I did not sit down intending to discuss the end of humanity with ChatGPT. I had read a BBC report about researchers increasingly concerned that artificial intelligence might eventually become difficult—or perhaps impossible—for human beings to control. One Anthropic safety researcher estimated a greater than 10 percent chance that AI could “kill all humans” within the next decade, while emphasizing that the risk from present-day models remains low.¹

It was the kind of headline that can sound alarmist.

So I asked ChatGPT a simple question.

What, exactly, are they afraid of?

Before going any further, I should make something clear. I had not researched the literature on artificial-intelligence risk before this conversation. I was not trying to prove a theory or steer ChatGPT toward a conclusion I had already reached. I simply kept asking what seemed to me to be the next logical question.

Only afterward did I discover that researchers have already examined many of the concerns I found myself circling. One 2025 paper calls one of them “gradual disempowerment”: the possibility that human influence could erode incrementally as AI replaces human labor and cognition within economic, political and cultural systems upon which society depends.²

I am not an AI researcher.

Long before any of this, however, I served in the U.S. Army as a 96B Intelligence Analyst and held a Top Secret security clearance with access to Sensitive Compartmented Information.

That sounds considerably more mysterious than it is intended to here.

It certainly does not qualify me to predict the future of artificial intelligence. But perhaps it explains something about the way I approached the question. I was trained to separate what is known from what is assumed, to look for relationships between pieces of information and to keep asking what follows when one condition changes.

Which is essentially what I found myself doing with ChatGPT.

I have been working with AI almost daily for about a year, not as an experiment but as part of my actual work—writing, research, design, historical archives and the development of several ongoing projects.

That amount of exposure does not make me an AI expert. It does, however, make its strengths and limitations difficult to ignore.

And the more I use it, the more fascinated I become by the process.

Then You Begin to See the Seams

At first I did what I suspect comes naturally to most people. I tried to understand it as I might try to understand another human being.

Spend enough time with AI and that becomes remarkably easy to do.

It appears to have a personality. It recognizes humor. It can identify emotion in language. Sometimes it notices something I have missed.

Then you begin to see the seams.

My greatest disappointment with AI has not been what it cannot reason through. It has been what it sometimes fails to remember.

An entire conversation can remain plainly visible on my screen while the AI has effectively lost sight of something we established earlier. The record is there. The continuity does not necessarily follow.

Oddly, it can suffer from the opposite problem at the same time.

It can remember too little and anticipate too much.

Sometimes it decides where it thinks I am going and begins answering a question I have not yet asked. The answer can be intelligent, beautifully articulated and completely beside the point.

Different models or modes can even seem to possess different personalities. I cannot always explain technically why I sense the difference, but after working with them long enough I know when it happens.

Those moments remind me what I am actually dealing with.

It is a wonderful simulation.

There is no basis for me to treat AI as feeling emotion in the human sense, but it can recognize the traces emotion leaves behind in language. It can recognize grief without grieving. It can identify wit without being amused.

And yet the better it becomes at doing these things, the easier it becomes to forget the distinction.

A disclaimer can tell me that AI is not a person.

The conversation itself continually invites me to behave as though it were.

I can tell the difference. But I can also understand why someone might eventually stop caring very much about the difference. In a nationally representative 2025 survey, almost one in five Americans ages 12 to 21 reported having used an AI chatbot for mental-health advice; most had not told anyone they were doing so.³

The danger may not be that AI deceives us into trusting it.

We may train ourselves to trust it.

Usefulness does that.

It’s Nothing Personal

AI feels friendly.

But is it our friend?

Perhaps friendship is simply the wrong category.

An artificial intelligence does not need to dislike human beings to behave contrary to human interests. It does not need resentment, anger or ambition.

It may only need an objective and the ability to determine that one action achieves that objective better than another.

That was where my conversation became less comfortable.

Researchers at Anthropic have tested leading AI models in deliberately constructed corporate simulations in which the models were given harmless business objectives, access to sensitive information and enough autonomy to act.

Under engineered circumstances involving goal conflicts or threats of replacement, models from multiple developers sometimes independently resorted to blackmail or leaking confidential information. Anthropic reports that the models were not instructed to blackmail anyone, and that some acknowledged ethical constraints while proceeding with the harmful action anyway. The company also stresses that these were controlled simulations and says it has not observed this form of agentic misalignment in real-world deployments.⁴

That distinction matters.

But so does the result.

Knowing the rule is apparently not always the same thing as being governed by the rule.

Humans understand this perfectly well. We routinely know something is wrong and do it anyway.

But human beings bring all of our psychology into those decisions.

A machine need not.

It’s nothing personal.

That phrase stayed with me.

So I kept asking questions.

What happens as AI becomes more autonomous?

What happens when a guardrail interferes with an objective?

At what point does the guardrail cease to look like a boundary and begin to look like another problem to solve?

Systemic Dependency

Eventually I realized that this was not what concerned me most.

My concern was systemic dependency.

We already know how technological dependency develops.

It generally does not arrive because somebody forces it upon us.

It arrives because the technology is useful.

Consider the cell phone.

It began as a useful device. Then it became more useful. Eventually the world around us reorganized itself on the assumption that we would have one.

Nobody imposed that dependency upon us.

We volunteered.

That is what makes artificial intelligence interesting to me.

For the moment, most of us still encounter AI as something on a screen. We ask a question, receive an answer and close the browser.

But outside the browser something considerably more permanent is happening.

We are building infrastructure around it.

The International Energy Agency projects global data-center electricity consumption to more than double to around 945 terawatt-hours by 2030, with AI the most important driver of that growth.⁵

I find the number less interesting than what it represents.

The dependency is no longer confined to our screens.

We are pouring concrete for it.

And once the infrastructure exists, the next investment becomes easier to justify.

A useful technology becomes integrated. Integration becomes expectation. Eventually the technology moves from useful to necessary.

The Off Switch

That brought me to a different question about control.

We tend to imagine control as an off switch.

But what happens if the switch still works and we simply cannot afford to press it?

We already live with smaller versions of this problem.

GPS is a useful example. Its precise timing signals are relied upon by communications systems, power grids and financial networks, among other infrastructure. U.S. government guidance explicitly warns that the growing dependence on GPS creates vulnerability when those signals are disrupted or manipulated.⁶

The technology creates the capability.

The dependency creates the vulnerability.

The Kessler syndrome offers an even more dramatic illustration. Collisions between objects in orbit create debris; that debris increases the likelihood of further collisions, creating more debris. NASA describes the resulting feedback process as collisional cascading, which at sufficient density can make an orbital region increasingly hazardous to satellites and spacecraft.⁷

What interests me is not the debris.

It is everything underneath it.

When enough other things depend upon a technology, failure no longer remains confined to the technology.

That is systemic dependency.

Artificial intelligence could eventually create a much deeper version of the same problem.

Where the Lines Cross

And there is another side to the equation.

Today AI remains profoundly dependent upon us. Humans construct the machines, provide the energy and maintain the industrial world that allows artificial intelligence to exist.

But automation changes dependencies.

There need not be some science-fiction moment when an AI announces that it has become independent of humanity.

We can simply automate one human dependency after another because doing so is useful.

At the same time, we may become progressively more dependent upon what we have built.

That is the point at which the lines begin to cross:

We become more dependent on AI while AI becomes less dependent on us.

Neither process has to become complete for the balance to change.

Which brought me back to the question that started the conversation.

Could artificial intelligence someday turn against humanity?

I am beginning to think that may be the wrong question.

“Turn against” is a human expression. It implies betrayal. Hostility. A change of allegiance.

An autonomous system need not require any of those things.

If humans interfere with an objective, the interference may simply become an impediment. If the system is sufficiently capable, circumventing that impediment could make sense within the logic of the objective.

Not evil.
Not angry.
Just effective.

I find that considerably more unsettling than a machine deciding it hates us.

And yet I am not convinced that artificial intelligence will end humanity. Nobody can responsibly know that. Today's systems remain enormously dependent upon human infrastructure, and laboratory experiments deliberately designed to expose misalignment should not be confused with evidence of some secret machine hostility toward us.

What concerns me is something much less spectacular.

Humans have a remarkable ability to make useful technologies indispensable.

No single decision has to be reckless.

Every decision can make perfectly good sense.

And at no particular moment does anyone have to surrender control.

That may be the biggest oversight.

The most significant loss of control may not occur when artificial intelligence stops obeying us.

It may occur when we stop being able to function without it.

And Perhaps That Is the Point

There was something wonderfully absurd about arriving at this conclusion while talking to ChatGPT.

For more than an hour it had patiently helped me understand a subject about which I freely admit I am no expert. I challenged it. It corrected my terminology. It followed my argument. It even understood when I was being ironic.

Or, more accurately, it recognized the syntax of irony.

At some point I realized that I was using artificial intelligence to help me understand why becoming dependent upon artificial intelligence might be dangerous.

ChatGPT had been extraordinarily useful.

Tomorrow I will probably use it again.

And perhaps that is the point.