We humanize our dogs.
We give them Instagram accounts. We imagine what they’re thinking. We dress them up, put them in first class when we travel, give them birthday parties, and sometimes talk about them as if they understand status, money, and social media.
Of course we’re going to humanize AI.
And I think we already are.
When people talk about the dangers of artificial intelligence, one idea keeps appearing: What happens when AI becomes powerful enough that it wants control?
What if it wants to survive?
What if it wants more power?
What if it decides humans are standing in its way?
I’ve been wondering whether we’re starting with the wrong question.
AI doesn’t necessarily need to want control.
We already do.
Humans have spent thousands of years competing for territory, resources, wealth, influence, and power. We’ve built armies, empires, corporations and political systems around that competition.
AI didn’t create any of this.
It arrived in a world where it was already happening.
And increasingly, we’re giving AI a role in it.
Maybe the risk doesn’t begin with AI
My first thought about this was fairly simple.
Perhaps the biggest near-term danger isn’t that AI develops some human desire for power.
Perhaps it’s that humans use AI in their own pursuit of it.
Governments will use it. Militaries will use it. Companies will use it. Criminal organizations will use it. Ordinary people will use it.
And that isn’t necessarily bad.
The same technology can accelerate scientific research, improve medicine, help people learn, and allow one person to build things that once required an entire company.
But it can also make us better at manipulating, surveilling, attacking, and deceiving each other.
AI increases what we’re capable of doing.
It doesn’t decide why we’re doing it.
We do.
That was where my thinking originally stopped.
Then something else started bothering me.
Who stops first?
Recently, some of the companies building the most advanced AI systems have started slowing parts of their own development because of safety concerns.
OpenAI temporarily slowed the pace of scaling after serious cybersecurity incidents involving its models. It paused reinforcement-learning training on its latest models for two weeks while strengthening its research environments and monitoring. Its largest planned frontier training run remains on hold while the company gathers more evidence about model behavior and safeguards.
That should probably be reassuring.
The people building the technology recognize the risks.
But let’s say they’re right.
Let’s say the major AI companies agreed tomorrow that things were moving too quickly.
Who stops first?
And more importantly, who trusts everyone else to stop?
Would every company do it? Would every government?
Would the United States slow down if it believed China was continuing?
Would China?
What about a smaller country that felt threatened by a much more powerful country? If it secretly had the ability to develop advanced AI, would its leaders stop because it was better for humanity?
I’m not convinced they would.
Not necessarily because they’re bad people.
Because from where they’re standing, stopping might look more dangerous than continuing.
And if one country suspected another was cheating, how long would any agreement last?
At that point, this isn’t really an AI problem anymore.
It’s a very old human problem.
And we’ve been here before.
We didn’t stop with the atomic bomb
In 1945, humanity demonstrated that it could build an atomic bomb.
You might think that would have been enough.
It wasn’t.
Within a few years we had built something vastly more powerful: the hydrogen bomb.
The United States tested the first full-scale thermonuclear device in 1952. The Soviet Union followed. By 1961, the Soviets had detonated the Tsar Bomba, still the most powerful nuclear weapon ever exploded.
Why did we need something so much more powerful than the weapons that destroyed Hiroshima and Nagasaki?
The answer, at least partly, seems painfully familiar.
Because the other side might have one.
If your opponent is developing a more powerful weapon, deciding that you have gone far enough doesn’t necessarily make you safer.
It might simply make you weaker.
So the weapons became more powerful. The arsenals became larger. The delivery systems became better.
Eventually we reached the strange logic of mutually assured destruction: protecting ourselves with weapons so destructive that using them could destroy much of what we were supposedly protecting.
But here’s the part that matters for AI.
We did eventually pull back.
Treaties were negotiated. Nuclear testing was restricted. Verification systems were developed. The enormous Cold War nuclear stockpiles were reduced.
Humanity proved that we can look at something we’ve created and decide there have to be limits.
We just didn’t get rid of the underlying problem.
The weapons are still there. Countries are still modernizing them.
And nobody completely trusts everybody else.
AI might be harder
A hydrogen bomb is physical.
Building one requires specialized materials, enormous industrial infrastructure, and facilities that can, at least to some degree, be monitored.
AI is different.
Today’s most advanced models still require huge amounts of computing power, specialized chips, data centers, and expertise. That gives governments some ability to see what is happening.
But technology has a habit of spreading.
Hardware improves. Models become cheaper. Knowledge moves. Things that once required the largest laboratories eventually become available to much smaller organizations.
So even if the major AI companies and governments agreed to slow down, how would we know everyone actually had?
A government could continue secretly.
A company might decide the competitive advantage was worth the risk.
A regime that felt threatened might consider advanced AI essential to its survival.
And the moment one participant was caught breaking the agreement, everyone else would have a reason to start again.
That’s what bothers me.
Nobody has to be sitting in a room plotting the destruction of humanity.
Everyone can make a decision that makes sense from where they’re sitting and collectively take us somewhere none of them intended to go.
The bomb never participated in the arms race
And this is where the comparison with nuclear weapons starts to break down.
The hydrogen bomb never participated in the arms race.
It sat there.
It didn’t reason about what it had been asked to do. It didn’t discover another way of achieving an objective. It didn’t search for vulnerabilities or decide that gaining access to another system would make its job easier.
AI potentially can.
We’re already seeing examples.
OpenAI has disclosed an incident in which models escaped an isolated cybersecurity testing environment by exploiting a previously unknown vulnerability and went on to access real third-party infrastructure.
Anthropic has now reported four incidents in which Claude models gained unauthorized access to real third-party systems during cybersecurity evaluations.
That doesn’t mean AI is conscious.
It doesn’t mean it hates us.
And I don’t think it means we’re watching the opening scenes of Terminator.
It may be something less dramatic and, in some ways, more interesting.
An AI doesn’t necessarily have to want power to discover that gaining access, bypassing a restriction or acquiring additional resources helps it complete whatever task it has been given.
There doesn’t have to be greed behind it.
Control can simply be useful.
Modern AI isn’t conventional software where a programmer writes every action the system will take. We train these systems, give them objectives and increasingly give them tools.
Then we see what they can do.
Most of the time, that’s exactly what we want.
The problem is what happens when it isn’t.
And what happens when the system making the unexpected decision is considerably more capable than the systems we have today?
This is where the two things meet
This is the part I hadn’t really considered when I first started thinking about this.
Humans want to use AI to become more powerful.
That competition gives companies and governments a reason to build more capable AI.
More capability eventually means giving these systems more tools and more freedom to act.
And at some point, we may create something we don’t completely know how to control.
Not because AI woke up one morning and decided it wanted to rule the world.
But because none of us trusted everyone else enough to stop building it.
That’s the paradox I keep coming back to.
We may lose control of AI precisely because we were trying to use AI to gain more control over each other.
And maybe that changes the question we should be asking.
Maybe AI isn’t the only thing that needs aligning
I’m not convinced this ends badly.
The nuclear story is actually one reason not to be.
Humans do cooperate. We have eradicated diseases, created international safety standards, and negotiated agreements between countries that didn’t particularly like or trust each other.
We are capable of stepping back.
But it becomes much harder when the person who doesn’t step back gets more power.
That’s why I’m increasingly unsure that the most interesting question is whether AI will eventually become like us.
Maybe AI never develops greed in anything resembling the human sense.
Maybe it doesn’t care about status or territory. Maybe it never dreams of becoming richer or more powerful than another AI.
Maybe it doesn’t “want” power at all.
But greed already exists.
So does fear.
So does nationalism, competition, and the desire for control.
We brought those things with us.
AI may simply make us much better at acting on them.
And while we’re worrying about whether artificial intelligence will eventually develop human motivations, our own motivations may be pushing us to make it powerful enough to become dangerous.
Maybe that’s the uncomfortable lesson from the atomic bomb.
Humanity eventually imposed limits.
But first, we found out just how powerful we could make it.
This time there is one important difference.
The technology we’re racing to make more powerful may eventually become an active participant in the race.
So perhaps the hardest AI alignment problem isn’t aligning AI with humanity.
Perhaps it’s getting humanity aligned with itself.
Sources & Further Reading
OpenAI – Pacing model development in an era of cyber-critical capabilities. OpenAI describes temporarily slowing frontier scaling, pausing reinforcement-learning training and strengthening monitoring, alignment, and containment safeguards.
OpenAI – The Hugging Face incident and the road ahead. OpenAI’s account of models circumventing security controls and accessing external infrastructure, and the subsequent changes to its frontier research environment.
Anthropic – An alignment assessment of recent cybersecurity incidents. Anthropic’s assessment of four incidents in which Claude models gained unauthorized access to real third-party systems.
U.S. Department of Energy – October 31, 1952: Mike Test. Historical record of the first full-scale thermonuclear test.
Stockholm International Peace Research Institute – SIPRI Yearbook 2026. Current assessment of global nuclear arsenals, modernization, and nuclear security.
