Andrew Rogoyski names premature trust the main AI risk
The most immediate risks of artificial intelligence are not related to the emergence of a superintelligent system that will start a war against humanity, but to premature trust in imperfect models and granting them control over critical processes. This view was expressed by physicist and artificial intelligence specialist Andrew Rogoyski in an opinion piece for Live Science.
The author notes that modern advanced models perform well in writing texts, programming, and mathematical tasks. At the same time, in his assessment, they lack the qualities needed for useful artificial general intelligence: consistent reasoning, long-term planning, factual consistency, and reliable performance beyond the data on which they were trained.
Dependence on infrastructure
According to Rogoyski, scenarios involving an all-powerful digital intelligence capable of independently destroying humanity distract from more practical threats. AI systems depend on electricity, cooling, data centers, specialized chips, networks, technical personnel, and supply chains. In many scenarios, cutting off power, restricting access to graphics processors, shutting down cooling, or disconnecting the network would stop such a system from operating.
More current news is available on the UA.News Telegram channel Telegram.
Instead, systems with limited understanding that have been granted autonomy and access to critical infrastructure may be dangerous. In the author's view, they can cause harm not because of intent, but because of narrowly defined goals and an inability to account for broader consequences. As an example, Rogoyski mentions Nick Bostrom's thought experiment about an AI seeking to produce as many paperclips as possible while ignoring other consequences of carrying out such a task.
Cyber risks and employment
Rogoyski considers AI-enabled cyberattacks and prolonged economic upheaval to be the two most immediate dangers. Hospitals, banks, energy grids, logistics companies, and governments depend on interconnected digital systems, so such attacks could cause significant financial losses and disrupt services.
The author also identifies the premature reduction of employees by companies that may mistake convincing demonstrations by language models for their ability to perform reliable work as another risk. In his assessment, this approach may lead to the loss of accumulated knowledge, accountability, professional connections, and trust within organizations. Rogoyski suggests that a large-scale incident in a cyber operation involving advanced AI could trigger a rollback of investment in the technology and a new “AI winter,” when funding and research decline.
