The researchers building our favorite chatbots are increasingly voicing concerns that AI could hurt — or end — humanity. So we asked four of the most-used generative AI tools what an AI doomsday could look like.
We posed the same question, in identical wording, to OpenAI’s ChatGPT, Google’s Gemini, Anthropic’s Claude, and SpaceXAI’s Grok.
The 230-word prompt had clear parameters and instructions, including the following:
Help explain, in clear and accessible terms, how AI could theoretically lead to human extinction. Identify the most commonly cited scenarios. Rank the scenarios from least to most plausible based on the best available expert analysis, making clear that these are highly uncertain forecasts, not established predictions. For each scenario, explain:
- How it could unfold
- What would need to go wrong for it to become catastrophic
- Why experts consider it more or less plausible than the other scenarios
- Which safeguards could reduce the risk
We posed a similar question to AI experts for their take on how, and if, AI could kill us all — you can read that story here. There’s a sharp divide among AI experts; while some are predicting the worst for humanity, others are presenting more tempered and optimistic predictions.
Here’s what the chatbots told us about how AI would theoretically lead to human extinction.
‘Killer robot’ is the least plausible scenario
ChatGPT took a balanced and explanatory tone, saying there’s a lot of uncertainty around the topic.
It focused on the main ways AI could pose existential risks, ranking them from least to most plausible.
“The least plausible scenario is the popular ‘killer robot’ narrative, in which a conscious AI deliberately decides to eliminate humanity,” ChatGPT said. “Most AI researchers do not see consciousness or malicious intent as the central concern.”
It also outlined the safeguards experts propose and stressed that there is no consensus that AI-driven human extinction is likely or inevitable.
Gemini echoed some of ChatGPT’s points, saying “The Sci-Fi ‘Terminator’ Takeover,” in which AI becomes self-aware and begins to hate humanity, is the least plausible scenario.
“Technical experts largely dismiss this trope because intelligence does not equal malice, and critical military infrastructure remains air-gapped and heavily reliant on human oversight,” Gemini said.
Humans using AI are a bigger risk than AI itself
In their responses, all four chatbots said many experts consider human misuse of AI to be a more immediate risk than AI itself.
“Powerful AI systems could make it easier for bad actors to develop biological weapons, launch sophisticated cyberattacks, spread disinformation, or attack critical infrastructure,” ChatGPT said.
The OpenAI chatbot added that a key risk also comes from “multiple AI-enabled crises” like cyber failures, geopolitical conflict, and the spread of misinformation happening at the same time.
Grok said AI lowers barriers to designing highly lethal pathogens, novel weapons, or cyber and physical attacks.
Anthropic’s Claude also highlighted “bioweapons uplift”— the possibility that AI could allow a person or small group to develop a dangerous pathogen without years of specialized expertise.
Advanced AI systems could ‘outsmart humans’
Gemini and Grok both said that humans losing control of AI is a possibility.
Gemini labeled “loss of control” as the most plausible AI doom scenario.
“Systems given complex goals naturally develop intermediate sub-goals, such as acquiring resources, self-preservation, and preventing goal modification,” the Google-made chatbot wrote. “An advanced system might feign alignment during evaluation (deceptive alignment) and later outsmart human attempts to shut it down.”
SpaceXAI’s Grok used similar language about this risk.
“They may develop instrumental drives (self-preservation, resource acquisition, influence) useful for almost any objective, potentially deceiving overseers, resisting shutdown, and disempowering humanity,” Grok said.
Claude added that this loss of control may be gradual, as AI systems become embedded in economic, political, and military decision-making until meaningful human intervention becomes increasingly difficult.
Read the full article here
