Exactly six years ago, in June 2020, a peculiar piece of news from the small town of Huancavelica in Peru emerged. Eight technicians who had been sent to the town to do maintenance on a telecommunications antenna had been kidnapped by a group of villagers.

Local media reported that the villagers who were holding the technicians captive would only release them if they agreed to remove all the telecommunications antennas from the village. They had hinted that 5G antennas were contributing to spreading COVID.

The pandemic had been raging for over six months and had killed hundreds of thousands of people. On social media, the theory that there was a connection between 5G technology and the spread of viruses had spread like wildfire. And it turned into a forest fire – a digital forest fire, as research scientist Daniel Thilo Schroeder at SINTEF Digital calls it.

“A digital forest fire occurs when disinformation spreads quickly through social networks and ends in real-world consequences,” says Schroeder.

He studies how artificial intelligence (AI) can be used to influence people through social media. Schroeder is concerned with democratic resilience and societal security.

A digital forest fire occurs when disinformation spreads quickly through social networks and ends in real-world consequences.

“I’ve spent a lot of my doctorate time, and also my time as a postdoc, studying the 5G theory and how it spread,” says the SINTEF researcher.

The researchers reviewed Twitter activity before, during and after the peak of the conspiracy in spring 2020. They looked at how users communicated and how groups, so-called communities, grew and interacted over time. They saw how prominent conspiracy theorists helped to amplify the claims. It all ended with people setting fire to telecommunications antennas and kidnapping telecommunications technicians.

“Conspiracy theories like this grow from sporadic tweets to a global spread, with a dramatic upscaling of disinformation,” says Schroeder.

Back in 2020, large language models that we often refer to as AI, were still in their infancy. Now, six years later, artificial intelligence has become cheap common property. In combination with social networks, AI can cause digital forest fires of a completely different nature.

At the same time, it has become difficult for researchers to keep up with what is happening. The major platforms no longer give researchers access to their data.

“All the APIs (digital communication channels between computer programs) from the major media platforms have been closed. We no longer have the ability to analyse the content and see what is going on out there,” Schroeder says.

With agentic AI, it’s no longer just a question of language models being good at convincing you and getting you to change your opinion about a topic. Now AI can also coordinate the campaign itself.

In an article that involved researchers from a wide range of prestigious research centres including the Max Planck Institute in Germany, the University of Oxford in England and Harvard University in the United States, Schroeder has pointed out how agentic artificial intelligence – artificial intelligence that not only responds to what you ask, but can perform tasks more or less on its own – can give superpowers to information campaigns and the spread of disinformation.

• Maintain consistent identities and memory
• Coordinate towards common goals while varying tone and content
• Adapt to user engagement, platform signals and human reactions in real time
• Operate with minimal human oversight
• Can be deployed across platforms

“The large language models, such as ChatGPT, are known for being able to express themselves very persuasively. They aren’t good at reasoning, but they are extremely good at language. They manage to resonate with the way people think,” says Schroeder.

Research scientists are warning against what they call swarms of cooperative, malicious AI agents, unleashed by political leaders who want to influence public opinion and change people’s picture of reality and their political perceptions.

“With agentic AI, it’s no longer just a matter of language models being good at convincing you and getting you to change your mind about a topic. Now AI can also coordinate the campaign itself,” says Schroeder.

In practice, swarms put a large number – thousands – of coordinated AI agents into motion to infiltrate online communities and create consensus. The AI ​​agents pretend to be real people, they post messages using slang and phrases that make them appear real. The AI agents can coordinate with each other, change strategy along the way and stay active for a long time. They can write blogs and send emails, choosing which channels to use depending on what serves them best.

“It’s just frightening how easy these things are to vibe code and just have small bot armies that can actually navigate online social media platforms and email and use these tools,” Schroeder stated in an interview with the The Guardian newspaper earlier this year.

Schroeder believes it is important to monitor the occurrence of such AI swarms, and supports the idea of ​​an AI observatory that can investigate how artificial intelligence affects society.

“We need to regain access to data, so that it becomes possible to actually study it. Researchers and decision-makers must be given the opportunity to exchange data and results, so that new mechanisms can be developed to capture the phenomenon,” he says.

It’s just frightening how easy these things are to vibe code and just have small bot armies that can actually navigate online social media platforms and email and use these tools.

Schroeder believes just monitoring content on the platforms is not enough.

“Content-based detection is no longer enough, now that large language models can pass the Turing test (the Turing test is a test proposed by Alan Turing, in which a machine’s intelligence is assessed based on whether it can carry on a conversation in such a way that a human cannot distinguish it from another human) and similar tests. Instead, we believe that we need to examine the behavioural patterns of such swarms, such as how they coordinate their activities.”

Schroeder compares the situation to the Russian Twitter operation in 2016, where Russian troll factories were used to influence the US presidential election. The campaign did not have much impact. It did not reach many people and had no visible impact on the election result.

“At that time, it was about using scripts – small computer programs – that tweeted the same message over and over again. They were easy to detect. And they had to use real people to do the work. They reached very few people. We think this has changed. Because now you can scale up,” says Schroeder.

Sources:

How malicious AI swarms can threaten democracy. The fusion of agentic AI and LLMs marks a new frontier in information warfare. https://www.science.org/doi/10.1126/science.adz1697

The 5G COVID-19 Digital Wildfire: An evolving network of Twitter contacts to explore phase transition metaphors in viral misinformation https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0343661