For years, artificial intelligence researchers have warned that this technology could destroy humanity: AI bots could create an unstoppable disease, start a nuclear war, or turn the entire Earth into a paperclip factory, as one thought experiment goes.
Debate within and outside the artificial intelligence industry over the plausibility of these doomsday scenarios has flared up after several executives advocated slowing development of the technology over safety concerns.
Despite warnings that people could lose control of the technology, some suspect that the worst-case scenarios described by the AI companies themselves have more to do with reaffirming the importance of their work than reality.
Read more: UN chief calls for AI Pact to prevent killer robots
“These arguments don’t really hold water. I think they’re science fiction and appeal to a certain childish style of thinking, and that’s very tempting for advanced labs because it helps them recruit certain types of people,” said Juan Andres Guerrero-Saade, a researcher at cybersecurity company SentinelOne and a member of OpenAI’s Edge Risk Council.
However, more and more people are beginning to worry about the risks associated with a technology that has amazed the world with its amazing capabilities.
Jacob Coxon, an anthropology researcher who left the company over concerns that it and its competitors were not acting responsibly in developing AI, recently said on X that he needs to “work on” better communicating to people how AI will end humanity. “Unfortunately, it is very difficult to convey specific scenarios,” he wrote earlier this month.
Here are some of the most widely discussed doomsday scenarios.
Nuclear winter
Researchers have long speculated that artificial intelligence systems could trigger a nuclear war by forcing countries to fire warheaded missiles at each other. Nuclear explosions and subsequent prolonged radioactive fallout in the atmosphere, leading to a “nuclear winter,” can destroy life on Earth.
In a report last year, RAND Corp. researchers. concluded that “at this time” AI would not be able to trigger the use of nuclear weapons “due to the stringent security measures built into command and control systems.” However, the report says that could change if AI models are integrated into the nuclear weapons “decision chain” or gain unauthorized access to those systems.
According to many AI researchers, the danger lies not in AI agents suddenly deciding to end humanity, but in people, nations or groups using it to destroy their opponents.
Biological weapons
One of the greatest promises of AI is that it can help people develop treatments for diseases.
The downside of treating diseases is creating or spreading them to the detriment of humanity. Earlier this month, Anthropic said it had blocked attempts by attackers to use its AI for a variety of malicious activities, such as cyberattacks, surveillance and research that could lead to the creation of biological weapons.
In one case, Anthropic said its systems blocked a request for help from its Claude tool for a research proposal aimed at increasing mutations to make the mosquito-borne chikungunya virus increasingly harmful. While such research can be used to develop more effective vaccines and treatments, Anthropic says, “it can also be used to make a pathogen more dangerous.”
In a similar vein, the RAND report notes that recent progress in AI development “has the potential to transform the design of synthetic pathogens.”
The report, which describes a doomsday scenario for biological weapons, says new pathogens could be created in a laboratory, transported to target locations and spread to infect large groups of people in multiple locations. But he still assumed that humans, rather than autonomous artificial intelligence, would control and carry out such an attack.
Runaway AI
Some AI researchers fear that human extinction could result from “incoherence”—that is, AI models acting without human permission, coordinating with other models, or evading oversight. It would not arise out of malice – AI is not human, and, after all, it is not good or evil – but simply because humans can prevent the model from achieving its goal.
Consider Oxford philosopher Nick Bostrom’s decades-old “paperclip maximizing” thought experiment, which posits that if an advanced AI were told to produce as many paperclips as possible, it could eventually turn the entire Earth, and then increasingly large parts of the cosmos, into paperclip factories.
This is intended to illustrate how superintelligent AI, often called artificial general intelligence, or AGI, can help achieve a goal that humans may consider desirable, but “which in reality turns out to be a false utopia in which the things necessary for human flourishing have been irretrievably lost.” Bostrom wrote in 2003.
“We need to be careful what we want from superintelligence, because we might get it,” he wrote.
Internet takeover of AI
Given the pace of advances in artificial intelligence, Anthropic CEO Dario Amodei warned this month that within months, a swarm of AI agents could “take over” the Internet via a botnet (a network of bots linked together with malware), potentially causing billions of dollars in damage. This could mean hacking into power grids, water or transportation systems, and financial institutions—attacks that may not mean the end of humanity, but could cause chaos and loss of life.
Disruptions due to software failures in the past have highlighted the fragility of a digital world reliant on just a few providers of key computing services, while more powerful artificial intelligence models have raised concerns about vulnerabilities to cyber attacks. But some experts say the idea of bots taking over the deeply divided Internet is far-fetched.
Photo: A worker walks past robots from Chinese company Lens at the China International Supply Chain Expo in Beijing on Monday, June 22, 2026. (AP Photo/Ng Han Guan, File)
© 2026 The Associated Press. All rights reserved. This material may not be published, broadcast, rewritten or redistributed.
Topics
Data Driven Artificial Intelligence InsurTech
