How I Use AI

The honest question, for me, was never whether AI touches my work. It does. The question is whether my judgment stays in charge of it. That is the line I care about, and everything below is an attempt to say where it sits.

I have come to treat these tools as collaborators rather than machines I switch on to produce an output. AI isn’t a person, but it serves a role, and a collaborator shapes what you end up thinking about. What it raises, how it pushes back, even the things it fails to find - all of that leaves a mark on what I weigh and how I put the pieces together. I want to be candid that this happens, because the alternative is to pretend my ideas are untouched by such influence. This isn’t true, and I have a hard time trusting people who make claims about ideas being 100% their own.

The risk I already run

People sometimes set the risk of being influenced by AI against a picture of themselves thinking in perfect isolation. That picture has never described me. I am always being shaped by the people I argue an idea out with - a co-author who recombines my pieces in an order I would not have chosen, a colleague whose question exposes the soft spot in my reasoning, a student whose confusion tells me the explanation is not yet clean. I have gotten an enormous amount out of those exchanges. Hearing my own argument out loud, in someone else’s hands, is often the first time I see what is actually wrong with it.

Every one of those conversations can push my thinking somewhere I would not have reached alone. That is the standing risk of talking to anyone at all, and it is also where most of my good ideas come from. Working with AI sits inside that same category. The risk is real. It is close to the one I already take every time I think in the company of other people, and I do not see a principled reason to treat the machine as a special case.

What that requires is ongoing discipline. I stay open to being influenced, because the openness is what makes the exchange worth having. But I keep a second habit running alongside it: I watch how my thinking moves, ask why it is moving that way, and come back later to check whether the shift actually holds up. Openness without that second habit is just drift.

How I keep it honest

AI makes mistakes. I do not want to wave that away - it is a real concern, and a confident wrong answer is harder to catch than a hedged one. But people make mistakes, too. A collaborator can be plainly wrong. My own reading of their good feedback can be wrong. Error has always been in the room; it did not arrive with the algorithms, and I deal with it here much the way I always have. I want to see more than one line of reasoning and then choose among them on the merits. I look for an answer, but I also want the explanation that leads to it and the material underneath. An answer I cannot trace is one I cannot stand behind.

In data work this is not optional, and it is where the discipline gets concrete. I code in R, and I do not ask a model for the endpoint. I ask for the code that generates it. Then I ask for the intermediate products - a summary table, a sanity check on a few figures - and I step through the code a piece at a time to see what each step is doing and whether it adds up to the thing I actually meant. I plot the output and check that the picture agrees with the number, because the two disagreeing is usually how you find the bug. Building several cuts that have to cohere is where a lot of the quality control actually happens.

The mechanical errors are the easy ones. The harder mistakes survive a clean run of the code. They are mistakes of interpretation. Years ago I ran a DuPont decomposition across a set of firms, and one of them swung from outstanding to dismal from one year to the next. The arithmetic was right. What I was looking at was timing - heavy investment landing in the years before the returns it eventually produced, so the ratios lurched while the underlying business had not actually reversed. A mechanical pass reports that swing as an anomaly. You need the context to read it as an accounting artifact, not a performance indicator.

The same trap shows up in variance and regression work. A role, or a region, can look like it explains a great deal of the outcome when what it really reflects is that many different people sit inside it, and their individual differences are doing the explaining. The model cannot separate the two. Which one I am actually looking at is a question about the business, and the statistics alone will not answer it. That judgment is the first thing to disappear when analysis gets handed off without the context it needs.

None of this is special to AI. It is what I would do collaborating with any analyst, and breaking a workstream into pieces I can inspect is how errors have always been caught, the small ones and the large ones alike. There is a side benefit to all of this: getting that far into the work often turns up something I did not have going in - a question I would not have known to ask. Validation work serves as an audit. And it also refines my understanding of how the world works. That has been true whether the collaborator across the table is a person or a model.

The lines I do not cross

Some of the work I will not hand over, and I am clearest about this in teaching. The point of an assignment is the thinking the student does on the way to the answer, not the answer itself. A tool that shortcuts that thinking does not help the student; it defeats the exercise. I hold that line for my own work too. The part where a messy problem gets pulled apart into its real drivers - what is actually causing the margin, why the growth is or is not self-funding, which number is doing the real explaining - that decomposition is the work. It is the part I am trusted to do, and delegating it doesn’t make much sense to me.

So the intellectual foundation stays mine: the questions I choose to ask, and the judgment about what matters. Those come from my own knowledge and experience. I keep editorial control and I keep the accountability that goes with it. If something under my name is wrong, that is on me, and I would not have it any other way.

Where the tools actually sit

Within those limits, the tools do real work, and it helps to be specific about the roles they play. Once I know what I want to say, a model is a capable editor - it tightens a paragraph, catches a clumsy transition, offers a phrasing I can take or leave. It is a useful sounding board for an argument I am already holding, a place to press on a claim and find where it gives. And it is a genuine help in research, surfacing a source or a connection I would have missed - though I check what it surfaces, because owning a citation means confirming it is real. What these roles have in common is that the thinking is already mine before the tool touches it. When I reach for one too early, to fill a gap I have not worked through myself, it tends to go badly. The polish can be the tool’s. The argument has to be mine.

This is the practice I follow in my research and writing, in the advisory work I do for clients, and it is the standard I hold my students to as well.

A standard I expect to revise

I do not think my current practice is the last word. The tools are changing, my own sense of where they help and where they get in the way is changing with them, and I would rather update this statement as that happens than defend a position I happened to hold this year. Consider it a working commitment, open to the same scrutiny I try to apply to everything else.