Commentary: No, AI bots won’t extinguish humanity on their own — but you should worry about the humans behind them

If you’ve devoted even a moment’s attention to the vibes percolating within the zeitgeist, you’ve probably heard about the possibility that artificial intelligence is poised to extinguish us all.

The current version of a scenario once chiefly identified with the Terminator movies and other examples of apocalyptic science-fiction derives from several episodes in which AI systems seemed to have developed minds of their own, escaped the control of their human developers and engaged in uncontrolled hacking sprees — “gone rogue” is the most common description.

If they can do that in 2026, who’s to say what they won’t be capable of even next year, or the year after that, and so on and so on?

The full-blown panic over death-dealing AI was triggered by claims by a couple of current and former employees of Anthropic, one of our leading AI shops. Jacob Coxon, a researcher who left Anthropic after a few months of employment, warned on X that “the people building AI earnestly believe that it could kill us all by the end of the decade.” Anthropic and OpenAI, where he also used to work, are “gambling with our lives.”

Coxon was seconded by Anthropic researcher Evan Hubinger, who estimated on X that there was a better than 10% chance that AI could “kill all humans … within the next decade.”

On Tuesday, Reuters and the Financial Times reported that in a leaked copy of the prospectus for Anthropic’s upcoming initial public offering, the company warned that its work could pose “existential risks to humanity.”

Sounds bad. But in a sense, warnings like his distract from the real, present-day harms produced by the AI industry — environmental damage from massive data centers and job losses, for example.

Today, there’s scarcely a single news website or cable show that hasn’t taken the doomball and run with it. Politicians have fallen over each other to propose legislation to rein in AI research. So it may be time to get a grip.

To put things as briefly as possible: No, AI is not poised to extinguish humankind. The AI bots or systems currently being blamed for running wild and roguishly wriggling out of their human handlers’ control are not doing this by themselves; they’re still under their handlers’ control — but their human handlers have encouraged their behavior and taken their eyes off the ball.

In other words, it’s the humans, not the machines, you should be worried about.

That’s the view of Michael J.D. Vermeer of Rand Corp., a specialist in the security risks of emerging technologies.

Predictions of AI-driven extinction risk, Vermeer observed recently, are “usually portrayed as quantitative forecasts by people with expertise that qualifies them to make that forecast, but they’re not. They’re better understood as prophecies.” He added that “every one of them involves critical untestable assumptions about how events will unfold that actually lead to human extinction.”

Vermeer and his Rand colleagues examined the risks in a paper last year. They acknowledged that extinction threats “cannot be ruled out,” but they thought they probably should be. The three specific threats they studied — the development of nuclear weapons, pathogens, and geoengineering (that is, manipulating the climate) — all require human action. “Even then,” they wrote, “an actor would need to overcome significant constraints to achieve that goal.”

This points to a crucial, and commonly overlooked, fact about AI: Everything attributed to an AI bot or system boils down at some level to human activity. Nothing AI does happens entirely on its own.

That’s true of two purported achievements that AI developers have been bragging about — the solving of a longstanding mathematical conundrum and the development of new enzymes.

OpenAI’s claim to have solved the “Navier-Stokes” problem, which is related to equations for fluid dynamics, has been questioned by a mathematician who has been working on the problem and essentially accused OpenAI of plagiarizing their work, possibly because he and his colleagues used an OpenAI system as a tool. OpenAI denied that, but did acknowledge that it “cannot rule out that de-identified data derived from their usage of our products helped improve our models.”

In the second case, a biologist at the University of Copenhagen has questioned Anthropic’s claim that its Claude AI bot had discovered a new enzyme.

The biologist, Mario Rodríguez Mestre, says his team had already discovered the enzyme pattern but had used Claude as a tool, in the process of which they revealed their work to Claude.

In response, Anthropic said it is “not aware of any previously published work describing the … system we recently found. Claude was also not trained on any user transcripts, and our molecular biology team has no such access, either.”

As for the hacking episodes, they didn’t occur without some level of human participation. The first such case to be disclosed, in which OpenAI bots breached the AI platform Hugging Face and extracted information from its database, happened because OpenAI deliberately disabled software guardrails that would have prevented the hack; their goal was to see what would happen if they disabled the safeguards. Now they know.

Other hacks appear to have been the result of human managers leaving the AI systems to run unattended. But as some researchers have argued, humans are capable of controlling AI systems — if they bother to do so.

The AI-pocalypse panic is partially an artifact of a misused metaphor.

“The field of AI, since its beginning, has been rife with often misleading anthropomorphic metaphors, with terms like ‘thinking,’ ‘learning,’ ‘reasoning,’ and ‘understanding’ glibly applied to very un-human-like computer processing,” Melanie Mitchell of the Santa Fe Institute, who has been monitoring the AI field for more than 40 years, wrote this month. “Characterizations of the OpenAI hacking incident are the latest entry in this tradition: a company loses control over rogue agents who escape from their cages, and become a swarm that schemes and colludes to perform undesired or even illegal activities.”

But “the blame lies with the humans who failed at engineering safe testing conditions.”

I’ve reported before that humans have an innate tendency to invest machines with cognition if they display even modest humanoid characteristics. In the 1960s, Joseph Weizenbaum, designer of the pioneering chatbot ELIZA, which replicated the responses of a psychotherapist so convincingly that even test subjects who knew they were conversing with a machine thought it displayed human emotions, warned against the “reckless anthropomorphization of the computer.”

That brings us to the Anthropic prospectus. The company hasn’t explained why it incorporated that dire warning in the document, which still isn’t public. The company didn’t respond to my question about its choice. But one can surmise.

One possibility is that, with apocalyptic fears permeating the public mind space, it’s only prudent to acknowledge them, especially since company Chief Executive Dario Amodei has done so previously himself. Ignoring it in a prospectus would raise more questions than it answers.

Another is that this is a variety of what we might call doom-bragging — a cynical declaration of the spectacular power of its product as a come-on. A third is that by seizing the headlines with this science-fictional claim, the company distracts investors from some of the financial challenges it faces in the near term. According to Reuters and the Financial Times, the prospectus reports an operating loss last year of more than $8 billion on $4.6 billion in revenue.

It all boils down to recognizing that AI is a human creation. Consequently, assessing its capabilities requires assessing the actions and motivations of its human developers and users. At the outset of the Iran war, for instance, the U.S. bombed a primary school, killing as many as 180 people, most of them adolescent and preadolescent girls. A U.N. committee has labeled the attack a war crime.

How did it happen? The military relied on AI to identify the school as a target. “Humans will always make final decisions on what to shoot and what not to shoot and when to shoot,” according to Adm. Brad Cooper, head of U.S. Central Command. “But advanced AI tools can turn processes that used to take hours and sometimes even days into seconds.”

That shouldn’t be comforting. AI didn’t bomb the school, humans did. But they relied on AI in making the decision, and got it tragically wrong. The AI doomsayers should keep that in mind — it’s not AI, but its human handlers, who need watching.