AI: Human Abuse Seen as More Urgent Than Loss of Control

Four of the most widely used chatbots were asked to rank the ways in which AI could lead to human extinction. Their responses converge on one point: malicious use by humans is seen as the immediate risk, while loss of control over advanced systems is presented as a plausible scenario. The cliché of the conscious killer robot is placed at the very bottom of the list.

Loss of Control: A Plausible Scenario According to Gemini and Grok

Two of the models surveyed believe that a loss of human control over advanced systems is possible. Gemini describes it as the most plausible apocalyptic scenario and outlines systems with complex objectives that could develop sub-goals such as resource acquisition, self-preservation, and preventing modifications to their objectives. Google's chatbot adds that an advanced system could feign its alignment during evaluation, then thwart human attempts to shut it down. Grok mentions similar instrumental motivations—self-preservation, resource acquisition, influence—that could deceive supervisors, resist shutdown, and disarm humanity. Claude specifies that this loss of control could occur gradually as AI integrates into economic, political, and military decisions, complicating any significant human intervention.

Human Misuse: Biology, Cyber, and Combined Crises at the Forefront

All four chatbots report that many experts consider the abusive use of AI by humans to be a more immediate peril than AI itself. ChatGPT cites cases where powerful systems could facilitate the development of biological weapons by malicious actors, sophisticated cyberattacks, the spread of misinformation, or attacks on critical infrastructure. It also mentions the possibility that AI could simultaneously trigger multiple crises, such as computer failures, geopolitical tensions, or the circulation of false information. According to Grok, artificial intelligence facilitates access to the creation of highly dangerous pathogens, new types of weapons, or both digital and physical attacks. Claude highlights the risk of "enhanced biological weapons," with AI enabling an individual or small group to create a dangerous pathogen without requiring years of specialized skills.

ChatGPT Considers the Killer Robot Scenario Improbable

ChatGPT ranks the killer robot narrative—a conscious AI deciding to eliminate humanity—as the least plausible scenario and emphasizes the extent of uncertainties. It states that most AI researchers do not place consciousness or malicious intent at the core of their concerns. Gemini holds the same position regarding a Terminator-like takeover, citing widespread rejection by technical experts, noting that intelligence does not equate to malice, and that critical military infrastructures remain isolated and closely monitored by humans. ChatGPT further indicates that there is no consensus on the likelihood or inevitability of extinction due to AI and mentions having outlined protective measures proposed by experts.

Methodology: Four Models, a Structured Prompt, and Acknowledged Uncertainty

Four models—OpenAI's ChatGPT, Google's Gemini, Anthropic's Claude, and SpaceXAI's Grok—received a 230-word prompt asking them to explain in accessible terms how AI could lead to human extinction, identify the most cited scenarios, and rank them from least to most plausible based on the best available expert analyses. The message specified the highly uncertain nature of these projections and requested, for each scenario, a description of the unfolding, elements that could go wrong, reasons for the plausibility rating, and relevant protective measures. Meanwhile, AI experts are divided between catastrophic visions and more tempered perspectives, while researchers involved in the development of chatbots increasingly express their concerns about potential threats to humanity.