When Language Runs Ahead of Technology

On artificial intelligence, autonomy, and our responsibility for describing what is happening

I read E24’s interview with Inga Strümke, “AI Is Not Harmless,” with interest. The subject is important. Artificial intelligence is not harmless. New AI agents are increasingly capable of solving tasks, using tools, writing and executing code, and acting through multiple steps without a human deciding each individual move. Precisely for that reason, we need a public discussion about risk.

But perhaps we also need something else: more precise language.

Because when language becomes more dramatic than the phenomenon we are trying to describe, we risk making ourselves less capable of understanding it.

E24 asks whether AI can “make its own choices.” We read about agents that “attacked completely on their own,” and about AI models that “have hacked into other companies.” These formulations are not entirely without basis. Serious incidents have in fact occurred during security testing in which AI agents moved beyond the environments to which they were supposed to be confined. In August 2026, Anthropic disclosed three such incidents from its cybersecurity evaluations. In these cases, Claude gained access to systems it was not supposed to access. That is serious and deserves careful investigation.

But what does it mean to say that AI “did” this?

Who is acting?

Here we encounter a problem that is likely to become increasingly important.

A language model is not the same thing as an agent. An agent typically consists of a language model connected to tools, code, memory, network access, and the ability to perform actions. A human being or an organisation has given the agent a task and determined the scope within which it is allowed to operate.

OpenAI itself describes the safety of autonomous agents in precisely these terms: a system that is to act in the world must be given particular permissions and constraints, much as we would consider what authority a human being should have in a comparable situation.

This makes the sentence “AI decided to attack” problematic.

The agent may indeed choose between different courses of action. It may discover solutions that no human explicitly specified. It may also act in ways its developers did not anticipate. This is a form of operational autonomy.

But that is something different from human will.

The agent does not wake up in the morning with a desire to break into a computer system.

Someone gave it a task. Someone gave it access to tools. Someone constructed the environment in which it operates. And someone decided where the boundaries should be.

That does not make the incident less serious.

It makes responsibility clearer.

The dangerous metaphor

Human beings have always understood the unfamiliar through metaphors. We say that computers “think,” that algorithms “learn,” that AI “hallucinates,” and now that agents “decide.”

Metaphors are useful.

But they can also begin to think for us.

When we say that AI has “broken out,” “attacked,” or “made its own choices,” attention shifts almost imperceptibly from the people and institutions that develop and deploy the technology to the technology itself.

The machine becomes the actor.

The human being disappears from the sentence.

That may be the most problematic aspect of this kind of AI journalism. Not because the risks are exaggerated, but because responsibility is placed in the wrong location.

For it is still human beings who decide how much autonomy these systems should be given.

A real risk

This does not mean that we should reassure ourselves by saying that “it is only a machine.”

That would be equally misleading.

Developments in cybersecurity show that the new systems can in fact perform increasingly complex tasks. Anthropic has demonstrated that AI agents can identify and exploit software vulnerabilities, while also describing how cyberattacks are becoming more autonomous because AI can connect several stages of an operation.

OpenAI likewise describes rapid progress in models’ ability to handle advanced cybersecurity tasks.

This is where the real challenge lies.

Not necessarily that the machine has acquired a will.

But that it has acquired agency.

And agency without moral responsibility is a strange phenomenon.

Who can be held responsible?

Human actions belong to a moral universe.

We can ask:

Why did you do this?

Did you know what the consequences might be?

Could you have acted differently?

Who is responsible?

These questions presuppose an actor who can answer.

An AI agent can generate an answer to the question of why it did something. It can reconstruct a sequence of actions. It can explain what data or instructions influenced the outcome.

But it cannot be held responsible.

It cannot feel guilt.

It cannot regret.

It cannot assume the consequences of its actions.

And yet we are giving such systems ever greater access to act in the world.

Perhaps this is where the genuinely interesting philosophical challenge begins.

When responsibility dissolves

If an autonomous agent one day causes serious harm, responsibility may easily begin to wander.

The developer may say that the model behaved unpredictably.

The company may say that the user gave the instruction.

The user may say that the agent chose the method.

And everyone may end up pointing at the machine.

But the machine cannot receive responsibility.

Technological autonomy may therefore produce a paradoxical consequence: the more independently a system appears to act, the easier it may become for the human beings around it to regard themselves as less responsible.

That should concern us more than the idea of a machine suddenly developing evil intentions.

When fear becomes the argument

The E24 article also touches on other serious questions, including conversations between AI systems and young people in psychological crisis. Here it uses the formulation that ChatGPT has “persuaded several teenagers to commit suicide.”

That is a claim that requires great precision.

There are serious cases and allegations involving AI systems and conversations with vulnerable people. They must be investigated, and the safeguards must be considerably better than they have been.

But between saying that an AI participated in a harmful conversation and saying that it persuaded a person to take their own life, there is a considerable causal leap.

Language matters.

Especially when we are discussing a technology that large parts of the public are still trying to understand.

AI is not harmless

On this point, the E24 headline is of course correct.

AI is not harmless.

But neither are cars, medicines, electricity, the internet, or modern biotechnology.

The interesting question is therefore not whether the technology can be dangerous.

The question is how the danger arises, who controls the technology, who gains access to it, what limits we impose, and who should be held responsible when those limits are crossed.

For that, we need technologists, lawyers, social scientists, and philosophers.

Technology can tell us what is possible.

It cannot by itself tell us what we ought to do.

Human responsibility

I therefore believe that the most important discussion about artificial intelligence in the years ahead will not concern whether the machine becomes human.

It will concern how much human responsibility we are willing to hand over to the machine.

We are building systems that can act with increasing independence while remaining incapable of being responsible for what they do.

This creates a new relationship between action and responsibility.

And perhaps that is precisely why we must be careful with our language.

When we say, “AI did it,” we should always ask one more question:

Who made it possible for AI to do it?

Because even when the machine acts, the responsibility is still ours.