What Is Happening in the AI Debate Right Now – My Attempt at Making Sense of It
Artificial intelligence is back in the main news. And the tone has changed.
People like Sam Altman of OpenAI and Dario Amodei of Anthropic are talking about the further development of artificial intelligence with a sense of urgency that is hard to miss. They are talking about risks, control, alignment, and increasingly about whether development at the frontier may need to slow down.
That can be unsettling.
The very people whose companies have made these systems increasingly powerful are now warning that their further development could become dangerous. The obvious story is quickly told: first they built it, and now they themselves are becoming afraid of what they have created.
Maybe that is what is happening.
I find another, less dramatic explanation at least equally plausible. It is not, however, a reassurance.
I am neither part of one of the major AI labs nor an alignment researcher. I have no privileged information about what is happening inside these organizations. What I am attempting here is therefore exactly that: an attempt at making sense of what is publicly visible. How do I understand, in September 2026, what we are currently observing?
A principle that has guided my own work for a long time helps me think about it:
Do first. Then understand immediately, and change immediately if necessary.
The “do first” matters.
Some things can only be understood after we have done them
Especially in Europe, we tend to look at the American technology industry with a certain degree of skepticism. “Move fast,” “fail fast,” and particularly Mark Zuckerberg’s famous “move fast and break things” can easily sound like the expression of a culture that acts first and thinks about the consequences afterwards.
There is something to that criticism. But I think it is too simple.
When something is genuinely new, we cannot understand everything in advance.
We can spend years developing models, analyzing risks and discussing possible consequences. At some point, somebody has to build something, try it and observe what happens. Some knowledge only becomes available through engaging with something that actually exists.
That is true of software. It is true of organizations. And it is particularly true of today’s AI systems.
So for me, the problem is not doing something before fully understanding it.
The crucial parts of my old principle are the two inconspicuous words:
understand immediately and change immediately.
How large can the gap become?
Doing, understanding, correcting – it is a loop.
I do something. In doing so, I learn something I could not have known before. That new understanding may change my model. And if the model no longer fits what I have done, I change it.
As long as this loop remains tight, “do first” can be an entirely reasonable way of learning.
It becomes problematic when the individual steps begin to drift apart.
We keep doing while understanding falls behind.
Eventually, we understand that an earlier assumption was wrong, but by then so many other decisions depend on it that changing it has become expensive.
Or we recognize a problem only when correction is barely possible anymore.
We know this well from ordinary software. A small design decision goes unquestioned, other components are built around it, users become accustomed to it, data accumulates. Years later, we discover that the original assumption was not particularly good.
Then we call it technical debt.
Usually, that is annoying. Rarely is it a catastrophe.
With artificial intelligence, the same question is now being asked at a different scale.
Perhaps this is exactly the question we are hearing now
I do not know what Sam Altman or Dario Amodei know internally. Nor do I want to psychoanalyze their public statements.
But I do not necessarily have to interpret their current sense of urgency as:
“We have lost control of our systems.”
I can also read it differently:
“We have to make sure that our doing does not run too far ahead of our understanding.”
That is a much less spectacular statement.
I would still take it very seriously.
The development of powerful AI has become extremely fast over the past few years. At the same time, new capabilities need to be tested, their consequences understood, safety mechanisms developed, and societal rules established.
Perhaps what we are seeing, then, is not the sudden discovery of a previously unknown danger.
Perhaps what is changing is the assessment of how large a gap between development and understanding we can still afford.
That would also explain why the question of pace is becoming increasingly prominent.
Not necessarily because “slower” is inherently safer.
But because understanding takes time.
What is special about AI is not that we do not understand everything
There are many things we do not fully understand.
We operate economies even though nobody can reliably predict how they will develop. We develop medicines whose effects only become fully visible through studies. We build complex software systems whose actual behavior no single person can understand in every detail.
Complete understanding has never been a prerequisite for human action.
But AI adds an uncomfortable question:
How certain are we that we will still be able to correct things later?
For me, this question captures a significant part of the alignment problem.
If a new software release is bad, we can roll it back.
If an architectural decision turns out to be wrong, a migration may be expensive. But usually, we can migrate.
With increasingly powerful AI systems, the discussion is about whether there may eventually be capabilities and forms of deployment for which we can no longer take this kind of reversibility for granted.
I cannot judge how realistic the most dramatic of these scenarios are.
But we do not have to regard them as certain in order to consider the underlying question reasonable.
As the potential consequences of an error become larger, I need to ask earlier whether my established method of learning through correction will continue to work.
That is not AI panic.
It is a fairly ordinary question of reasonable action.
Perhaps we should also lower the temperature of the cultural debate
Europe and Silicon Valley are often portrayed as opposites.
Here: precaution, regulation and skepticism.
There: speed, experimentation and innovation.
There is truth in that distinction too. But perhaps it is not particularly helpful for understanding AI right now.
The European idea that we should first understand a fundamentally new technology before developing it runs into an epistemic limit: some of that understanding can only emerge through development itself.
The American idea that we can develop quickly and correct problems afterwards runs into a different limit: it only works reliably as long as correction afterwards remains possible.
Perhaps what we are seeing now is both limits becoming visible at the same time.
That would be less dramatic than the story of AI developers suddenly becoming afraid of their own creation.
But it would by no means be insignificant.
What I want to record in September 2026
Perhaps, a few years from now, we will look back at today’s warnings and consider them exaggerated.
Perhaps we will wonder why we did not take them more seriously.
I do not know.
And that is precisely why I do not want to pretend that I can already tell the correct story about this period today.
What I can record is my current understanding.
I do not see people like Altman and Amodei suddenly discovering that artificial intelligence might be dangerous. That discussion has been going on for years.
But I do see signs that something in the weighting is changing.
The question is no longer only:
What can we do next?
Alongside it, a second question is becoming louder:
Are we understanding fast enough what we are doing?
And behind it stands a third, perhaps decisive question:
If our understanding changes – can we still change what we are doing in time?
Perhaps that is a useful translation of what is currently making its way through the news in rather alarming terms.
It takes some of the drama out of the debate.
It does not make it any less important.
Comments
Write a comment