How I use AI

I work on my own. AI tools let me read more, faster, and turn a week of reading into something I can actually use. What they do not do is think for me. The theses on this site, the funded-versus-financed distinction, the argument that democracy needs its own economy, the read on what a given moment in Brussels actually means, are mine. I arrived at them by thinking, usually over months, often in conversation with people who disagree with me. A model is not good at that work, and I am not asking it to do it.

So rather than let you guess, here is what is actually happening.

Where AI helps

• Reading at volume: searching long documents, comparing versions of a text, pulling the through-line out of a stack of reports
• Transcribing recordings and tidying my own notes
• Checking a figure against its source before I use it
• Editing passes on drafts I have already written, mostly cutting
• Code and layout for this website

Where it stays out

• The argument. Every thesis here is one I reached myself.
• Client advice, in any form
• Sources. Every citation on this site points to something I opened and read. If a model surfaces a number, I check it against the original before it goes anywhere.

What I never put in

• Client, partner and grantee material that is confidential. That does not go near a general-purpose model, whatever the convenience.
• Named individuals and organisations from my research, where I can avoid it. Before I ask a model to do an assistant task, I strip out the names and identifying details and work with role labels instead. The reasoning does not need to know who; it needs to know the shape of the thing. I also keep training turned off on the tools I use, so my inputs are not fed back into a model.

The environmental cost

These tools run in data centres that draw real electricity and water. Most of that cost sits with the companies training the models, not with me at my desk, and I am wary of the theatre where a solo advisor claims to be saving the planet by writing shorter prompts. What I can do is use the tools with some discipline. I reach for a heavier reasoning model only when the work genuinely needs it, not to summarise a PDF. I keep prompts specific, which produces shorter and more useful answers. For a plain fact I use ordinary search, which costs a fraction of a generated one. None of this is a large saving on its own. It is the same discernment I would want applied to any tool with a cost attached, which is to use it where it earns its place and leave it alone where it does not.

Why this matters beyond my own desk

The tools most people in this city reach for are trained overwhelmingly on English-language text and on American legal and political assumptions. European policy runs on twenty-four languages, two legal traditions and a habit of negotiated compromise with no obvious equivalent in the training data. A model can summarise a trilogue outcome. It cannot weigh what a member state will accept in November against what it said in June. That judgement is the job, and it stays with me.

There is a second reason to be clear about this. The first reader of anything I publish is increasingly not a person. It is a system summarising me for someone with thirty seconds to decide whether I am worth their attention. I would rather that summary be built on a straight account of how I work than on a guess.

If something here is wrong, it is my mistake. Tell me and I will correct it.