How I Use AI
There’s a lot of talk about whether AI touches output or not. Some contend that any use of AI makes things worse, and there is a tendency to think that human touch means the quality must be higher. I think it’s more complicated than that.
AI performs well for some things, and much depends on the project and how it’s applied. It can be useful for pressure testing a first draft. It can be good at turning more disorganized thoughts into structured options to build on. It’s a constant back-and-forth conversation between you and what feels like a collaborator. It’s not a person, so it’s not exactly collaboration. It’s a different form of creative work.
Why it still has to come from us
With all that said, the heart of the work still needs to be from us. I think there are two reasons.
First, authenticity matters. People want to know what we think and how we’ve made choices to present ideas. When all the thinking - the content, the judgments, and the curation - gets outsourced to something that’s not us, the work stops being ours. It also tends to be worse. The question isn’t whether we’re using the right tools, but how we use them to produce our best work. Our best work is an expression of ourselves that makes people feel something or think about something differently.
Second, communication requires trust. Today, it’s more important to audiences to have a sense of the journey taken to arrive at a point. Our thinking has always been a hidden result of the people we’ve spoken with, of the things that we’ve seen, and of experiences that shape the salience of our different thoughts. So much of our thought process is opaque, even to ourselves. For some reason, it seemed easier to trust when we were talking with humans. Now that AI is involved, we want to know how it was involved in the journey.
People are right to want to see what that process is. It’s always been important, and AI has sharpened the value of that context. Even when the entire journey is complex and difficult to present in full, I think it’s a good thing when we can share what we can with those who are curious.
The risk I already run
For a lot of us, creating work doesn’t come from sitting alone in a room, spewing thoughts. I have long worked with other people, including co-authors, editors, and thought partners. The risk of being influenced by them, along with the benefit of seeing another perspective, has always been a tension. I think that comparison is apt when we consider the choices writers face with AI: who they reach out to, when they reach out, and what they want to preserve as entirely their own versus what they want to open up to new enrichment.
I think that in some areas in my writing, opening up to other people or opening the work up to AI changes my thinking, and that is what makes the ideas richer and better, and sometimes a better expression of the point I’m trying to make. In those cases I try to keep it. There’s always the risk that my thinking has shifted more than I want, or perhaps in a direction I don’t like; in those cases, it’s important to be mindful about preserving the important elements. These types of decisions happen a lot with writers who work with others, who talk to others, who work with editors. In the end, what we’re all after is to make something better and to refine our own thinking, too.
Creating better work requires being open to some influence. It also requires preserving a voice and a coherent vision. And of course, it requires taking responsibility for the words expressed on the page.
How I keep it honest
AI is something like a diligent worker and partner, but, like other people I work with, it sometimes gets things wrong. I deal with it much the way I deal with double-checking myself and double-checking others when we talk about things. You want to see reasons; you want to see sources. When we talk about logic, especially in calculations, we want to trace it, not just peer into a black hole with no audit trail. The pathway to conclusions is just as important as the conclusions themselves.
I code in R, and when I co-create code with either people or a machine, we explicitly put down where the calculations are and how we are thinking about them:
- In a plan
- In code, which brings the calculation to life
- In quality check plans that make sure we execute on our intent
- In the final appraisal of results
This feels like a lot of work, and it’s very much like the academic process. It is not because we don’t trust ourselves but because we really want to make sure that we’ve put quality into the answers. There’s such a need for careful coherence linking an idea, an approach to investigating it, and its implementation. In each one of those stages, there can be mistakes or slight misinterpretations. Often, it doesn’t quite capture the full, rich, messy context of the world. For that reason, there’s always a need to proceed cautiously, whether you’re dealing with people or with AI - to have a very clear idea of what you’re after, how you’ve translated it, and a path back to check what we’re doing. That’s what I seek to do with both people and AI.
Not too long ago, I ran a DuPont decomposition across a set of firms - taking a lot of data and re-computing it into a year-by-year sequence to decompose where performance comes from for a company. One of the companies swung from outstanding to dismal from one year to the next, and all the math was right. What we realized was that for some companies, where there’s a lot of upfront investment and returns come later, a big change in measured performance might reflect a mismatch between the timing of investments and when returns materialize. This is a classic contextual problem with all kinds of accounting data. You really need to understand the context of the industry, of a particular business, even the particular firm’s decisions, to grasp the right insight. That’s something that even great analysts struggle with.
The same kind of problem arises when using variance decomposition to understand performance drivers. We found that differences across regions and among specific managers were strongly associated with performance levels. In this case, managers were very often tied to region (but not always). When we tried to isolate these two different variables, we had to reconcile this overlap. Understanding how much overlap there is and its context helped us interpret the results with care.
When we see the numbers, that additional layer of insight and understanding is critical. Rather than making quick conclusions (even when numbers are calculated correctly), good investigators need to think through the insights carefully. It involves thinking through the evidence, how it arrives, and what we might be missing.
The care required to produce quality work has always been important, and it remains so, even in the presence of AI augmentation. It’s a cautionary tale that you can’t just put things in and expect truth to come out. We’re always trying to understand the full implications of contextualizing calculations in a messy, complex reality. Not all things can go into an equation, so careful interpretation is required to know what we can learn from our investigations.
The lines I do not cross
There are some lines not to cross when we’re teaching or learning something. We’re trying to test students’ ability to do something. It might be something that an AI can already do, but the point of it is that we believe it’s valuable for someone to know it - it’s not just a task, but a capability that contributes to other skills, too. When it comes to knowing something cold and testing ourselves, there’s value in demonstrating it unaided. Outsourcing that obviously skirts the whole idea of what we’re trying to do.
When you’re developing judgment, outsourcing it misses the whole point. Whether or not we’re able to judge the calculations in context correctly (e.g., what’s really driving the margin of a company, whether this growth is sustainable, and what makes us think that) is something that should come from within. That’s very different from relying on AI to help us organize our files, compile summaries, and reformat the data in ways that let us see things differently. Complex work often requires doing both kinds of work, and we should be intentional about where we bring AI in.
How I actually use AI tools
It depends on the projects, the richness of the context, and what I’m trying to deliver. It can be used in a multitude of different roles depending on what’s appropriate. As an editor, AI can help rephrase passages that seem clumsy and sweep for grammar and spelling errors once an idea is down. Also, it’s a great tool for organizing source materials. I’ve also found it very useful for more mundane tasks, like refactoring code and making global formatting changes.
It can also be a sounding board for taking the opposite side of an argument. It is not to say that we adopt those arguments wholesale or give up our positions, but understanding other plausible possibilities sometimes surfaces things we missed. That improves the work and often our own thinking as well.
Researching and uncovering different sources can also be a very valuable contribution. We must trace through all the provenance and understand where such sources came from. That’s true whether it comes from an AI or another person. Making sure we understand how this connects to our material is important. Researchers can find sources, but determining the implications and how they connect to the work is something that authors do.
A standard I expect to revise
As the technology evolves and as we experiment with different creative workflows and learn new things, things will change. All of us have to learn how to use this technology effectively, and we have to experiment and make our own decisions about where it fits best in our creative process.
I keep thinking, experimenting, changing, and openly discussing, rather than committing to just one way of doing things. Overall, I think it’s best to be both careful and courageous as we move forward as investigators, creators, researchers, and thinkers.
Updated August 17, 2026