Most of the debate about AI in critical infrastructure is about whether the algorithm is good enough. In our experience that is the wrong first question. The systems that fail in control rooms are usually not the ones that were technically wrong — they are the ones nobody trusted, or trusted too much.
Our paper Human-AI interaction in safety-critical network infrastructures, published in iScience as part of the AI4REALNET consortium, is about that second problem: what a well-defined working relationship between a human operator and an AI decision system needs to contain.
Why infrastructure is a special case
AI is spreading through industries where a wrong answer is expensive but recoverable. Power grids, railways and air traffic are not like that. They are operated continuously by people who are accountable for what happens, under regulation, with consequences measured in outages rather than in lost revenue.
At the same time these are exactly the systems where the operator’s job is getting harder: more renewable generation, more volatility, more decisions per hour than a human can reasonably evaluate. So the pressure to automate is real, and so is the reason to be careful about it.
What the paper argues
The paper sets out where human-AI collaboration in network operations currently stands and where it needs to go. It covers the forms this interaction can take — from an assistant that ranks options to progressively higher levels of automation — and the conditions each form requires: explainability that an operator can act on, a clear division of accountability, and the ability to learn from how the human and the system correct each other over time.
It is a research paper, not a product claim. What it gives us is a shared vocabulary with the operators we work with for a conversation that otherwise tends to happen in generalities.
Why we build this way
The design of FLEX follows directly from this line of work. The system proposes; the operator decides. Every recommendation carries the power flow that justifies it and the constraint it protects. Regulatory and physical limits are enforced inside the optimisation rather than checked afterwards, so an action that breaks a rule is never put in front of a human in the first place.
That was a research position before it was a compliance requirement. Since the EU AI Act’s high-risk obligations became enforceable in August 2026, it is also the thing grid operators ask us about first - which is covered in more detail on our trust and compliance page.
Read it
The paper is open access, use this link.
