Going viral

Fake news spreads like an epidemic. Here’s how to stop it.

More than five years after the outbreak of Covid-19, most of us still remember the shape of the pandemic. It started small and distant. Then it passed from one person to another quickly, and we often did not even know we were spreading it.

Today’s misinformation epidemic has the same shape: Fake news spreads quickly, and those “infected” by a rumour are often unaware it is untrue before passing it on to others.

The “pink salt trick”, a weight-loss method claiming the effectiveness of consuming Himalayan salt, was circulating on TikTok. It is not backed by science and is risky for those with certain health conditions. Another common claim — that chemical sunscreen is toxic — is untrue and stops people from protecting their skin against sun damage.

The spread of fake news is not unique to the 21st century. But the internet, social media and especially AI are turning it into a crisis.

The World Economic Forum says fake claims are one of the greatest threats the world faces today, alongside geoeconomic conflict, societal polarisation and climate change. The World Health Organization has called the spread of health-related misinformation an “infodemic”.

That is why researchers are keen to understand how fake news spreads and what can be done to stop it. To do so, they do not just compare misinformation to a virus, but actually treat it like one.

Using the same models built to study outbreaks like the Covid-19 pandemic, researchers can simulate the online spread of false claims.

The research done by Jun Zhuang, an associate dean for research at the University at Buffalo, informed the simulations in this story. “A person cannot learn a rumour unless they are exposed to it by someone else, making it a person-to-person contagion process,” he says.

When misinformation “goes viral”, it is not just a metaphor. It is what’s actually happening.

A computer can replicate these interactions over and over, turning them into a simulation. It is not reality — “These models have to simplify reality to make the maths work,” Zhuang says — but it helps researchers make sense of and fight epidemics of misinformation.

The most basic simulation is a runaway rumour. If it is interesting enough and the people sharing it are well connected, the false claim will spread widely.

With no debunkers — just the unexposed , spreaders and stiflers — nearly everyone in this simulated online community will encounter the misinformation.

In 2019, South Korea came close to a real-world version of this scenario. A YouTube channel featured a US man who claimed to have cured his cancer with fenbendazole — a drug used to control parasites in animals. Then a South Korean entertainer who heard about this said he was treating his own cancer with fenbendazole. The news spread rapidly among cancer patients in South Korea. Despite a lack of evidence that it can treat cancer, demand for fenbendazole spiked.

Eventually, debunkers entered the scene. After 20 days, the South Korean Ministry of Food and Drug Safety banned the use of fenbendazole to treat cancer. And after 54 days, the Korean Cancer Association made an official statement correcting the social media claims .

But here is the thing about debunkers: It really matters when they show up.

A delay is costly. For example, if a debunker is introduced a week after the rumour starts, 45 per cent of the population would have already believed it.

Even after the debunkers correct the record, 13 per cent would have seen the false claim but not the correction.

In South Korea, rumours about fenbendazole circulated for years. They started to fade only after the South Korean entertainer who spread them died of the cancer he was trying to treat.

It matters that so many people believed the misinformation, even if it was just for a time. Some concerned cancer patients acted on the claims before any correction reached them. Others developed liver injury after self-administering the drug. While there is no official count of South Koreans who died because they chose fenbendazole over recommended treatments, studies show that cancer patients who use unproven remedies over standard treatments have roughly twice the risk of death.

Zhuang’s research found that earlier corrections are always better. When a debunker appears much sooner — after one day instead of a week — the spread of misinformation dies quickly.

This kind of quick debunking happened in 2021, after an Indian politician tweeted that a new, dangerous Covid-19 variant had originated in Singapore. Within a day, Singapore’s Ministry of Health corrected the record. Other officials in both countries also challenged the misinformation online, as did mrbrown, a well-known Singaporean blogger. The fake news was quickly quashed, just like in a comparable simulation.

The speed of debunking was key in this case. But something else happened that was worth noticing: The correction came from a diversity of sources at once, which most likely sped up the debunking process.

When The Straits Times adapted Zhuang’s model to have debunkers deployed from various locations, rather than just one, it showed that scattered debunkers fought the rumour far more efficiently.

But speed and multiplicity of sources help debunking and the spread of truth as much as they help fake news — which explains how artificial intelligence can and is already exacerbating the misinformation epidemic.

“AI is turbocharging the ability to create more content, to spread content and to get that content in front of people,” says Benjamin Ang, head of the Centre of Excellence for National Security, Future Issues and Technology at the S. Rajaratnam School of International Studies.

As researchers work to understand how AI impacts the transmission of misinformation, simulations can also help make sense of this fast-moving picture.

Consider this: Some research shows that we are more likely to believe fake news created by AI. Does that mean we are more likely to share it? Researchers are not sure. But it is a reasonable bet that the more convincing a rumour is, the more people who see it will pass it on. So, using the model, we can explore this question: If AI-generated misinformation spreads faster because it is more convincing, what would that look like?

If AI-powered misinformation does spread more quickly, it is a big problem for debunking. Faster spreading requires faster debunking.

Here is another challenge: AI can weaponise the multiplicity and diversity of misinformation sources. A coordinated AI-generated launch can include dozens of accounts posting the same claim simultaneously, rather than one person and a single organic post. Look how quickly the misinformation spreads when it starts in multiple places.

Singapore has recently been the target of this kind of distributed campaign. In July, an investigation discovered a TikTok network of AI-generated female presenters posting videos that looked like real news updates. Ninety per cent of them contained false or misleading information. Account after account followed the same script, just tweaked slightly.

The people behind the accounts “don’t know which one will get traction”, Ang explains, “but it’s so easy to just throw them all out there and see which one sticks”.

So why not deploy this power for good, by creating AI debunkers? AI is good at misinformation, but not great at debunking. It is decent at detecting the simple cases – one study found that an AI model correctly flagged 90 per cent of false headlines as false. But its effectiveness is limited. AI debunking struggles with videos, doctored images or claims that depend on context. Since AI is not that good at debunking, we need to do it ourselves.

As with Covid-19, tackling misinformation requires interventions at all levels across society. Governments combat the spread through laws like Singapore’s Protection from Online Falsehoods and Manipulation Act, which lets officials move quickly to correct false claims, and by pressuring technology companies to adjust algorithms that boost fake news.

But governments and platforms can do only so much. “In many ways, a policy solution also has to be an individual solution,” says Ang, “because a policy solution is actually trying to change how individuals behave.”

Changing individual behaviour is hard. Ang draws the Covid-19 parallel directly, saying: “We don’t wash our hands after we sneeze. We sneeze in front of other people. The same thing happens with misinformation. We don’t check what we read and we don’t check before we share.”

Edson Tandoc Jr, a professor in communication studies at Nanyang Technological University, says debunking misinformation is at the heart of this change.

That is easier said than done, especially with those closest to us. Tandoc remembers that when a relative shared misinformation in a family group chat, he hesitated to speak up.

“Our societies tend to focus a lot on maintaining harmony and emphasising family ties,” he says, “and a big part of that is a respect for the elderly and seniority.” Still, he spoke up because he knows debunking misinformation works. “Based on my research, I felt that I should speak up.”

The comparison between combating viruses and fighting misinformation may well carry into the possibility of vaccination, too.

Researchers are working towards a “vaccine” for susceptibility to fake news — they call it “prebunking”. Watching a 90-second video on common misinformation tactics makes people less likely to fall for false claims afterwards. The research is promising — Google even tested the theory on 120 million people ahead of the 2024 European Union election — but it has not been deployed anywhere permanently.

We are still in the part of the misinformation pandemic where there is no widespread vaccination. Which means that, for now, the work still falls on us — checking, correcting and being willing to be corrected in turn.