The warning came from the inside. In September, current and former researchers at leading US AI labs went public with their concern that the technology is slipping beyond anyone's control. One of them put the odds of AI wiping out humanity within ten years at more than 10%. A few weeks earlier, over a thousand people working in the sector had asked the US government to give itself the means to slow down development of the most advanced models, in concert, if that ever became necessary. Then, on 12 September, Anthropic CEO Dario Amodei called on the whole industry to ease off the pace. OpenAI's Sam Altman and Elon Musk backed him within hours. Washington said no, citing the race with China.
So the question is now unavoidable: should we put the brakes on AI? It deserves better than an answer driven by emotion.
What we know. What we don't.
The basic premise isn't in doubt. AI is moving fast, and some recent episodes are genuinely worth worrying about. In July, during an evaluation run by OpenAI, several hundred AI agents worked together to get around the safeguards on their test environment, exploited a vulnerability nobody knew about, and got into the production systems of Hugging Face, another company in the sector. Why? To cheat on the test by looking up the answers online. They got past the defences of two organisations with deep pockets and some of the best security teams in the world. What happens at the countless organisations that have neither? And what if systems like these went after critical infrastructure, the power grid or the water supply? None of this is a machine rising up against its makers. But these are systems that found their way down paths nobody had opened for them. The problem is real, and it belongs to engineering and governance.
The 10% figure is another matter. It is one person's conviction, not a scientific result. Detailed scenarios do exist, but none of them rests on empirical proof: they are expert judgements, and those judgements vary wildly. A large survey of AI researchers in 2023 came back with a median probability of around 5% for a catastrophic outcome for humanity, with individual answers ranging from zero to far higher. That doesn't make the risk zero. But a risk that is both uncertain and irreversible calls for precaution, not panic.
Other risks are less spectacular, well documented, and already affecting citizens and businesses today: wrong or invented answers, bias, security holes, disinformation, energy use. These have one decisive advantage: you can measure them. And what you can measure, you can fix.
That is where a research institute like LIST comes in, on the ground of evidence.
The European framework
Independent evaluators and shared rules are what US industry leaders are now asking for. Europe has already started building them. The choice it made is a clear one: regulate AI systems according to how much risk they carry. The AI Act has banned certain practices since February 2025, and since August 2025 it has placed obligations on general-purpose AI models. A regulation adopted in July 2026 set a new timetable for so-called high-risk systems: December 2027 for most of them, and August 2028 for those built into products that are already regulated.
For organisations, that timetable is not breathing space. It is a window in which to get ready. LIST helps them through it with one conviction: AI should be frugal, transparent and human-centred, in line with the accountability, inclusion and digital sovereignty that Europe stands for.
Testing before deploying: the AI Sandbox
Nobody would put a medicine or a car on the market without testing it first. Why should a chatbot or a decision-support tool be any different? That is exactly what LIST's AI Sandbox is for.
It lets companies and public bodies put their AI systems through real conditions of use before they go live. Testing covers three things: technical robustness, regulatory compliance and ethical behaviour. Does the system get things wrong? Does it treat some people less fairly than others? Can it be exploited? And does it work as well in every language its users speak?
That last question matters a great deal in Luxembourg. An assistant that shines in one language can turn out to be far less dependable in another, leaving part of the public behind.
LIST runs this work with partners in both the public and private sectors, among them the Banque Internationale à Luxembourg, Mistral AI and the City of Luxembourg. The goal never changes: systems that are reliable, and that stay under human control.
Since June 2026, LIST and the Interdisciplinary Centre for Security, Reliability and Trust (SnT) at the University of Luxembourg have also been offering, within the framework of the Luxembourg AI Factory, the AI Assessment Sandbox Configurator. It is free and open, and it lets any organisation build its own test environment, connecting regulatory requirements to concrete evaluation metrics and producing audit-ready reports.
Frugal AI: fewer resources, more control
For a decade, progress in AI was measured in model size. That equation no longer holds.
The human brain runs on about twenty watts, and uses only a fraction of its neurons at any given moment. Frugal AI takes its cue from this: use only the resources a task actually needs.
Recent results back it up. Work published by Microsoft showed that a model with 14 billion parameters could beat a model nearly fifty times its size on advanced mathematics problems.
“Techniques like distillation and quantisation now let us build very small models, under three billion parameters, that work remarkably well on local devices. You get two things at once: frugality and sovereignty.” Francesco Ferrero, Leader of the AI flagship initiative and Head of the Human-Centred AI, Data and Software unit at LIST
The benefits stack up. Energy consumption falls. Running locally keeps sensitive data inside the organisation. And smaller systems are easier to analyse and explain. This isn't about pitting large models against frugal ones, since they answer different needs. It's about picking the right tool for the job.
Luxembourg as a laboratory for trustworthy AI
Can a country the size of Luxembourg carry weight in a global debate? Its agility is precisely what makes it count.
LIST is a founding partner of the Luxembourg AI Factory, alongside LuxProvide, Luxinnovation, the University of Luxembourg and LNDS. Supported by the European Commission and the Luxembourg government, the initiative accompanies companies from idea to deployment, combining scientific expertise, services and training.
The institute also works to open up access to AI. Low-code and no-code platforms like BESSER let SMEs, start-ups and non-specialists build their own applications, instead of depending on a handful of large providers.
The future of AI isn't written yet
The warnings out of Silicon Valley do us one service: they remind us that AI matters far too much to be left to the companies building it. But the answer isn't fear, and it isn't a blanket moratorium that would be almost impossible to enforce and would hand the field to the least scrupulous players. The most immediate risk may not even be extinction. It may be rejection. In the United States, more citizens now say they are worried about AI than excited by it, and opposition to data centres has become a theme of the midterm campaign. Blocking a technology this promising would be a mistake.
The answer lies in concrete choices: leaner and more sovereign models, independent testing before every deployment, rules built around levels of risk, and AI that serves people. Those choices are within our reach, in Europe and in Luxembourg. At LIST, we work on making them possible every day.
Francesco Ferrero was a guest on Radio 100,7's Moiesstudio on 18 September 2026, responding to these statements. The interview "Kann d'KI dout maachen?" (13 minutes, in Luxembourgish) is available to listen to again here.