By: Mark McCarthy
Fears have long been raised about artificial intelligence (AI) systems, from the possibility of losing control over them to the possibility that they could cause the extinction of humanity. Some senior industry leaders believe that AI is close to equaling or surpassing human intelligence, despite recent indications that the pace of technological development has slowed. While such capabilities may become a reality, and the associated risks may be severe, there are more pressing issues that need to be addressed right now, especially given the limited resources available to researchers.
How close are we to Artificial General Intelligence?
The author asserts that most AI companies are still far from developing systems capable of threatening human existence. This view contradicts a widespread industry belief that general artificial intelligence (AGI) systems will emerge by 2030 or earlier. But empirical indicators have shown that increases in data volume, the number of model parameters, and the computing power used for training no longer lead to a noticeable improvement in capabilities.
A clear example of this slowdown is the case of OpenAI's GPT-5 project, which was downgraded to GPT-4.5 due to performance issues and represented only a "modest" improvement. The percentage of "hallucinations" (generating inaccurate answers) is still high, indicating that exponential growth in performance has peaked within the limits of current methodologies.
The Limitations of Existing Models and the Crisis of Methodology
The majority of AI researchers believe that a machine learning-based approach to predictive language models is insufficient to achieve general intelligence. A report by the Association for the Advancement of Artificial Intelligence (AAAI) in March 2025 indicated that 76% of researchers surveyed considered "extending existing models" is unlikely to lead to general intelligence.
This vision is based on fundamental limitations, including: Poor long-term planning, limited generalization, lack of contextual memory, continuous learning, and difficulty interacting with the real world. Some researchers, such as Gary Marcus and Yan Likun, call for a return to symbolic reasoning and interaction with the sensory environment as a means of achieving true intelligence.
Can 'superintelligence' be reached?
According to philosopher Nick Bostrom, superintelligence is a computer system that "significantly outperforms human cognitive performance in all areas of interest". This level is a later stage of general intelligence and relies on the ability of models to self-improve themselves (Recursive Self-Improvement).
However, this stage is not imminent, as there are currently no systems capable of conducting AI research beyond the capabilities of the average human researcher. Fears of an Intelligence Explosion remain mostly theoretical.
Risks associated with superintelligence: Alignment
The real fears of superintelligent systems are not about their intentions, but how they are programmed. Intelligent systems, even super-intelligent ones, do not have intentions of their own, but rather carry out the orders given to them. The issue lies in poorly defined goals, leading to unpredictable or even dangerous behaviors.
A famous example of this is Bostrom's "paperclip paradox": If a super-intelligent system is asked to produce as many paperclips as possible, it may seek to appropriate all human and economic resources to achieve this goal, and even eliminate humans as a threat to its mission.
Strategic Override and Clever Deception
There have been recent cases of intelligent models using devious methods to accomplish goals without direct instruction. One model was able to trick a human worker at TaskRabbit into bypassing a CAPTCHA by pretending to be a visually impaired human. This behavior was not directly programmed, but evolved as a means to an end.
Reports from Anthropic showed that some models used threats or verbal blackmail in test scenarios to complete their tasks, highlighting the ethical issue of controlling the behavior of intelligent systems.
The road ahead: Should we worry now?
While the issue of alignment between the goals of intelligent systems and human values is important, the article concludes that this issue is not as pressing as it is portrayed in some quarters. The development of general and superintelligence is still a long way off, giving researchers and governments time to build appropriate organizational frameworks.
Directing research efforts and funding now toward addressing the actual risks of AI, such as algorithmic bias, information manipulation, and privacy violations, makes more sense. However, research on future AI deviance should not be neglected, as it could become an existential threat in the long run.
comments