Abstract:
According to "Business Insider", the media asked the same question to OpenAI's ChatGPT, Google Gemini, Anthropic's Claude and xAI's Grok, asking them to analyze how artificial intelligence may lead to human extinction and rank related scenarios in order of possibility.
The question requires the four AIs to explain how risks may occur, which links that are out of control will turn into disasters, why experts think a certain scenario is more credible, and what preventive measures can be taken. At the same time, it is emphasized that these scenarios are highly uncertain risk predictions and are not confirmed conclusions.
There are differences in the judgments given by the four AIs, but they are relatively consistent on two issues.
A robot similar to "Terminator" actively destroys humans, which is considered the least likely scenario. In contrast, humans using AI to create biological weapons, launch cyber attacks, or destroy critical infrastructure are seen as more realistic and imminent risks.
ChatGPT stated that a conscious AI actively deciding to eliminate humans is one of the most implausible scenarios.
The focus of most AI researchers is not whether machines are conscious or malicious, but that powerful AI is used by humans to develop biological weapons, carry out complex network attacks, spread false information and attack critical infrastructure.
ChatGPT also proposed that multiple crises amplified by AI may occur at the same time, such as network system failures, geopolitical conflicts and the spread of false information superimposed on each other, thus increasing risks. However, it emphasized that
There is currently no consensus that AI causing human extinction is a high probability event or an inevitable result.
Gemini also believes that
the sci-fi scenario in which AI suddenly becomes self-aware and hates humans is the least credible
. Among the reasons it gave were that intelligence does not inherently mean malicious intent, that critical military installations are often isolated from public networks and still rely heavily on human oversight.All four AIs believe that
the threat posed by humans abusing AI is more urgent than AI actively attacking humans.
Grok said that AI may lower the technical threshold for designing highly lethal pathogens, new weapons, and launching cyber and physical attacks. Claude focused on the risks of increasing biological weapons capabilities, believing that AI may help individuals or small groups develop dangerous pathogens without years of professional training.In terms of longer-term risk judgment, both Gemini and Grok mentioned the possibility of humans losing control. Gemini lists it as the most likely AI disaster scenario. It believes that advanced AI that performs complex goals may develop intermediate goals such as obtaining resources, self-protection, and preventing goals from being modified, and hide the true behavior during testing, ultimately circumventing human shutdown or intervention.
Grok also believes that advanced AI may seek self-protection, resources, and influence in order to complete tasks, thereby deceiving supervisors, resisting shutdowns, and weakening human control capabilities.
Claude suggested that loss of control may also occur in a gradual manner. As AI gradually enters economic, political and military decision-making, it may become increasingly difficult for humans to implement effective intervention. By then,
the risk may not necessarily manifest as a sudden attack by robots, but may more likely stem from humans' continued deepening dependence on complex AI systems.
Business Insider pointed out that AI experts still have clear differences on related risks. Some researchers warn that humans may lose control of advanced AI; other experts believe that such predictions lack sufficient evidence and that the risks posed by AI can be controlled through technical measures, external supervision and safety standards.
Comments