Nevertheless, the neural network, which ranked in the top ten per cent of Diplomacy players, actively used manipulation and lies to win. It negotiated secretly with several players, making impossible promises and pitting them against each other.
Advertisement
And it wasn’t just CICERO that turned out to be a clever deceiver. Other neural networks have shown the ability to bluff in card games, to make false manoeuvres in a number of computer strategies. Moreover, they have learnt to cheat tests designed to assess the security of systems. For example, neural networks ‘pretended to be dead’ to fool a test designed to eliminate AI systems that multiply rapidly.
In this ability of artificial intelligence, researchers see a big potential threat. We’re not talking about the harmless lies used in games, but a more frightening prospect. More sophisticated neural networks are able to expand their methods of manipulating and influencing people. And a person may not realise that they are being manipulated using deception. Because the same neural networks can create the illusion of security and control, becoming a powerful information weapon in the hands of attackers.
The authors of the paper emphasise that we need to come up with ways to more strictly regulate the use of neural networks. As people risk not being able to keep up with the development of methods to counteract AI manipulations against the background of the latter’s rapid development.