How I work

How I use AI in my research

A large part of my work is not about getting AI to write. It is about getting smarter, and about noticing when I or the models get something wrong. Here is how it works in practice, and why I have built the whole way of working around criticism rather than confirmation.

Why I use AI: to learn

I come originally from learning, cognition, behaviour, and communication. Those are the areas I have spent most of my working life on. But many of the questions I look into cross the boundaries between fields. So I use AI as a kind of private tutor and sparring partner.

When I run into areas I don't know much about, I ask the models to explain theories, research, mathematics, physics, biology, or other fields to me. And I don't just ask them to explain the material; I ask them to connect it to the problems and questions I am working on.

This is one of the things AI is unusually good at. People tend to become specialists. If you have spent a whole working life on biology, you probably haven't also spent one on physics, mathematics, or machine learning. AI, by contrast, can move across fields and help find analogies, structures, and patterns that would otherwise be hard to spot.

That does not mean the analogies are necessarily right. The opposite, in fact. A large part of the process is testing them, challenging them, and trying to refute them. Some of my ideas turn out to be bad. Others rest on misunderstandings, especially when I move into areas where I have limited background. Here AI is often just as valuable for telling me why an idea is wrong as for helping me develop it.

I try to stay very aware of my own limitations. I hold no formal academic position. I work independently, with no research group, no funding, and no institution behind me. That makes it even more important to be critical of my own results. So I have deliberately built workflows where the models are not there to confirm my ideas but to challenge them. I actively ask them to find errors, weaknesses, alternative explanations, and overlooked assumptions, both while a hypothesis is forming and in the hostile reviews that follow.

The goal is not to be right. The goal is to find out what is right. If an idea does not survive criticism, it should be discarded or revised. Science is not about being right from the start. It is about gradually being wrong in better and better ways.

How I work in practice

My work rarely happens by sitting down and writing a paper from beginning to end. It is more of an ongoing process where ideas arise, get examined, get challenged, get connected, and are gradually developed over time.

I have built a tool, Zeph.io, that lets me talk directly to the models. So a large part of the work happens as conversations rather than as traditional writing. I actually write relatively few prompts. Instead I talk. Often these are not finished thoughts or well-structured questions, but ideas, observations, intuitions, or connections that surface in the moment. Sometimes they are unclear. Sometimes they turn out to be wrong. Other times they become the start of something interesting.

So I try to capture thoughts as they arise. Many ideas don't come when I am sitting in front of the computer. They come when I am out walking, driving, swimming, or doing something else entirely. When that happens, I often dictate the thoughts straight to a model on my phone. Later the notes get moved into the relevant projects and papers. In that way AI works not only as a research assistant but as a kind of external memory.

My research is organised as a collection of standalone papers. Each paper has its own working session, its own notes, and its own discussions. At the same time they are tightly connected. An idea that arises in one paper can turn out to be relevant for several others. So I have built a system where papers can exchange notes, observations, hypotheses, and references with each other. I think of it as a dashboard, or an internal messaging system between the projects. It makes it possible to work on many parallel threads without losing the connections between them.

I also use AI tools with access to my local working environment. That means the models can help read, organise, analyse, and update material directly in the project structure. Every change is version-controlled as it happens, so the development can be followed and documented.

It is really a personal research environment built around AI, not just a chat window:

The result is a way of working that resembles traditional writing less and an ongoing dialogue between ideas, notes, experiments, and criticism more. The goal is not to produce many papers. The goal is to build a system where good ideas have a better chance of surviving and bad ones are caught sooner. In a sense the process is as much the innovation as the theories themselves, and that is probably also why I can work across so many fields without drowning in the complexity.

Am I not afraid of relying too much on AI?

Yes. It is probably one of the things I think about most.

What I fear most is not that my ideas turn out to be wrong. That is a natural part of research. What I fear are the banal mistakes, the kind AI has become known for: a paper that does not exist, a reference that is invented, a misread source, a simple error that should have been caught. Not because they necessarily change the conclusion, but because they can make people dismiss the rest of the work.

So I deliberately try to be more sceptical of the models' answers than I would be of many human explanations. I use several models to check each other. I actively ask them to find errors. I ask them to flag areas where they are uncertain. I run hostile reviews. And I use external tools to validate references and bibliographies.

My basic approach is that no single model gets to stand alone. When several independent models reach the same conclusion, my confidence rises. When they disagree, I look into why. This does not make errors disappear. It just makes them less likely.

The important shift is this: I do not blindly trust the models. I have built a process that tries to catch them when they get something wrong. Paradoxically, that is exactly why I dare to use AI as much as I do. Not because the models are perfect, but because they make it possible to run a kind of continuous quality control that would be very hard for one person to do alone.

At the same time, AI has become an important learning tool for me. I often work in areas where I have no formal background, and there I use the models to explain theories, research, and concepts at different levels of complexity. Sometimes that means translating technical literature; sometimes it means turning a complicated mathematical or physical explanation into something I can actually work with. I don't see it as a shortcut, but as access to teaching. Earlier it would have taken months or years to build up enough background knowledge to even ask the right questions. Today I can get an overview of a field faster, and then spend the time on checking whether the ideas hold.

It does not mean I become an expert in everything. The opposite. The more I learn, the more aware I become of how much I don't know. Maybe that is the most important role AI plays in my research. Not giving answers, but making it easier to spot my own misunderstandings.

The short note on tools and method is on the Disclosure page. To see what the work has turned into, the papers are on the home page.

← Back to plain-English home · Disclosure · Dansk version →