The Royal Society Talk on Misleading Information, Fact Checking, and AI Resilience

Overview

This Royal Society talk examines real world misinformation and how professionals verify claims across multiple modalities. It highlights how large language models (LLMs) interact with evidence, how visual content can mislead, and how charts can distort data, all within a framework aimed at building resilience against false information in both humans and machines.

What you will learn

Key ideas include the limitations of counter-evidence when a claim is new, the fallacy-based approach to debunking misrepresented science, and the need for robust evidence retrieval and source evaluation. The talk also discusses automated benchmarks for text, image, and chart verification and concludes with an ethical reflection on truth, belief, and critical thinking.

Introduction

The speaker opens by framing misinformation as a real world problem dividing information across text, images, and charts. The goal is resilience against false content and responsible AI that does not amplify misleading claims.

Text based misinformation and fallacy detection

The talk uses a misrepresented COVID-19 claim as a running example. It explains how professional fact checkers build context around a claim, examine the provenance of the source article, and evaluate whether premises actually support conclusions. A new computational formalism is proposed to detect fallacies when a claim is based on a cited study. In experiments, LLMs are prompted with fallacy definitions and contexts, showing high performance when premises are provided, but struggle as premises must be generated. A biased effect is observed when evidence is included in prompts, causing models to be biased toward the claim.

Images and contextualization

Turning to images, the talk discusses how a picture paired with a caption can mislead about location and event. A five-question framework guides automated image contextualization: determine the visual claim, identify the claimant's intent, verify image originality, retrieve evidence, and check caption alignment with image. A data set and method that leverages LLMs to extract answers from human fact-check articles is introduced, showing varying performance across source identification, time stamping, and geolocation.

Misleading charts and defenses

The final portion covers misleading charts. Definitions of misleaders and strategies to surface chart distortions are given, along with a plan to reduce vulnerability of multimodal models through inference-time corrections. Two data sets, MisVs real-world charts and a synthetic set, underpin experiments comparing zero-shot LLMs, rule-based linters, and neural classifiers. Results reveal vulnerability of LLMs to misleaders, while two corrective strategies show promise: converting charts to tables for question answering, and redrawing charts from correct tables. The speaker emphasizes that these corrections should not degrade performance on honest charts due to potential noise in the conversion process.

Concluding reflections and future directions

In closing, the talk stresses continuous active learning, the need for credible cross-domain content, and the risk of deception via manipulated documents. The speaker invites action to cultivate critical thinking, cross-cultural perspective, and robust, trust-worthy AI systems that help users navigate a dynamic information landscape.