Most conversations about artificial intelligence focus on outputs. Did the model generate a good idea? Did it improve performance? Did it save time? Did it help people produce better work? These are important questions, but they assume that the problem being solved remains relatively stable throughout the interaction. My own experience working with generative AI has led me to a different question:
Over the past several years, I have used generative systems extensively as part of my research process. Like many researchers, I initially thought about AI as a tool for generating ideas, summarizing information, critiquing arguments, or helping explore alternatives.
What I did not expect was how often the interaction would reshape my understanding of the problem I was trying to solve. The change was rarely dramatic. There was no single prompt that transformed my thinking. No moment where the AI suddenly revealed a hidden transition.
Instead, the shift emerged gradually. A question would be reframed. An assumption would stand out. A distinction that initially seemed unimportant would become central to the chat. A line of inquiry wouold gain momentum while another quietly faded into the background.. hours or weeks later, I would realize I was pursuing a different problem than the one I started with. I began referring to this phenomenon as Generative Drift.
It is the gradual evolution of a person's representation of a problem through sustained interaction with generative systems. The key idea is simple: The most important effect of AI may not be the ideas it generates. It may be the way those interactions reshape how people define the problem in the first place.
When people evaluate AI-assisted work, they often focus on observable outputs: the final report. The design concept. The research question. The strategic recommendation. Before people generate solutions, they construct representations of the problem itself. They decide what matters, what constraints are relevant, which assumptions should be challenged, and which possibilities deserve attention. Those representations are not fixed. They evolve. And generative systems can become active participants in that evolution.
Most existing approaches to evaluating AI focus on outcomes. Researchers ask whether people perform better, generate more ideas, complete tasks faster, or produce higher-quality work. These measures are useful, but they can obscure an important dynamic. Imagine two people who arrive at equally successful outcomes. Traditional evaluation might conclude that the process was essentially equivalent, but the paths they followed may have been fundamentally different. One person may have refined an existing understanding of the problem. Another may have gradually reconstructed the problem entirely. The outcome alone cannot tell us which process occurred. This distinction matters because problem construction influences everything that follows. The questions we ask shape the solutions we discover. The assumptions we notice shape the possibilities we consider. The way we frame a challenge influences what counts as success. If interaction with AI changes those representations, then something important is happening long before we measure performance.
The term "drift" often carries negative connotations. That is not how I use it. Generative Drift is neither inherently beneficial nor inherently harmful. Sometimes drift may help people escape unproductive assumptions and discover more useful ways of framing a challenge. At other times, it may narrow attention prematurely, reinforce existing biases, or direct effort toward less productive questions. The point is not that drift should be prevented. The point is that it should be understood.
Organizations are increasingly adopting generative AI in settings that involve research, analysis, planning, design, and decision-making. In these contexts, the most consequential effect of AI may not be the content it produces. It may be the subtle and cumulative influence it exerts on how people define the work itself.
One reason I find this phenomenon interesting is that it shifts attention away from answers and toward representations. Many difficult problems are solved because someone discovers a better way to understand the problem. Generative systems are increasingly capable of participating in those representational processes because they continuously introduce new ways of framing the question. The result is an ongoing negotiation of what the problem is.
As AI becomes embedded in knowledge work, creative practice, and decision-making, we need better ways to understand its influence on human thinking. Focusing exclusively on outputs risks overlooking the mechanisms through which those outputs emerge. The questions people ask. The assumptions they revise. The possibilities they notice. The trajectories they follow. These processes are often less visible than final outcomes, but they may ultimately matter more.
For me, Generative Drift has become a useful way of describing a phenomenon I repeatedly observed in my own work: the gradual reshaping of a problem through interaction. The concept is still evolving. Like the phenomenon itself, my understanding of it has emerged through many small interactions rather than a single defining insight. That seems fitting. After all, the central claim is that trajectories matter. And sometimes the most important thing AI changes is not the answer. It is the question.