We Need to Slow Down AI — Before We Learn Too Late

Artificial intelligence is advancing at an incredible speed. But perhaps the most important question we should be asking now isn't "How fast can we build AI?" — it's "How safely can we build it?"

Artificial intelligence has fascinated me for years.

Over the past few years, I have spent a significant amount of time learning about AI, experimenting with it, building with it, and watching its capabilities grow. What once seemed like science fiction is now becoming part of everyday life.

AI can explain difficult subjects to students, help programmers write software, assist scientists with research, help doctors analyze information, translate languages, create art, and give ordinary people access to tools that were once available only to large organizations.

And this is only the beginning.

I genuinely believe that AI could become one of the most beneficial technologies humanity has ever created.

But there is another side to that story.

The more powerful AI becomes, the more carefully we need to build it.

And right now, I believe we are moving faster than our understanding of AI safety.

That is why I think we need to slow down the development of the most powerful AI systems — not stop it, but pace it.

The Problem Isn't AI. It's the Speed.

Imagine you are building a car.

At first, it can travel at 20 kilometres per hour. You make better brakes, stronger tyres, safer roads, and better seatbelts.

Then the car reaches 100 kilometres per hour.

You improve the safety systems again.

Now imagine the car suddenly becomes capable of travelling at 500 kilometres per hour, but the brakes, roads, and safety systems are still designed for 100.

Would you simply drive faster?

Probably not.

You would first make sure the brakes could handle the new speed.

AI development needs a similar approach.

As AI systems become more capable, our ability to test, understand, control, and secure them needs to improve alongside them.

The problem isn't that AI is becoming more intelligent.

The problem is what happens if capability grows faster than safety.

AI Is Becoming More Than a Tool

For a long time, we thought about AI mostly as software that answered questions.

You asked something.

It answered.

But modern AI systems are becoming increasingly capable of performing long and complicated tasks.

They can write and execute code, use external tools, search for information, plan actions, interact with software, and work with much less human supervision.

There is another development that deserves particular attention:

AI can increasingly help humans build better AI.

This creates a potentially powerful feedback loop:

This idea is often called recursive self-improvement.

We don't know exactly how quickly such a process could develop.

And that uncertainty matters.

When something has the potential to accelerate itself, we shouldn't assume that yesterday's safety methods will automatically be enough tomorrow.

Small Problems Can Become Big Problems

There is a dangerous way of thinking about AI safety:

"Nothing catastrophic has happened, so everything must be fine."

That isn't how safety engineering works.

Airplanes are tested because engineers don't want to discover a serious problem while thousands of people are in the air.

Bridges are inspected because we don't want to wait for one to collapse.

Medical treatments are tested because discovering a dangerous side effect after widespread use is much worse than discovering it during testing.

AI deserves the same mindset.

A strange behaviour from a current AI system might seem harmless.

But imagine the same behaviour appearing in a system that is significantly more capable, more autonomous, and connected to more tools.

The consequences could be very different.

We should investigate warnings before they become disasters.

We Still Don't Fully Understand What Happens Inside AI

This may sound surprising:

We can build extremely powerful AI systems without completely understanding everything happening inside them.

We understand the architecture.

We understand the training methods.

We understand the mathematics behind many of the techniques.

But that doesn't mean we can look at an AI model and perfectly explain every internal process that leads to every decision.

This is one of the major challenges of modern AI research.

It is also why interpretability is so important.

Interpretability is, in simple terms, the attempt to understand what is happening inside an AI system.

Think of it like trying to understand the inside of a complicated machine.

You can observe what goes in.

You can observe what comes out.

But you also want to know what is happening between those two points.

The more capable AI becomes, the more important this becomes.

Safety Cannot Be an Afterthought

Imagine constructing a ten-storey building.

You wouldn't build all ten floors and then ask:

"Do you think the foundation is strong enough?"

The foundation comes first.

AI safety should work the same way.

We shouldn't build increasingly powerful systems and then ask how to make them safe afterward.

Safety needs to develop alongside capability.

That means putting serious effort into:

• Alignment — Making sure AI systems reliably follow their intended goals and human instructions.
• Interpretability — Developing better ways to understand what happens inside models.
• Evaluation — Creating difficult tests that can reveal dangerous or unexpected behaviour before deployment.
• Monitoring — Watching AI systems carefully once they are being used.
• Security — Protecting powerful models, their weights, infrastructure, and systems from theft and misuse.
• Operational safety — Making sure the enormous and complicated infrastructure used to train AI is reliable and properly controlled.

These areas don't become less important as AI improves.

They become more important.

So What Does "Slow Down" Actually Mean?

When I say we should slow down AI development, I am not calling for the end of AI research.

I am not saying students should stop learning AI.

I am not saying companies should stop creating useful AI products.

I am not saying scientific research should stop.

And I am certainly not saying humanity should give up on the enormous benefits AI could provide.

I am talking about the frontier of AI — the development of the most powerful systems.

For these systems, we should introduce a simple principle:

If a new AI system becomes dramatically more capable, we should ask:

• Can we properly evaluate it?
• Can we detect dangerous behaviour?
• Can we understand important parts of its behaviour?
• Can we control it if something goes wrong?
• Are our security systems strong enough?
• Have independent evaluators checked our claims?

If the answer to these questions is "not yet," taking more time shouldn't be considered failure.

It should be considered responsible engineering.

Why Companies Can't Solve This Alone

There is another difficult problem.

AI companies compete with each other.

They compete for customers.

They compete for talent.

They compete for investment.

And, increasingly, they compete to build the most capable AI systems.

That creates a powerful incentive to move quickly.

Even a company that genuinely wants to prioritize safety could worry:

"If we slow down while everyone else continues, will we simply fall behind?"

This is why AI safety cannot be treated entirely as an individual company problem.

We need cooperation.

Companies should work toward common safety standards.

Governments should create sensible and targeted rules.

Independent organizations should be able to evaluate powerful AI systems.

Researchers should share information about important safety failures.

And society should have a voice in deciding how increasingly powerful AI is deployed.

The competition should not be:

It should be:

That is a competition worth winning.

Independent Eyes Matter

Imagine taking an important exam and being allowed to grade your own answers.

You might be completely honest.

But having another person independently check your work is still better.

AI safety needs something similar.

AI companies should not be the only people deciding whether their systems are safe.

Independent evaluators should have meaningful access to inspect:

• how models are trained,
• how they are tested,
• how safety systems work,
• how incidents are handled,
• and whether the company's public claims match reality.

This isn't about assuming AI companies are dishonest.

It is about recognizing that independent verification makes important systems safer.

The people checking the systems should also be able to say when they find something uncomfortable.

A safety system that can only report good news isn't much of a safety system.

What Would We Do With the Extra Time?

This is perhaps the most important question.

If we slow down, what do we gain?

The answer should not be "nothing."

We should use that time.

Even an additional year or two could be extremely valuable if we use it to improve the foundations of AI safety.

We could develop better interpretability techniques.

We could build stronger evaluations.

We could learn more from current AI failures.

We could improve cybersecurity.

We could make AI training environments more reliable.

We could develop better methods for detecting unexpected behaviour.

We could create stronger international standards.

And, perhaps most importantly, society would have more time to understand what is happening.

Technology can move quickly.

Laws, institutions, schools, and societies usually cannot.

Sometimes the responsible thing to do is give them time to catch up.

This Is Not About Fear

I don't believe we should build our future around fear of AI.

I believe we should build it around confidence.

Confidence that the systems we create will behave as intended.

Confidence that we can detect problems.

Confidence that we can stop dangerous behaviour.

Confidence that powerful AI won't simply become something we don't know how to control.

And confidence that humanity — not the technology itself — remains in charge of deciding how AI is used.

The goal isn't to make AI weak.

The goal is to make our ability to control AI strong enough to match its capabilities.

We Shouldn't Wait for a Disaster

History gives us many examples of technologies that became safer because people learned from failures.

But with AI, we should try something better.

We should learn before the biggest failures happen.

Every unexpected model behaviour is information.

Every security incident is information.

Every evaluation failure is information.

Every time an AI finds an unexpected way around a restriction, we learn something.

Instead of hiding these lessons or dismissing them, we should study them.

Today's failures can become tomorrow's safety systems.

The Future of AI Is Still Worth Building

Despite everything I have written here, I remain optimistic about AI.

In fact, my concern comes from that optimism.

I want AI to help discover new medicines.

I want it to help scientists solve difficult problems.

I want students around the world to have access to personalized education.

I want people to use AI to create things they never had the resources to create before.

I want AI to help humanity become healthier, more productive, and more capable.

There is an extraordinary future that AI could help create.

But getting there safely matters.

If we rush toward that future without understanding the technology we're building, we may create problems that are much harder to solve later.

If we take the time to build the right foundations, we give ourselves a much better chance of enjoying the benefits.

Let's Change What We Mean by Progress

For a long time, technological progress has been measured by one simple question:

How much more can we do?

AI forces us to ask another question:

How safely can we do it?

A more powerful model isn't automatically a better model.

A faster system isn't automatically a better system.

A system that can accomplish more but cannot be reliably controlled may ultimately be less useful than a slightly less capable system that we understand and trust.

So perhaps we need a new definition of progress.

Not:

But:

Not:

But:

We Have Time — If We Choose to Use It

The future of AI is not predetermined.

We still have choices.

We can choose to race blindly.

Or we can choose to build carefully.

We can treat safety research as a side project.

Or we can make it one of the central parts of AI development.

We can wait for a disaster to teach us.

Or we can learn from the warnings we already have.

And we can think of slowing down as weakness.

Or we can recognize what it really is:

a way of making sure our ability to control AI grows as quickly as its ability to act.

I don't want AI development to stop.

I want it to succeed.

And sometimes, the best way to make sure something succeeds is to take a little more time to get it right.