit, using Boolean search with hashtag community mapping, coded language detection and emoji co-occurrence analysis.

• Where it maps. GeoDS, the UKRI-funded Geographic Data Service run by UCL and Liverpool, supplies Modelled Ethnicity Proportions and SmartStreetSensor footfall at LSOA level. Cross-referenced with Pulsar, it maps hotspots against demographic patterns and vaccination uptake.

• Who is engaging. SDDS MOSAIC, held at the University of York, provides YouTube viewing histories, app usage and linked survey data from around 10,000 UK Android users. It is accessed through the SafePod Trusted Research Environment, and it is what makes individual consumption pathway analysis possible.

Separately, each stream tells you something partial. Together they let you follow a claim from the moment it is generated to the moment it changes what somebody does about a vaccination appointment.

From research to action

• Policy: report for the UKRI Data Access Taskforce and Ofcom, plus a parliamentary inquiry submission on online safety and AI governance.

• Academic: two to three journal submissions establishing the Synthetic Discourse Analysis framework.

• Community: six workshops across all phases, a digital literacy toolkit via UKHSA, and train-the-trainer sessions for 15 to 20 community moderators.

• Industry: a white paper on API requirements for synthetic content detection, testing with Pulsar TRAC, and roundtables with the Alan Turing Institute.

Get in touch

Dr Sam Martin Smart Data Research UK Fellow, Manchester Metropolitan University (Digital Society Research Group) Senior Research Fellow, Vaccine and Society Unit and Pandemic Sciences Institute, University of Oxford.

DOWNLOAD

Poster: Download the full A4 poster (PDF)

Synthetic Realities is funded by UKRI Smart Data Research UK and hosted by the MMU Digital Society Research Group. Mentored by Dr Adi Kuntsman and Professor Keeley Crockett. Built on the LISTEN methodology, the VAIL framework and Oxford Vaccine Group research.