Using Machine Learning to Automate Misinformation Detection on Social Media

In under five minutes, anyone can create an account, record a video and post content to millions of people. There is no mandatory verification, no gatekeeping, no editorial oversight...And yet the consequences can be global.

‍

From vaccine misinformation to political campaigns, social media has demonstrated how quickly false claims can spread (often faster and further than corrections). Research shows that when people feel uncertain or fearful, they are more likely to believe false information (Wu et al., 2019). Meanwhile, platform algorithms amplify content that generates engagement and misinformation tends to do exactly that

‍

So the question becomes: Can machine learning help us detect misinformation at scale?

‍

Why Misinformation Spreads So Effectively

‍

Misinformation isn’t just incorrect information, it is information designed to persuade.

‍

Three structural forces make misinformation particularly powerful:

1. Scale: Anyone can post anything without verification.
2. Speed: Videos and posts can go viral within hours.
3. Impact: Even when misinformation is later retracted, belief often persists. Vaccine hesitancy, for example, remains recognised as a major global health threat

‍

Historical examples highlight the stakes. During the 2016 EU referendum, misinformation campaigns were widely discussed and the UK government later compelled Facebook to hand over advertising data.

‍

What Does Misinformed Content Look Like?

‍

One of the first challenges is definitional: how do we even detect misinformation automatically? Rather than beginning with fact-checking, we began by analysing linguistic patterns.

‍

Credibility Signals

• Referencing professionals, celebrities, or “trusted sources”
• “As a medical researcher…”

Structure

• Storytelling formats
• Strong opinionated framing
• Persuasive rhetorical construction

Language

• Personal pronouns
• Fear-inducing terms
• Emotional amplification

‍

The psychological component is critical. Misinformation often exploits emotional states rather than simply presenting incorrect facts.

‍

The Core ML Challenge: Claim vs Opinion

‍

One practical step toward automation is distinguishing claims from opinions.

‍

Consider:

‍

“As a medical researcher, I can tell you this supplement cures chronic fatigue.”
(Claim)

‍

vs.

‍

“As a medical researcher, I believe this supplement may help with chronic fatigue.”
(Opinion)

‍

If a model can reliably separate factual claims from personal opinion, it can prioritise verification pipelines more effectively.

‍

The Models Tested

To explore this, multiple classifiers were trained and compared

‍

• Random Forest
• Support Vector Machine (SVM)
• Logistic Regression
• Transformer-based models

‍

The goal was to validate claim–opinion separability using:

‍

• Psycholinguistic features
• Sentiment
• Metadata (views, engagement, downloads, shares)

‍

The Result: 99.9% Accuracy… Almost

‍

Initial results appeared extraordinary. All three classifiers performed extremely well, with Random Forest achieving the highest performance and misclassifying only a handful of examples. However, when results look too good to be true, they often are.

‍

The Hidden Problem: Metadata Bias

‍

A closer inspection revealed something important:

‍

The models were heavily relying on metadata i.e. views, engagement, downloads rather than linguistic features

‍

In other words:

• High-performing videos were being classified as claims
• Engagement patterns were acting as shortcuts
• Linguistic nuance was largely ignored

‍

This is problematic because:

• Views can be manipulated or purchased
• Viral content is trend-dependent
• We don’t want misinformation to go viral before detection
• Engagement does not equal truth

When metadata was removed or controlled for, accuracy remained high (~99%), but performance became sensitive to dataset structure and distribution bias

‍

This exposed a fundamental issue: Generalising misinformation detection beyond controlled datasets is extremely difficult.

‍

The State of the Art (Pre-LLMs)

‍

By 2022–2023, misinformation detection pipelines had clear limitations

‍

• Required expensive, manually labelled datasets
• Models were narrow and task-specific
• Every new task required new fine-tuning
• Detection often reduced to sentiment or keyword correlation
• Heavy reliance on external fact-checking websites
• Difficulty defining what even constitutes a “fact”

Misinformation is not just incorrect data, it is a social phenomenon. Traditional classifiers struggle with that complexity.

‍

The Silver Lining: Enter GPT-4

‍

With the release of GPT-4, a new question emerged: Can LLMs simultaneously detect misinformation and verify claims in context?

‍

Unlike traditional models, large language models:

• Can reason across context
• Can access structured external information
• Can distinguish nuance between opinion, speculation, and factual assertion
• Can evaluate claims dynamically rather than relying purely on static features

‍

This doesn’t solve misinformation overnight. But it changes the architecture of possibility.

‍

Instead of:

• Classify → Flag → Send to fact-checker

We can now explore:

• Detect → Extract claim → Cross-check → Explain reasoning

In one integrated pipeline.

‍

Where This Research Leaves Us

‍

The takeaway is not that we have solved misinformation. Even with LLMs, misinformation remains a profoundly difficult problem.

‍

This isn't just a modelling challenge, It is an ecosystem challenge. Social media platforms optimise for engagement, algorithms amplify emotionally charged content and confirmation bias nudges us toward information that reinforces what we already believe. Even when corrections are issued, belief persistence is well documented.

‍

An AI system can flag content, It can extract claims... it can even cross-reference sources.

‍

But it cannot single-handedly fix:

• Incentive structures that reward virality
• Algorithmic filtering that reinforces echo chambers
• Human psychology that favours emotionally resonant narratives
• The ambiguity of what counts as a “fact” in contested domains

‍

If anything, this research reinforces a broader conclusion:

‍

We don’t just need better models. We need better public understanding of how algorithms shape what we see.

‍

Misinformation is ultimately a socio-technical problem... one that sits at the intersection of technology,  behaviour and education.

‍

‍