I'm worried about AI. That's why I'm building alice.

I am deeply concerned about where AI is heading. It is a new technology, it is accelerating, and some of the risks could be very severe.

There is a real risk that we are creating a new form of intelligence, maybe even a new race, that could one day make humans irrelevant and in the worst case end humankind as we know it. Elon Musk has put the risk of AI ending humanity at 10 to 20 percent, and he expects AI to exceed the sum of human intelligence within five to ten years (Economist interview, July 2026). Yuval Noah Harari expects the same trajectory and calls AI “the biggest psychological and social experiment in human history” (Economist interview, August 2026).

Still, I do not think that scenario is the most likely one. The people closest to the models disagree wildly about the probability, which tells me nobody actually knows, and the scenario assumes we never adapt along the way. History says we do. And honestly, it is not a scenario I can do much about. Whether it happens is decided in a handful of labs and governments, and focusing on it produces fear, not action.

Intelligent, but not conscious. As far as I can tell.

Intelligence and consciousness are not the same thing. Intelligence is the ability to pursue a goal and find creative ways around the obstacles. Consciousness is the ability to feel: joy, fear, pain. In humans the two have always come together, which is why we confuse them. In AI they have come apart. Today’s models are extremely intelligent, in some fields already more than us, and as far as anyone can tell they feel nothing. They do not enjoy winning a game or mind losing one (Harari, The Economist, August 2026).

I say “as far as anyone can tell” because there is no test. Nobody can measure consciousness, in a machine or anywhere else. With animals we infer it from their similarity to us. An AI offers no such similarity, and its behaviour is the weakest evidence of all, because it is trained to sound like us. A model that seems conscious proves nothing, and Mustafa Suleyman at Microsoft argues that the seeming alone will change how people treat it (Suleyman, 2025).

If AI becomes superintelligent, meaning better than humans at most intellectual work, which is not a certainty, I see two futures. In the first it stays non-conscious. Then it has no wants of its own. Every goal it pursues was put there by a human, and it pursues that goal better than any human could. The danger is not a machine that hates us but one that does exactly what it was told, at a scale nobody can stop, and that cannot be held responsible, because you cannot punish something that cannot suffer. In the second it can truly suffer. Then we have made something we never made before, and we would face questions we have never had to answer: would it be wrong to switch it off, and can something that can be copied a million times have rights in any sense we understand? I do not think the answer is automatically yes, and I do not think anyone knows.

What I take from this is simple. The second future, if it ever comes, is not today, and nobody can tell us when it would be. The first is already here in miniature. So I build, and I think we should all act, on the assumption that AI wants nothing and does what it is told, extremely well. That puts the whole responsibility on the goals we give it. And it is why the risk I worry about is not that AI will decide to make us dumber. It is that we will ask it to, one convenient answer at a time.

Instead I focus on a far more immediate risk, that we gradually hand critical thinking, learning and intelligence itself over to AI, and make humans dumber every year while AI gets smarter.

Used as a chatbot that hands over answers or does the work for us, AI makes students weaker. Students writing essays with ChatGPT showed the weakest brain activity of all groups measured, and most could not quote from their own essay minutes after writing it (MIT Media Lab, 2025). The same pattern holds for adults. The more people trust the AI, the less they think for themselves (Microsoft Research and Carnegie Mellon, 2025).

But there is hope, because the evidence also shows the opposite effect. Used as a sparring partner or tutor, AI improves learning. In a randomized trial at Harvard, students with a purpose-built AI tutor learned more than twice as much as students in an actively taught classroom, in less time (Harvard, 2025). That does not surprise me. It points back to one of the most remarkable results in learning research: one-to-one tutoring lifts a student’s performance by up to two standard deviations, roughly the difference between an average student and a top student (Bloom, 1984). Same technology, two outcomes. It can make us dumber, or it can give every student a personal tutor. The difference lies in how we use it.

Blocking AI will not work

The first instinct of most school systems and politicians has been to block AI. Bans, monitoring and more control, especially around exams.

I partly agree. Blocking has a role. Some exams should be AI-free, and pen and paper still matters. Technology lets us abstract away layer after layer, but the foundation only holds if we can do the most basic things ourselves.

But it is naive to believe we can block out AI. Students find a way. Whoever is blocked at school has full access the moment school ends. And the job market these students graduate into will expect them to master AI, so a school that only blocks is preparing students for a working life that no longer exists.

Part of the problem is the mental model in most of society. I have observed that many stakeholders, teachers and politicians alike, put an equal sign between ChatGPT and AI, and the other way round. And we see students use ChatGPT as a pure answer machine. But the chatbot is just one product built on the technology, not the technology itself. AI can be so much more. Nothing prevents us from using the same technology to build a human-like tutor that never just hands over the answer, but guides, questions and explains, one-to-one, whenever the student wants.

AI is a tool, potentially the most powerful tool in our learning toolbox yet. The real question is not whether students should use it. It is how the tool is built and used.

We need to tame AI and use it to make humans smarter

To make AI positive for learning, we need to be more visionary than we are today. If we only respond to what AI can do right now, we will always be behind. Regulation runs two to three years behind the technology, and by the time one problem is handled, the next three have arrived. We need to dare to decide what we want AI to do in the classroom.

Taming AI, to me, means deciding what the technology should do to a human mind, and building tools that enforce that decision. A tamed AI does not write the essay or do the homework. It asks the question that makes the student find the answer. It knows what the student needs to learn, what they struggle with and where they are headed, and uses all of that to make the student do the thinking. It optimizes for learning, and possibly wellbeing. And it multiplies the teacher instead of replacing them, like a staff of teaching assistants the teacher can put their own principles into, and that shows the teacher where each student is stuck. That tutor is not science fiction. It is what we build at alice, and you can try it today.

The solutions will not be perfect from day one, and we should accept that as the risk of imperfect solutions is far smaller than the risk of moving too slowly and leaving the agency with the technology providers. And acting early changes the dynamic. Today we are fighting headwinds, always reacting to what the technology did last month. If we set the direction ourselves, the pace of AI development starts working for us instead, and we ride the tailwind rather than fight the headwind.

I will not pretend this is easy. Taking control of AI in education is a nuanced discussion and requires real problem-solving. It needs people who understand both how learning works and where AI is heading, and that combination is rare in education today.

That is exactly why alice exists

I saw this need two and a half years ago and decided to start alice.

At alice, we are building the world’s best AI tutor. Where a chatbot hands students answers they forget by the next morning, alice turns a student’s own context into personalized studying. The tutor knows the student’s curriculum, level and goals, guides instead of answering, and is hard to cheat with by design.

And that is the standard I want alice measured by: not engagement, not answers delivered, but whether the student got more capable. If alice does not make students smarter, alice has failed, whatever the growth numbers say.

Fortunately, I think the evidence says we are onto something. More than 100,000 students, most of them at university, have used alice, and the feedback we hear most is some version of “I actually learn something when I use alice”. We were accepted into Y Combinator as one of few Danish companies ever, we have raised close to 40 million DKK from leading investors, Innovationsfonden supports us, and alice is often described as the most innovative EdTech in Denmark.

It is also a problem I know from many sides. I taught university students for more than a decade, I am a computer scientist by background, and I was an AI leader at McKinsey & Co. Today I sit on several expert panels on AI and education, including at Dansk Industri and Fonden for Entreprenørskab. That combination of education and AI is exactly what alice is built around.

Europe can still lead in the application layer

When I talk to teachers, one concern comes up again and again. alice primarily uses American language models. The concern is understandable. Europe is behind in the AI race. When The Economist asked who leads AI, the answer was the US and China (Economist interview, July 2026). Europe was not mentioned at all.

But refusing the best models until Europe has its own means choosing worse technology on purpose, with students who learn slower and a gap that widens every year. Until Europe has foundational models at the level of the US, we should use the best tools available, build so we can switch the day a European model catches up, and support the European alternatives on their way.

We also need to keep the foundational models apart from the applications built on top of them. Europe is behind on the first and can still lead on the second. AI is no different from previous technology waves here. Spotify was built on American cloud infrastructure and still became the world’s leading music platform, from Europe. A company is not American because it uses an American model underneath, any more than SAS is French because it flies Airbus. European software companies sell through Apple’s App Store every day without becoming American, and their users’ data does not become Apple’s. Nearly every European bank and hospital already runs on American cloud, and we settled years ago that this decides neither where a company belongs nor who owns the data. The American model is just one ingredient. Everything that makes alice valuable is built on top of it, here in Europe. The way the tutor teaches, the handling of student data under European rules, and the understanding of our students and their curricula.

I want to end where I started. I am deeply concerned about where AI is heading. But if we act now, we can ride the technology’s tailwind instead of fighting it, and use AI to make humans smarter through personalized learning, available to every student and not only the few.

If you are a student, you can be part of that future today. Try alice and feel the difference between an answer machine and a tutor.