AI tools barely move the productivity needle so far — is that because they're built wrong, or because the models underneath just aren't good enough yet?

This explores whether GenAI's uneven productivity comes from how we build and use the tools, or from limits in the models themselves. The corpus leans toward design as the bigger lever right now, with one important exception. The gains are real but small: survey data puts today's aggregate productivity boost at about 1.1%, close to expert estimates What is generative AI's actual aggregate productivity gain today?. That gap between what demos promise and what workplaces see is what both camps are trying to explain.

The design camp has the strongest argument, and part of it is decades old. Simkute and colleagues show that GenAI is repeating failures that aviation automation research documented long ago. Workers stop making things and start checking them. Workflows break apart. Interruptions disrupt focus. Easy tasks get easier while hard tasks stay hard Does generative AI productivity actually improve or decline in practice?. Mollick makes a sharper claim: models are already more capable than people can use, and the chatbot interface wastes that capability. In one study, financial professionals got real gains from GPT-4 and then lost them to the mental effort of working through a chat window. Less experienced users were hurt most Is the AI capability gap really an interface problem?. Larsen-Ledet names one cost precisely: tools make you keep re-explaining context the system should already know Can generative AI reduce overhead work that adds no value?.

The model side still matters, though not quite in the way the question assumes. Some limits come from how the models work, so neither better design nor more scale fixes them. Models generate text one token at a time, and they can't take back what they've already written. That makes them weak at puzzle-like tasks where you have to try options and abandon them. The fix that works is to pair the model with a separate solver, which is a design choice that adds what the model lacks Why does autoregressive generation fail at constraint satisfaction?. The most surprising finding is that better models fail differently, not less. Weaker models visibly delete parts of a document. Frontier models quietly corrupt them while the document still looks intact Does model capability change how documents degrade?. So as models improve, a person's job of checking the output gets harder. That is exactly the 'role shifts to evaluation' problem Simkute describes. Better models can make the design problem worse.

The corpus also suggests that 'capability' may be the wrong measure. A mid-sized model trained on real multi-step work can compete with much larger models at a fraction of the cost Does model efficiency matter more than peak capability for real work?. Following through on a task matters more than raw intelligence. And the same tool affects people differently depending on how they use it. Interviews with knowledge workers found four separate skill outcomes: some skills grow, some are maintained, some erode, and some change in value How does generative AI actually change worker skills?. A better model can't fix that; how work is organized can.

The honest answer is that it's mostly design, but the proposed fixes are untested. Simkute's team turns their diagnosis into five design principles: continuous feedback, personalization, interfaces that show what's happening, stabilizing tasks, and clear division of work between person and AI. They openly call these research directions rather than proven solutions Can Human Factors design fixes restore GenAI productivity losses?. In some settings, like university assessment, the problem may be too tangled to solve cleanly at all. It has every hallmark of a 'wicked problem' Does GenAI assessment challenge fit wicked problem theory?.