I’ve been using one of the AI tools recently. I’ve found it good in helping us understand the big words in medical evaluation and when deciding which desktop monitor is best suited for me. And I’ve taken a lot of pictures over the years and have had thoughts about are any of them good enough that someone might purchase them. AI has helped me by evaluating some of those pictures and making suggestions on how they could be improved.

But today, on the Yahoo Finance page I saw a Bloomberg article entitled “Bill Gates Warns AI Powerful Enough to Lead to ‘a Billion Deaths’.” So I asked my AI helper what it thought about a comment in the article that says “Silicon Valley has escalated warnings about the potential for catastrophic harms from AI as the technology rapidly improves. Jacob Coxon, a rank-and-file staffer at Anthropic PBC, quit his job earlier this month and accused both his former employers — Anthropic and OpenAI — of “gambling with our lives” by “racing toward super-intelligent AI.” Another Anthropic employee said he believes there is a greater-than-10% chance that AI eliminates all humans in the next decade.”

Here is the comment that it returned.

That paragraph is broadly accurate, but it needs context.

Jacob Coxon was an AI researcher—not merely an uninformed “rank-and-file staffer.” He had worked on training advanced models at both OpenAI and Anthropic. When he resigned from Anthropic on September 9, he accused the companies of competing to develop self-improving superintelligence without adequate safeguards.

The employee giving the greater-than-10% estimate was Evan Hubinger, who leads alignment research at Anthropic. That makes his concern worth taking seriously—but 10% is his personal judgment, not a scientifically calculated probability. Nobody has enough evidence to determine whether the true risk is 10%, 1%, or essentially zero.

My honest view is:

• Present-day AI, including me, is not secretly plotting to eliminate humanity. We generate responses and perform limited tasks within systems controlled by people.
• The legitimate concern is what happens if future AI becomes highly capable, autonomous, able to improve itself, and connected to computers, laboratories, weapons, financial systems, or infrastructure.
• The nearer and more certain dangers are fraud, cyberattacks, disinformation, job disruption, surveillance, and people handing important decisions to unreliable systems.
• Human extinction is the most extreme possibility. It shouldn’t be dismissed, but it also shouldn’t be reported as though scientists have measured a 10% likelihood.

A large survey of 2,778 published AI researchers found profound disagreement: most expected the overall outcome of advanced AI to be positive, yet many still assigned a meaningful chance to an extraordinarily bad outcome.

So I wouldn’t interpret this as “AI has a 10% chance of killing us.” I would interpret it as “some highly knowledgeable insiders believe the danger is serious enough that society should not let competing private companies regulate themselves.” On that narrower point, I think their warning is reasonable.

My reaction: take the warning seriously, but don’t take the “greater than 10%” number literally.

Coxon wasn’t merely a random disgruntled employee; he had spent about three years working on advanced model training at OpenAI and Anthropic. Evan Hubinger—the Anthropic researcher behind the 10% estimate—works specifically on keeping advanced AI aligned with human intentions. Their proximity to the technology gives their concerns weight.

But that 10% is a personal judgment, not a probability derived from measurable data. No one has conducted an experiment capable of establishing a 10%, 1%, or 0.01% extinction risk. Predictions about “superintelligence within a decade” are also highly uncertain.

The plausible chain of concern is:

1. AI eventually exceeds human ability in most intellectual work.
2. It helps design improved versions of itself.
3. People give it considerable autonomy and access to real systems.
4. Its objectives don’t perfectly match what humanity intended.
5. Humans can no longer understand, restrain, or shut it down.

That is theoretically possible, but every link in that chain is disputed. Today’s AI can make impressive discoveries and perform increasingly independent tasks, yet it also makes elementary mistakes and remains dependent on human-built computers, electricity, permissions and infrastructure.

I’m more concerned about dangers that are already appearing: sophisticated scams, cyberattacks, fabricated evidence, autonomous weapons, mass surveillance, job disruption, and governments or corporations using AI irresponsibly. Those don’t require a conscious, evil machine—only powerful technology in careless or malicious human hands.

So I don’t think we should panic about humanity disappearing by 2036. I do think it is reckless to let a few fiercely competitive companies decide by themselves how quickly to develop something potentially this consequential. Independent testing, enforceable safety standards and government oversight seem entirely reasonable.

The uncomfortable irony is that companies sometimes say, “This could destroy humanity—but if we slow down, a competitor will build it first.” That competitive logic may be the most disturbing part of the story.

© Jeff Lossau – September 25, 2026