New Multimodal Fact-Checking Dataset Spans 10 Languages and 22 Countries

A team of researchers has unveiled M4FC, an ambitious dataset designed to improve how AI systems tackle misinformation in images and text across different languages and cultural contexts.

What's in the dataset#

The collection brings together 4,982 images paired with 6,980 claims, all verified by professional fact-checkers from 22 organizations working across diverse geographic regions. Each claim appears in one or two of ten supported languages, making this one of the most linguistically and culturally representative fact-checking resources to date. The images themselves capture a wide spectrum of real-world contexts rather than focusing on a single region or narrative.

Six complementary fact-checking tasks#

Rather than addressing just one aspect of misinformation, M4FC covers six distinct but related challenges:

• Visual claim extraction — identifying what assertions are being made in or about an image
• Claimant intent prediction — determining the motivation behind a claim
• Fake image detection — spotting manipulated or AI-generated imagery
• Image contextualization — placing images in their proper setting and timeline
• Location verification — checking geographic claims
• Verdict prediction — the ultimate task of assessing whether a claim is true, false, or misleading

Bridging a research gap#

Existing multimodal fact-checking datasets have struggled with scale, language diversity, or breadth of scope. Some contain only hundreds of examples, others cover just one or two languages, and many zero in on a single task. M4FC addresses these limitations by combining size with multitask breadth and genuine cross-cultural representation.

How intermediate tasks improve accuracy#

The researchers tested whether combining results from earlier tasks — such as image contextualization or location verification — could strengthen the final verdict prediction. Their analysis reveals that layering these intermediate steps can meaningfully boost performance, suggesting that breaking down fact-checking into component tasks yields better outcomes than tackling the problem in isolation.