Science: AI-Generated Content Difficult to Distinguish by Humans

Bratislava, 7 September (TASR) - People have trouble recognising content generated by artificial intelligence (AI), with text and images being the most difficult to distinguish, whereas for speech, the success rate is around 60 percent, Robert Moro from the Kempelen Institute of Intelligent Technologies reported earlier this week, stressing that the ability to distinguish such content can be enhanced by training. In the future, there may be an increase in personalised attacks, as fraudsters use data from posts on social networks or messages obtained through phishing. "In this way, they can create responses that are mimicking one's own style, thus making them appear even more credible," Moro explained. Clinical psychologist and forensic expert Dusan Kesicky pointed out that age or education do not play a role in getting tricked, as the fraudsters attempt to induce emotional destabilisation in their victims. This can be done by causing them to fear for their close persons, or by bringing positive news about winning or appreciation of their investment. However, getting their victims to act in haste and without thinking is always their goal. According to Moro, AI models capable of conducting autonomous conversations or transforming faces in videos in real-time may also become commonplace. Even though large language models and speech generators still work best for English, many of them can already handle Slovak reasonably well. The latest AI models only need a few-seconds-long recording to be able to clone the required voice. Kesicky noted there is no universal recipe for completely avoiding fraud. Although training and experience helps, the tricks used by fraudsters are changing all the time, so it is a never-ending process. However, he believes it is important for this topic to stay in the public consciousness. Slavomir Psenak, Chief Technology Officer with the O2 mobile operator, said that up to 70,000 fraudulent calls are blocked everyday, but it is very difficult to tell which are legitimate ones. "The principle is that we cannot block any legitimate calls made by our customers," he added. jrg/mcs/am