Finnish professors assess artificial intelligence risks — YLE News
In Finland, Professor Arno Solin of Aalto University and Professor Pekka Abrahamsson of Tampere University assess the likelihood of the most severe artificial intelligence development scenarios differently. As YLE News reports, both researchers consider malicious use of the technology to be a current risk, while their assessments differ regarding the possibility of losing control over AI.
Malicious use of AI
One of the three scenarios considered by ChatGPT and Claude involves people or organizations using artificial intelligence for cyberattacks, developing viruses, or gaining access to weapons of mass destruction. Both language models called this option the least likely of the three because it involves human control rather than AI's own goals.
Solin believes that the main current danger lies precisely in the malicious use of the technology. According to him, AI can already help create more convincing online scams and personalized influence on public opinion. He also noted that AI may make it easier in the future to create bombs, biological weapons, or chemical weapons.
Abrahamsson also called the use of AI for malicious purposes the greatest current risk. In his view, bad actors may deploy an “army” of AI agents to take over human-created infrastructure.
Conflict between major powers
The second scenario suggests that an AI-driven arms race could disrupt the balance of power between major states and accelerate decisions to launch attacks. For example, an automated cyberattack on the critical infrastructure of one state could be interpreted as an act of war, potentially triggering a retaliatory strike and further escalation to the use of weapons of mass destruction.
More current news is available on the UA.News Telegram channel Telegram.
The language models assessed this scenario as intermediate in likelihood. The risk increases if decisions on responding to attacks are delegated to AI without human involvement. Abrahamsson stressed that a system's autonomy, broad powers, and access to a large volume of data without human oversight create a complex problem.
Loss of control over systems
According to the surveyed language models, the most likely of the three scenarios is a situation in which AI becomes so autonomous that people delegate more and more authority to it, while the system's goals cease to align with human goals. In that case, according to the scenario description, the system may be impossible to stop.
Abrahamsson expressed concern that an insufficiently understood AI model could create subsequent generations of systems that would be even harder to understand and control. Solin, by contrast, considers the likelihood of AI getting out of control to be low, but believes it is worth preparing for such a risk because of the high potential cost of an error.
According to Solin, modern language models operate on servers and in computing clusters, so people can stop the process. He noted that there is no evidence that AI has independently transferred its own code to another server and launched it there.
