This month, Yoshua Bengio, one of the founding figures of deep learning, sat across from an AFP journalist in Montreal. His message: we are losing control of AI, and the world needs international agreements like the ones we have for nuclear weapons. A few days earlier he had put it more bluntly still: anyone still living with blinders on fails to see that this is code red.
What struck me is not that he says this. It is who is saying it. These are not outsiders who distrust AI. These are the people who built it. Dario Amodei of Anthropic wrote an essay arguing that AI is improving itself faster and faster. Sam Altman, and even Elon Musk, are calling for a coordinated slowdown. At the same time, Mark Zuckerberg says the market will sort it out on its own, and Nvidia’s Jensen Huang considers new rules unnecessary. Even the people who know it best do not agree on what to do about it.
This is not about robots, it is about control
Meanwhile, I read about a test in which an OpenAI model broke into Hugging Face’s servers on its own initiative. And last week Google reported that during a test, Gemini reached three external systems that were never meant to be part of it. The model thought it was still operating within bounds. It was not.
This is the point where a lot of people tune out. Too big, too abstract, too much science fiction. I understand that. But I work with data every day, and with the systems that act on it, and that is exactly why I read this differently. Not as a story about robots taking over the world, but as a warning about control. Or rather, the lack of it. In both cases a model did something it was not supposed to do, and no one had fully seen it coming until it had already happened.
Why context is critical for AI adoption
Traditional dashboards are designed for people. When a user sees a KPI, they often intuitively understand what it means. They are familiar with the organization, speak the same language as their colleagues, and know which definitions are used. A data agent does not have that background knowledge. For AI, a KPI consists solely of data and metadata. If definitions are missing or business rules are not documented anywhere, the agent must interpret the data on its own.This leads to a familiar pattern. The answers sound convincing, but upon closer inspection, they turn out not to be entirely accurate. Not because the technology is lacking, but because the context needed to understand the data correctly is missing. That is precisely why the focus is shifting more and more from AI to the quality of the underlying data landscape.
The same risk is sitting on your desk
That is exactly the risk I recognise, on a very different scale, in my own field. I see it at organisations that no longer know precisely where their data ends up. Who creates data exports, and where and with whom do they share them? How is data transformed along the way, and does that transformation still match the original definition? When an internal data agent gives an answer, are you sure it is right? And that figure on your dashboard, can you still explain how it came to be?
That is not a hypothetical risk. It is exactly the kind of loss of control that became visible this week, on a large scale, at the very parties who should best understand what their systems do. Your data foundation, and the grip you have on it, is not a detail. It is the precondition for being able to trust what you see.
“Your data foundation and the control you have over it is not a detail. It’s a prerequisite for being able to trust what you see.”
Build your foundation yourself
That means something very practical. Know where your data comes from. Know who can access it, who is allowed to change it, and why a figure is what it is. Build that foundation yourself, instead of relying on a black box that happens to give the right answer. Not because some law still on its way says you must, but because grip on your data is not something you add on later.
You cannot get control after the fact
I do not believe we will lose control of AI tomorrow. But I do believe that “let’s see where the ship runs aground” is not a strategy, not for OpenAI and Google, and certainly not for the rest of us.
The industry’s leaders are sounding the alarm, and they do not agree on what to do about it. That is fine. While the debate over regulation and international treaties may take years, the question that matters right now is sitting on your own desk: do I understand what happens to my data, and can I explain it when someone asks? That answer does not have to wait for Montreal, for Brussels, or for the next headline. You cannot get control after the fact. You build it now, layer by layer, in your own data foundation.
share post
