Response to the CNN Article AI Will Off Humanity

Random Picture of Toy AI Project

There has been a lot of discussion lately about artificial intelligence ending humanity. Maybe it happens in one year. Maybe it happens in ten. Depending on whom you ask, there is some percentage chance that AI eventually wipes us out.

I am not going to lean completely one way or the other. I am an engineer, not an AI safety researcher, but I do have some thoughts about the logic behind that proposition.

A recent CNN article gave one of the more direct versions of the argument:

https://www.cnn.com/2026/09/17/tech/how-will-ai-exterminate-humanity-cec

One thing I continually hear, even when reporters interview leaders of major AI companies, is some version of this question:

"Could AI destroy humanity?"

And sometimes the answer is essentially: yes, it could, possibly within some number of years.

What I don't hear discussed nearly enough is how. That seems like an important part of the argument.

If someone says there is a meaningful probability that AI will exterminate humanity within ten years, I want to see the chain of events that gets us there. What capabilities does the AI need? What infrastructure does it control? What physical systems does it interact with? What humans cooperate with it? What safeguards fail?

Most importantly: how does the scenario go from "very dangerous" to "human extinction"?

What About Humanity Destroying Humanity?

There are more than eight billion human beings on Earth.

Humans already possess an extraordinary collection of dangerous technologies: nuclear weapons, biological agents, chemical weapons, cyberwarfare capabilities, autonomous and semi-autonomous weapons, drones, and conventional military systems.

Humans also kill other humans on a large scale.

The United Nations Office on Drugs and Crime estimated approximately 458,000 victims of intentional homicide worldwide in 2021.

The world population is approximately 8.3 billion people.

There are about roughly 6.1 billion people are adults. 

There are about 3.7 billion to 4 billion people

There were 20,162 recorded homicides in the U.S. in 2024 

There are about 400,000 to 458,000 people are victims of intentional homicide globally 

That is horrifying, but it also illustrates something important about scale.

Humans are unfortunately quite capable of killing humans, yet even with billions of people, organized militaries, governments, terrorist organizations, criminal organizations, weapons and centuries of warfare, exterminating humanity is an entirely different problem.

It reminds of when I was looking at the history of the Ebola virus and even then I think we had the opposite problem, top scientists yelled that there could be a major issue and no one listened.  Even today, there are outbreaks happen, not getting a lot of international attention.  From what I understand the USAMRIID has a repository of various viruses that if released could be very dangerous.  There are strick physical protocols for who handles what.  

https://en.wikipedia.org/wiki/United_States_Army_Medical_Research_Institute_of_Infectious_Diseases

There are also ways that people could attack critical infrastructure using technology that exists today. Power grids, communications networks, transportation systems and financial infrastructure are obvious targets.

These are serious threats. But "serious threat to society" and "extinction of humanity" are very different claims. That distinction matters.

What About AI?

Going back to the CNN article, one scenario involves AI helping someone develop or deploy a deadly biological agent.

That deserves serious consideration. But again, I want to understand the mechanism.

Dangerous biological agents already exist. Governments and research institutions already maintain high-security facilities, physical containment procedures and strict controls around certain pathogens.

An AI potentially lowering the knowledge barrier for biological attacks is a legitimate security concern. But even in a scenario where an AI tricks or assists someone, a long chain of things still has to happen.

Someone needs access to materials. Someone needs equipment. Someone needs laboratory capabilities.

Physical security systems have to fail. Monitoring systems have to fail. Other humans have to fail to notice what is happening.

And finally, whatever is created has to be capable of causing damage on a civilization-ending or species-ending scale.

Maybe such a chain is possible. I don't dismiss it.

But the chain itself should be part of the discussion.

We already live in a world containing dangerous biological materials. AI doesn't have to invent danger for danger to exist.

Why Ten Years? Why Ten Percent?

This is another part of the AI-extinction discussion that bothers me.

Why ten years? Why not eight? Why not six?

And when someone says there is a 10% probability of catastrophe, where does that number come from?

Why 10% instead of 8% or 17%?

Predictions about rapidly developing technologies are notoriously difficult.

OpenAI was founded in 2015. Anthropic was founded in 2021. The modern generative-AI industry itself has transformed dramatically in only a few years.

Go back ten years and ask how accurately even experts could have described the AI industry of 2026.

Could they have predicted today's companies, products, compute requirements, capabilities, limitations and economics with enough precision to assign reliable probabilities to them?

Maybe some could have predicted the general direction. That's different from predicting specific outcomes.

This doesn't mean long-term AI predictions are worthless. Companies absolutely should create five- and ten-year projections. Researchers should think about technological trends, capabilities and risks.

But there is a major difference between:

"AI capabilities will probably continue improving."

and:

"There is a 10% probability AI will exterminate humanity within ten years."

The second statement sounds much more mathematically precise than our knowledge of the future may justify.

That doesn't automatically make it wrong. It does make me want to know how that probability was calculated.

Autonomous Weapons Worry Me More

The article also discusses AI-controlled weapons and robots.

This scenario concerns me more because we can already see pieces of it happening.

Modern warfare has demonstrated how important drones have become. Russia and Ukraine have extensively used drones for surveillance and attacks. Now imagine increasingly autonomous systems identifying, tracking and attacking targets with less human involvement.

That doesn't require science fiction.

The technological progression is fairly easy to imagine because many of the pieces already exist.

AI-controlled physical weapons could potentially produce enormous casualties.

But even here, there is another leap between "AI makes warfare substantially more dangerous" and "AI kills every human being."

I want to understand that leap.

AI Takes Over the Internet

Another scenario is an AI escaping onto the Internet, replicating itself and operating independently.

This one raises a different set of questions for me. Replicate itself where?

Modern frontier AI systems require substantial computing infrastructure. They don't simply exist as a magical program floating around the Internet.

Could a sophisticated AI compromise servers, distribute components of itself and create something resembling a botnet?

What does it do? Create ad-supported websites? Send spam? Run cyberattacks? Steal computing resources?  Those things could certainly cause enormous economic and security problems. But again, we have moved from "AI exterminates humanity" to something more recognizable: an extremely sophisticated cyberattack.

I also find the compute question fascinating.

We regularly hear about AI companies building enormous data centers and acquiring huge numbers of GPUs because frontier models require tremendous computational resources.

Then, in the extinction scenario, the AI apparently escapes and becomes completely independent.

Where is it running? My Chromebook? My personal WordPress server?

If an AI can somehow transform millions of ordinary computers into a giant distributed intelligence, that itself would be one of the most remarkable technological breakthroughs in computer science.

I would like to understand that mechanism before jumping to the next step of the argument.

When Prediction Starts Looking Like Faith

This is where my skepticism really comes from.

Some AI-extinction discussions remind me of historical doomsday predictions—not because AI researchers are religious or because AI risks aren't real, but because the reasoning can sometimes become difficult to falsify.

Faith-based reasoning is a cognitive or philosophical approach where conclusions are derived from core religious or spiritual beliefs ("faith") rather than purely empirical evidence or secular logic

Predicting the end of the world without empirical evidence is a classic, extreme manifestation of faith-based reasoning, often driven by deep-seated psychological and sociological needs. 

But connecting imaginable events together does not automatically make the entire sequence probable.

The farther into the chain we go, the more assumptions we accumulate.

That is where I think discussions about existential AI risk need to become more concrete rather than less concrete.

Show me the mechanism, Show me the dependencies. Show me the failure points.

And show me how we get from a dangerous AI incident to eight billion people disappearing.

The Current State of AI

Maybe part of my skepticism comes from using this technology regularly.

I think large language models are one of the most interesting technologies I have encountered.  But I am still somewhat unimpressed by the gap between today's systems and some of the claims being made about tomorrow's systems.  Google and Yahoo search also seemed incredible when I first encountered Internet search in the 1990s. Technologies can be revolutionary without immediately becoming omnipotent.

There are still things I am waiting to see from AI:

• AI providing dramatically better automated support during IVR/customer-service calls.
• AI independently building a substantial, production-quality application that people actually use.
• A genuinely new operating system or major hardware platform substantially designed by AI.
• AI making a major breakthrough toward curing cancer or another major disease.
• AI meaningfully improving something messy and physical, such as urban traffic.
• AI systems that operate independently for long periods without constantly depending on human prompts, infrastructure and supervision.
• And while we're at it, fix the Windows blue screen and the spinning beach ball on my Mac.

I expect AI to improve dramatically.  I expect it to change software development, science, business, warfare and probably many things I haven't considered. And some of those developments could absolutely be dangerous. But extraordinary predictions deserve detailed mechanisms. If AI has a meaningful chance of exterminating humanity, don't just give me the percentage.

Document Cleaned up Open AI - Chat GPT