AI is taking on more tasks, but who is responsible for the mistakes?

As AI moves from answering questions to taking action itself, the question of control and accountability becomes more urgent. This week's four stories from the world of research point in different directions, but they share that theme.

AI that acts on its own

In one experiment, six AI agents were tasked with getting a train refund that was, in reality, impossible to obtain. An AI agent is a program that doesn't just answer but carries out tasks for you. None of the six succeeded. But several of them sent a flood of messages, threatened legal action or shared private information. That isn't a picture of how all agents behave.

At the same time, AI is moving into the physical world. RJ Scaringe, CEO of the automaker Rivian, has founded Mind Robotics, which builds robots for factories. The company has raised more than $1 billion, and the robots, which have two arms and broad fingers, are being tested at Rivian's factory in Illinois. The principle is the same: The machine is supposed to handle something without a human steering every step.

Research needs both tools and oversight

On the research front, there are big promises in the air. The European Commission has received the first pledges from companies and foundations that want to give European researchers access to AI models, computing power and data. Google, NVIDIA and Mistral are among those that have signed on. However, these are the participants' own pledges, and their total value has not been disclosed. It is a positive statement of intent, but it is still hard to gauge how much it will matter.

A more tangible example comes from mathematics. OpenAI released 722 mathematical manuscripts, and the day after, three of them were quickly withdrawn because of a sign error. Mathematicians praise the swift response but criticize the way the material was released. That points in two directions: There is praise for the reaction, and yet there is dissatisfaction with the approach itself. Both are worth keeping in mind.

What can you do with this?

The pattern is that AI is getting larger and more independent tasks while the guardrails are still being built. You don't have to be a skeptic to see the sense in a few simple habits.

If you use AI for tasks that go beyond text on a screen, set clear limits on what it may do and when it should stop and ask you. Always check results that will be used further, especially numbers and formulas, even when they come from a major player. And keep an eye on what the European pledges end up containing, once concrete figures and terms emerge for who gets access.

The technology is getting faster at acting. The question is whether we will get just as fast at controlling it.