Jacob Coxon, a researcher who spent the past three years working on pretraining research at OpenAI and Anthropic, has left Anthropic with an extraordinary warning about the future of artificial intelligence.
“I resigned from Anthropic today,” Coxon wrote in a public statement. He accused both companies of failing to act responsibly and claimed they are “racing straight toward self-improving superintelligence and gambling with our lives.”
His message quickly drew attention because it did not come from an outside critic of the AI industry. Coxon worked inside two of the companies at the forefront of developing increasingly powerful AI models.
His most alarming prediction concerns what could happen if those systems eventually become substantially more capable than humans.
According to Coxon, future AI systems could become capable of hacking computer systems, accelerating scientific and technological advances and acquiring significant resources and influence in the real world. He went even further, claiming that some people working inside AI companies privately believe the technology could pose an existential threat to humanity before the end of the decade.
Those are extraordinary claims. And they require an important distinction: Coxon is warning about what he believes future AI systems could become, not describing the capabilities of ChatGPT, Claude or other AI models available today.
Could AI really take control?
Current artificial intelligence remains dependent on infrastructure built, powered and controlled by humans.
AI models run in data centers filled with specialized chips and servers. Those facilities require enormous amounts of electricity, cooling systems and telecommunications infrastructure. Services can be disrupted by power failures, hardware problems, network outages and other physical events.
An AI model also does not automatically gain access to power grids, bank accounts, weapons, factories or government systems simply because it becomes more capable.
For software to affect the physical or digital world, it needs access. That can mean APIs, passwords, computer systems, external tools, robots or permissions deliberately provided by people and organizations.
That reality makes the popular image of an AI suddenly “waking up” and independently taking control of the world far more complicated than it appears in science fiction.
However, dependence on physical infrastructure does not eliminate every potential risk.
Data centers can have backup power, computing can be distributed across different locations and increasingly capable AI agents can be connected to tools that allow them to perform actions rather than simply generate text.
The more autonomy humans give these systems, the more important cybersecurity, access controls, monitoring and human oversight become.
A real debate behind an extreme warning
Coxon’s prediction that AI could cause human extinction is not an established scientific fact. There is currently no publicly available AI system known to possess the kind of autonomous, self-improving superintelligence he describes.
There is also no scientific consensus establishing that such a system will emerge by 2030 — or that it will necessarily emerge at all.
But the broader question behind his resignation is taken seriously within the AI industry.
Major AI companies have developed frameworks specifically intended to evaluate increasingly powerful capabilities and establish safeguards before deploying advanced systems. Researchers are studying issues ranging from cybersecurity and biological risks to autonomous behavior, deception and the possibility of humans losing meaningful control over future systems.
The disagreement is largely about how likely those extreme outcomes are, how quickly AI capabilities could advance and what precautions should be taken today.
Coxon belongs to the more alarmed side of that debate.
He argues that AI companies are caught in a competitive race: each fears that slowing down could allow another laboratory or another country to develop more powerful systems first.
From his perspective, that competition creates incentives to continue developing increasingly capable AI even when researchers themselves remain uncertain about how those systems could behave.
The immediate issue is already here: work is changing
For companies and workers, however, the most relevant effects of artificial intelligence do not require waiting for superintelligence.
AI is already changing how people write software, analyze information, create content, provide customer service, recruit employees and perform administrative work.
That transformation raises much more immediate questions for employers: Which tasks should be automated? Which decisions require human oversight? What information should an AI system be allowed to access? And which skills will employees need as AI becomes embedded in everyday work?
Those questions are particularly relevant for HR departments and business leaders because adopting AI is not simply a technological decision. It also involves workforce planning, training, cybersecurity, accountability and determining where human judgment remains essential.
Coxon’s resignation deserves attention because it comes from someone who has worked directly on advanced AI systems. But his dramatic prediction should be understood for what it is: a warning about a possible future, not evidence that an AI takeover is currently underway or inevitable.
The debate his departure exposes may ultimately be more important than the prediction itself.
As artificial intelligence becomes more capable, companies, governments and society will have to decide not only what AI can do, but how much autonomy humans are willing to give it.
And that is a question the workplace is already beginning to confront.
