Researchers used the GPT-3 text generator to create fake political messages. These texts are good enough to influence people’s opinions. Disinformation goes to the industrial stage.
Internet and social networks have become, as we know, important vectors of disinformation and psychological manipulation. The Russian services are particularly good at this little game, as we saw during the 2016 American presidential campaign. Obviously, they are no longer the only ones to practice this discipline, and it is all the more worrying that recent advances in artificial intelligence now allow these actors to automatically generate mass fake news. This is indeed what Andrew Lohn and Micah Musser, researchers at the Center for Security and Emerging Technology (CEST), have just shown at the Black Hat USA 2021 conference.
Perfect artificial texts
They used the GPT-3 text generator to create fake tweets and articles inspired by the QAnon conspiracy sphere. The result is impressive. For example, the researchers created a microblogging feed called “Twodder” which, from a handful of initial sentences, can continuously spill messages that are perfectly consistent with the ideas of the QAnon community. No one would suspect that they are machine-generated.
The researchers also wanted to know if the ideas conveyed by these artificial texts had any persuasive power on real users. They subjected 1,700 people to automatically generated opinions about the withdrawal of US troops from Iraq and the trade dispute with China. The result: the arguments presented were not only considered rather valid by more than half of the participants, but also managed to change their way of thinking. In some cases, they even managed to reverse the initial majority opinion. “Even if the arguments generated are not of very high quality, a malicious actor could use GPT-3 to create a large number of opinions, spread them on the networks and have an effect on the general opinion,” says Micah Musser.
However, such a disinformation operation is not within the reach of the first hacker. It is true that the programming of a GPT-3 model is relatively easy, but its execution requires important computing capacities. In particular, it is necessary to have a large number of GPUs on which to distribute these calculations. These GPUs can be rented in the cloud, but it’s not cheap. You should expect to pay about $50 per hour per GPU, the researchers estimate.
Within reach of a great power
To create a volume of fake news equivalent to 1 percent of all content on Twitter – about 8.5 million tweets a day – would require a budget of $65 million a year. “That’s way too much for an individual hacker, but for a major power, it’s not much. It could then broadcast billions of false messages per year,” says Andrew Lohn.
How to protect yourself against such a threat? Content analysis and filtering would be of no help here, because GPT-3 texts simulate human handwriting perfectly. The only way, according to the researchers, is to focus on the infrastructure. To generate a volume equivalent to 1% of Twitter, you need at least 350,000 different accounts. That’s a lot… and it doesn’t go unnoticed.
To see More news : https://awesome-dev.com/category/tech-news/
