Working with AI makes people better at their craft, but only if they still feel they could outdo it.

This explores what changes to GenAI tools and workflows keep them working alongside people rather than taking over their work, and how much evidence backs those changes. The short version: the corpus has strong ideas and suggestive patterns, but almost no tested interventions. It also suggests that 'competitive' may be the wrong thing to remove entirely.

The most surprising finding comes from a survey of 403 writers. Those who felt both a sense of rivalry with GenAI and a willingness to collaborate with it reported the strongest productivity and craft outcomes Does balancing rivalry and collaboration with GenAI boost writer productivity?. Writers who simply leaned on the tool did not do as well. A little friction, the feeling of 'I can do this better', may be part of what keeps people engaged. The authors proposed adding small deliberate frictions to the tool to encourage that stance. That idea has not been tested, and the survey shows a correlation rather than cause and effect Can micro-frictions boost rivalry without harming collaboration?.

How work is divided between person and tool appears to matter more than which tool is used. A field study of 244 consultants found three patterns. 'Centaurs' kept control of the method and grew in their own field. 'Cyborgs' worked in close back-and-forth with the AI and built new AI skills. 'Self-Automators' handed over both the thinking and the deciding, and gained nothing Do different GenAI collaboration modes produce different skill gains?. Interviews with Dutch knowledge workers show the same split from another angle. The same tool led some people to build skills, some to keep them, some to lose them, and some to value different skills than before, depending on which tasks it took over How does generative AI actually change worker skills?. So one design lever is to make it easy to stay in charge of the method and awkward to hand off the judgment.

Aviation research on automation offers the most concrete design proposals. Simkute and colleagues argue that GenAI repeats old automation problems. People shift from doing the work to checking it, workflows get broken up, interruptions destroy focus, and easy tasks get easier while hard ones stay hard Does generative AI productivity actually improve or decline in practice?. Their proposed fixes come from decades of cockpit design: continuous feedback, interfaces that show the state of the task, clear rules for who does what, and support that adjusts to the person Can Human Factors design fixes restore GenAI productivity losses?. Larsen-Ledet adds a quieter cost. Today's tools make people restate context the system should already have, which is busywork. Explaining your thinking can sharpen the work, while repeating yourself just wears you down Can generative AI reduce overhead work that adds no value?. A complementary tool would remove the second kind of explaining and protect the first.

This matters because the overall gains so far are small. Survey estimates put GenAI's contribution to total productivity at about 1.1% bick-blandin-and-deming-put-generative-ai-adoption-at-546-p. Much of the remaining value probably depends on the human-side design problems above rather than on better models. The open gap in the corpus is that none of these fixes has been tested experimentally. Anyone who runs those tests would be adding something the collection doesn't yet have.