Why AI Tool Defaults Trade Your Depth for Ease

A small moment from yesterday

A friend who builds product told me about a morning this week. He opened his AI assistant, typed three words, accepted the suggested prompt, took the answer, and shipped a draft. He said the strange part was not that it was fast. It was that, two hours later, he could not remember what the assistant had said, what he had said, or how the draft had ended up the shape it was. The session had been smooth. He had no fingerprints on it. That smoothness is now a design property, not an accident, and the research is starting to name what it costs.

What the research shows

A peer-reviewed experiment from Yonsei University, published in March 2026 in the International Journal of Human-Computer Studies, asked thirty-six people to do exploratory searches with a generative AI tool. The researchers, Nakyung Kim and Yong Gu Ji, varied two design choices. First, how the AI revealed its answer: all at once, or progressively, in stages the user followed along with. Second, how the prompt was formed: by accepting the system’s suggestion, or by writing one’s own with guidance. The combination of progressive answers and self-formed prompts produced more exploration, more engagement, and more of the kind of cognitive effort that researchers call germane, the productive sort that builds new understanding rather than wasting attention.

A second small study reinforces the pattern from a different angle. A team at the University of Michigan, publishing in late 2025 in the Proceedings of the ACM on Human-Computer Interaction, ran a within-subject experiment on AI-supported note-taking with thirty people. Three conditions: high automation with structured notes generated for the user, moderate assistance with a real-time summary alongside the user’s own notes, and transcript only. Moderate assistance produced the highest post-test learning scores. Full automation produced the lowest. Participants preferred the fully automated condition anyway, because it felt easier. The preference-versus-outcome gap is the punchline, and it suggests that the path of least resistance in AI design quietly bends people away from their own understanding.

What it doesn’t tell us yet

Both studies are small lab experiments, thirty-six and thirty people, single-session, unfamiliar topics. They do not show what happens over weeks of real knowledge work on real projects. They do not tell us whether the depth-for-ease tradeoff replicates at scale, or which populations feel it most, or how it interacts with skill and experience. What they do is point a finger at a design lever that field telemetry alone would not surface. This is a directional finding, not a settled rule.

One thing to notice in your day

Sometime today, watch what happens at the exact moment an AI assistant offers to do the next step. Notice whether the suggested prompt gets accepted, or whether a fresh one gets written. Notice whether the answer arrives all at once and gets scanned, or whether it forms in pieces and gets read along the way. Both of those small choices are the same design lever the Yonsei researchers manipulated. The interfaces almost always nudge toward one-click and full answers, because that is what feels good to use. The research suggests the other path is where the building actually happens. This is the same texture as the question of where coding-agent ROI shows up first, and the same gap that shows up in research on how AI reliance erodes self-belief over time. None of this is a verdict on the tools. It is a noticing prompt about which side of the default you are standing on. The honest practice today is to catch one moment, just one, and see which path the hand took.