Detecting, Measuring, and Countering Disinformation with Data and AI

The five topics connect into a pipeline, and each can also be taken on its own.

1. Multimodal detection identifies manipulated content. Temporal knowledge graphs and neuro-symbolic retrieval.
2. Ground and verify claims against evidence. Robustness evaluation and red-teaming.
3. Test how far the underlying language models can be trusted. Multi-agent simulation and early-warning.
4. Model how narratives spread and when to intervene. Agentic monitoring of non-human actants.
5. Watches the shared knowledge sources (LLM outputs and collaborative encyclopaedias) that feed the rest. Two or more topics can also be combined (for example detection with verification, or simulation with monitoring) into a larger system for understanding and countering disinformation.

1. Multimodal Detection of AI-Generated Disinformation

Generative models now produce convincing manipulated text, images, audio, and video at low cost. Most detection tools work only on text and in a single language, which leaves multimodal and model-generated manipulation largely uncovered.

This thesis will develop a multimodal detection pipeline that combines foundation models across text, image, audio, and video, and grounds their outputs against a temporal knowledge graph used as external memory, so that verdicts are traceable rather than opaque scores. A specific focus is generalisation: the system is evaluated on climate content and on transfer to other manipulation domains, such as political deepfakes, health misinformation, and influence operations. The outcome is a prototype multimodal detection service, integrated with a multi-agent architecture, together with a measured assessment of cross-modal and cross-domain robustness.

Work to be done

Literature Review

• Survey multimodal machine learning and foundation models for text, image, audio, and video.
• Review deepfake and synthetic-media detection, media forensics, and provenance/watermarking standards (e.g. C2PA).
• Study narrative characterization and manipulation-technique taxonomies (e.g. DISARM) and multilingual detection challenges.

Data and Requirements Analysis

• Assemble or curate a multilingual, multimodal corpus of authentic and manipulated content, with a manipulation-technique labelling scheme.
• Define detection requirements and quality criteria for citizen, journalist, and policymaker contexts.
• Specify metadata alignment (e.g. DISARM) for interoperability across the pipeline.

Model Design

• Design cross-modal fusion strategies combining per-modality foundation-model features.
• Incorporate provenance and synthetic-origin signals into the decision.
• Ground verdicts against a temporal knowledge graph used as external memory, to reduce hallucination and support explanations.

Prototype Development

• Implement the detection pipeline as one or more agents within a multi-agent architecture.
• Expose explainable outputs (evidence, contributing modalities, provenance) through a neuro-symbolic layer.

Evaluation

• Measure precision, recall, and F1 per modality and for the fused system.
• Expert feedback through pilot deployment with practitioners (i.e. journalist, fact-checkers, etc.)
• Assess cross-domain generalisation by transferring from climate to political, health, and defence-related synthetic media.

Analysis

• Compare against text-only and single-modality baselines.
• Analyse failure modes, multilingual gaps, and the limits of cross-domain transfer.

2. Temporal Knowledge Graph Construction and Neuro-Symbolic Retrieval for Grounded Claim Verification

LLMs generate fluent text but hallucinate, and automated claim verification is often shallow and easy to defeat by cherry-picking. Knowledge-graph and linked-data methods that are established in biomedical work are rarely applied to disinformation, even though they can preserve the context of a claim: who said it, where, under which framing, and against which evidence.

This thesis will develop an agentic pipeline for building and querying a temporal knowledge graph, and a GraphRAG-style neuro-symbolic layer that grounds LLM reasoning in that graph. The work covers entity extraction and linking, ontology alignment, provenance modelling, and temporal inference, combined with retrieval- and tool-augmented reasoning (e.g. ReAct-style agents using protocols such as MCP). A cross-domain objective is to show that the ontology and pipeline transfer from climate to other controversy domains, such as politics and medicine. The outcome is a prototype verification environment that combines vector retrieval, graph reasoning, and provenance to produce grounded, explainable claim assessments.

Work to be done

Literature Review

• Survey knowledge graphs, temporal knowledge graphs, and graph databases.
• Review entity linking, ontology alignment, and provenance modelling.
• Study GraphRAG and retrieval-augmented reasoning, and claim-verification pipelines (e.g. FEVER-style evidence retrieval).
• Study the design philosophies and evolutions of DISARM and STIX frameworks.

Domain Modeling

• Identify core entities and relationships, such as Claim, Actor, Narrative, Evidence, Scientific Baseline, Policy, Framing, and Source.
• Design an ontology and temporal graph model that preserves claim context and provenance over time.
• Extend the graph with domain concepts using LLM-based entity extraction, linking, and relationship extraction.

Knowledge Graph Construction

• Build agentic pipelines that extract structured knowledge from news media, social platforms, and fact-checking sources.
• Model temporal edges and versioning to support temporal inference and dynamics-aware memory.
• Apply human-in-the-loop validation to ensure accuracy and trustworthiness.

Neuro-Symbolic Verification Layer

• Combine vector retrieval, metadata search, and graph traversal with LLM reasoning.
• Use the temporal knowledge graph as external memory for context-aware, hallucination-resistant verification.
• Generate explanations as graph paths and provenance chains.

Cross-Domain Transfer

• Adapt the ontology and pipeline to a second domain (e.g. political or medical controversies).
• Assess reuse of analytical templates across domains.

Evaluation

• Compare metadata search, vector search, hybrid search, and GraphRAG verification.
• Measure verification accuracy, hallucination rate, temporal-reasoning quality, and explainability.
• Evaluate transferability and the effort required to move to a new domain.

3. Adversarial Robustness and Red-Teaming of LLMs against Disinformation

LLMs are both a defence tool and an attack surface: they can generate disinformation at scale, and they are themselves vulnerable to data poisoning, bias, and alignment drift, including through shared sources such as Wikipedia. Closed evaluation pipelines make results hard to reproduce or reuse.

This thesis will build an open, reproducible robustness benchmark that combines adversarial red-teaming, counterfactual and persona-based testing, and data-poisoning stress tests, and measures robustness, bias susceptibility, and alignment drift across several models. Evaluation covers three social contexts (citizens, journalists, and policymakers), with climate as the primary testbed and health, elections, and influence operations as transfer domains. The outcome is an open evaluation harness with reusable attack suites and an empirical study of model robustness.

Work to be done

Literature Review

• Review alignment methods (e.g. Constitutional AI), holistic LLM evaluation, truthfulness benchmarks, and risk taxonomies.
• Study adversarial red-teaming, persona-based evaluation, and jailbreak/prompt-injection techniques.
• Survey data- and retrieval-poisoning attacks, including via shared knowledge sources.

Threat Modeling and Requirements

• Define attack vectors, personas, and social contexts (citizen, journalist, policymaker).
• Specify robustness, bias, and alignment-drift metrics and reporting formats.
• Define reproducibility and openness requirements for the benchmark.

Framework Design

• Design controlled prompting, counterfactual tests, and adversarial scenario generation.
• Develop automated red-team agents that generate and adapt attacks.
• Integrate poisoning experiments targeting retrieval and shared-knowledge inputs.

Prototype Development

• Implement an open evaluation harness with reusable attack suites and a report/leaderboard.
• Support evaluation across several open and, where feasible, proprietary models.

Evaluation

• Measure robustness, bias susceptibility, and alignment drift across models and contexts.
• Quantify the effect of poisoning and adversarial inputs on truthfulness.
• Repeat the evaluation in a second domain (e.g. health or politics) to test transferability.

Analysis

• Assess reproducibility and the transferability of attacks across models and domains.
• Derive mitigation recommendations for deploying LLMs in disinformation-sensitive settings.

4. Multi-Agent Simulation of Narrative Diffusion and Early-Warning Signal Detection

Common disinformation metrics treat virality as harm, and monitoring is mostly reactive. Modelling how narratives spread makes it possible to signal risk before a narrative reaches wide visibility, so that countermeasures such as prebunks can be timed rather than issued after the fact.

This thesis will build a simulation and early-warning system that models narrative propagation across platforms and emits a signal as tipping points approach. It combines network-diffusion and complex-contagion models with hybrid agents (rule-based agents extended with LLM-driven generative agents that hold memory and act strategically), and supports counterfactual tests of prebunk timing, booster reminders, and targeted messaging. Alongside climate, the approach applies to political campaigns, health misinformation, and coordinated influence operations. The outcome is a prototype simulation environment and a validated early-warning component.

Work to be done

Literature Review

• Study epidemic and network-diffusion models, and complex-contagion frameworks (homophily, echo chambers, reinforcement).
• Review agent-based modelling, LLM-driven generative agents, and demographically conditioned agent samples.
• Survey cross-platform narrative laundering and algorithmic-personalisation effects.

Requirements and Data

• Assemble multi-platform narrative datasets and define audience segments and actor types.
• Define early-warning objectives, lead-time targets, and evaluation criteria.

Model Design

• Design a hybrid agent population combining rule-based and LLM-driven generative agents.
• Model network structure, algorithmic personalisation, and cross-platform spread.
• Incorporate heterogeneous agent reasoning (e.g. theory-of-mind-style social reasoning).

Early-Warning Pipeline

• Define signal features and tipping-point detection over simulated and real narrative streams.
• Calibrate thresholds and connect outputs to the intervention/prebunking layer.

Intervention Simulation

• Test prebunk timing, booster reminders, and targeted communication in simulation.
• Explore counterfactual scenarios under different socio-political conditions.

Evaluation

• Measure predictive accuracy, lead time, and calibration of the early-warning signal.
• Assess simulated intervention effectiveness against baselines.
• Evaluate transfer of the models across domains (climate, politics, health).

5. Agentic Monitoring of Non-Human Actants: LLM Poisoning and Collaborative-Knowledge Manipulation

Most tooling watches human posts and ignores two forces that increasingly shape shared reference knowledge: LLM outputs and the editorial dynamics of collaborative sources such as Wikipedia. Because retrieval systems and model training draw on these substrates, manipulating them can propagate quietly into downstream AI and public understanding.

This thesis will design autonomous monitoring agents that track manipulation of shared knowledge substrates: detecting coordinated or anomalous edits in collaborative encyclopaedias, and probing LLM outputs for injected or drifting disinformation across prompts, versions, and models. Detected anomalies feed a temporal knowledge graph and an early-warning component with provenance-rich, explainable alerts. Climate is the anchor case; the same agents apply to public-health guidance, geopolitics, and science reporting. The outcome is a prototype monitoring framework with validated detection on known manipulation cases.

Work to be done

Literature Review

• Review vandalism, edit-war, and coordinated-behaviour detection in collaborative knowledge platforms.
• Study LLM data- and retrieval-poisoning, model auditing, and drift detection.
• Survey provenance modelling and anomaly detection on temporal streams.

Requirements Analysis

• Define the non-human actants, signals, and data sources (edit histories, model-probing outputs).
• Specify alerting requirements, provenance needs, and governance constraints (e.g. EU AI Act regulatory sandbox).
• Define computational constraints for different deployment scenarios (i.e. non-profit, governmental, etc.)

Agent Design

• Design specialised agents such as Source Discovery, Edit-Monitoring, LLM-Probing, Anomaly-Detection, and Provenance agents.
• Define coordination and human-in-the-loop escalation between agents.

Detection Methods

• Develop temporal anomaly and coordination detection over edit streams.
• Systematically probe LLM outputs across prompts, versions, and models to detect injected disinformation and drift.
• Leverage algorithmic approaches that can be operationalized at scale on large knowledge platforms (e.g. Wikipedia).

Integration

• Feed alerts into an early-warning component and enrich a temporal knowledge graph with provenance.
• Produce explainable alerts suitable for journalists and public authorities.

Evaluation

• Measure detection precision and recall on known manipulation and poisoning cases.
• Analyse false positives and detection timeliness.
• Assess generalisation to other topics and domains, and discuss governance and ethics.