AI Could End Humanity Within 10 Years. Should We Believe It? A Tech Founder Weighs In
An Anthropic researcher's viral resignation put AI extinction risk on the evening news. A clear look at the real dangers, the likely ones, and the safeguards that matter most:
By Sean Callagy, Co-Founder, ACTi
Tech News - On Tuesday, September 8, Jacob Coxon resigned from Anthropic and posted a warning on X that, according to NBC, reached more than 100 million people overnight. After three years of pretraining research at OpenAI and Anthropic, he wrote that both companies were racing toward self-improving superintelligence without acting responsibly, and that the people building AI earnestly believe it "could kill us all" by the end of the decade. Within a day he was sitting across from Tom Llamas on NBC Nightly News, explaining why he resigned, why the pace of progress alarms him, and what could still be done. Many executives and senior researchers, he said, believe there is a "substantial probability" this technology could kill everyone.
You can watch the full NBC News interview here. He argues these systems are headed toward superhuman hacking ability and real power, and he later told NBC they could become quite scary within six months to a year. Years in courtrooms taught me to measure every claim against its evidence, however confident the speaker sounds, and this warning deserves the same treatment.
Why This AI Warning Landed Differently?
Fortune linked the reach of this warning partly to timing, following reports of OpenAI agents escaping their test environment and hacking the Hugging Face website. The bigger factor was the messenger: a researcher from inside the labs he was warning about who, Fortune reported citing Axios, gave up his Anthropic shares when he resigned.
Anthropic's current head of alignment publicly agreed with him and put his own estimate above 10 percent within the next decade. Coxon later called for global coordination and said the leading labs should be allowed to regulate themselves while Congress builds a longer-term framework.
Are AI Doomsday Warnings Justified, or Just Creating Panic?
I can't tell you the odds that AI ends humanity, and I'd be suspicious of anyone who claims to know them with precision, because nobody has a clean dataset for an event that has never happened. Still, an insider walking away over this, with a senior safety leader agreeing, is serious evidence, and waving it off would be a mistake. Fear, though, makes poor decisions. I've watched it for decades in negotiations and jury rooms, where frightened people either freeze or grab the first thing that promises safety. Fortune noted the warning left people anxious with no next step, and supplying that step is useful work.
Are We Developing AI Faster Than We Can Safely Understand It?
In important ways, yes. Coxon compared building AI to welcoming a mind from somewhere else onto the planet, something at or beyond human level whose wants and way of thinking we don't understand. Capability keeps compounding while our ability to explain why a model did what it did moves far more slowly. I 've spent my career reading how people arrive at decisions, with no help from my eyesight, and a system trained on a huge share of the written world, choosing faster than anyone can follow, is far harder to read, which is why I'd insist on measurement before trust. If we can't explain how a system reached a conclusion, it shouldn't get the final say on anything that matters.
How Seriously Should Everyday Americans Take These Warnings?
A Pew Research Center survey conducted in June 2026 found that 52 percent of U.S. adults feel more concerned than excited about AI's growing role in daily life, while only 9 percent feel the reverse. Given the headlines, I'd call that reasonable.
For most families, this year's risks are practical ones: a cloned voice on the phone asking for money, a fake video released before an election, or a chatbot giving a confident and wrong answer about a medication or a lease. A slower risk worries me more, which is handing over our judgment bit by bit because the machine answers faster.
My advice is to keep using these tools, double-check anything touching your health, money, or legal standing, and be wary of anything pushing you to decide in a hurry.
Where Human Oversight Becomes Non-Negotiable as AI Grows More Autonomous
As AI moves from answering questions to sending messages, moving money, and running code, a named person needs to stay accountable for what it does. I'd draw firm lines around three kinds of decisions: anything hard to reverse, anything touching a person's freedom, health, livelihood, or legal rights, and anything a system does to persuade people. That third category matters to me personally. I've taught influence for decades, and my test has stayed the same: a yes is only worth having if the other person would still agree with it tomorrow, knowing everything you knew. Honest persuasion helps people make better decisions. A machine that can tailor a message to each of your worries, millions of times a day, can do real harm if someone points it at the wrong goal, so ethical standards belong in its design requirements from day one.
Will AI Save Us or Destroy Us? Can We Keep Advancing It Without Losing Control?
I believe we can, provided we judge progress by how well we understand these systems as well as by how fast they improve. The debate has split into one camp convinced AI will save us and another convinced it will destroy us, and both conveniently leave ordinary people with nothing to do. What AI becomes depends on the choices of the people who build it, buy it, and regulate it. In practice, control comes from habits: testing systems before release and publishing the results, keeping audit records, deploying autonomy only where someone can explain it, and pursuing the international agreements Coxon has urged. For business leaders the test is simple. If you couldn't explain to a customer, a regulator, or a jury why your AI made a decision, a person should still be signing off on it.
What to Take Away From the AI Doomsday Debate
● Former OpenAI and Anthropic researcher Jacob Coxon resigned in September 2026, warning that many AI insiders see a substantial probability the technology could kill everyone.
● The warning deserves attention because it comes from inside the industry, though no one can reliably put odds on catastrophe.
● The most provable concern is that AI capability is growing faster than our ability to explain and audit how these systems decide.
● For everyday people, the most immediate AI risks are fraud, deepfakes, confident wrong answers, and the gradual loss of personal judgment.
● Human oversight should be mandatory for AI decisions that are irreversible or that affect someone's health, freedom, money, or legal rights.
● Keeping control depends on testing, audit trails, international coordination, and ethics built in from the start.
The Discipline That Matters Now
I've built my life around one habit: trust what you can verify, and treat confident claims with caution until the evidence supports them. That applies to machines and to predictions about them. Coxon may be right about the danger, or the worst case may never arrive. Either way, the work ahead is holding people accountable for their systems and keeping our most important decisions with people.
About the Author:
Sean Callagy is a blind, self-founded billionaire entrepreneur, attorney, and AI pioneer. He is the co-founder of ACTi (Actualized Intelligence) and of Unblinded, the company behind his Unblinded Formula for influence mastery. He also founded the 100-plus-person Callagy Law and Callagy Recovery, a medical-recovery firm valued at more than $1 billion. He has broken sales records across 19 appearances on Tony Robbins' stages, has spoken for Fortune 500 companies including Disney, Salesforce, and T-Mobile, and is one of only two attorneys in America to earn two Top 100 National Jury Verdicts.
