When a company brings generative AI into a brand-new kind of work, it tends to fail in the same old ways automation always has.

This explores the predictable ways generative AI breaks down, or quietly makes work worse, when organizations bring it into new kinds of work. The corpus suggests that many of these failures aren't new. Simkute and colleagues trace GenAI's uneven productivity to the 'ironies of automation' first documented in aviation decades ago. People move from doing work to checking it. Workflows get broken apart. Interruptions destroy focus. Easy tasks get easier while hard ones stay just as hard Does generative AI productivity actually improve or decline in practice?. Pilots and plant operators ran into the same problems. The surprise is that each new field seems to rediscover them from scratch.

A quieter failure mode is the hidden labor AI adds. Larsen-Ledet separates two kinds of explaining. 'Mundane articulation' means restating context the system should already know. 'Productive articulation' means explaining that actually sharpens your thinking. Current tools pile on the first kind, which can wipe out the time they were supposed to save Can generative AI reduce overhead work that adds no value?. A related gap shows up in design: generative UI tools skip about a quarter of the design reasoning they're given, and a third of functional requirements. The output looks finished, but it doesn't do what was asked Do generative UI tools actually implement their stated design rationales?. In both cases the automation seems to work, and the cost only shows up later, for whoever has to catch the gap.

That raises the question of whether errors stay visible at all. One analysis finds that we can partly measure whether AI mistakes are visible, contained, or reversible, but nothing yet measures the whole system of people, tools, and institutions where those mistakes actually play out How can we measure whether AI errors stay visible and recoverable?. Fragility also turns up in unexpected places. AI-text detectors that catch over 99% of rewrites drop to 3.8% the moment a new model ships, because word-choice patterns shift Why do AI detectors suddenly fail when models update?. When agents are chained together, they fail in their own ways: they swap roles, give empty replies, get stuck in loops, and drift off-topic Why do autonomous LLM agents fail in predictable ways?.

The failure you may not have expected is about people over time. In software engineering, generative AI is absorbing the entry-level tasks juniors once learned from and moving them into senior-plus-AI workflows. That removes the hands-on struggle through which expertise used to form Does generative AI prevent juniors from getting entry-level work?. This is the classic irony of automation playing out over a career instead of a single shift: the system depends on human experts to check it while getting rid of the work that produced them. Meanwhile, a lot of real-world adoption stalls before reaching any of these failures. At Argonne National Laboratory, use stayed small and conversational, not built into workflows How are national lab staff actually using generative AI?. A historical look at agent technologies, from GPS onward, argues that missing ecosystem conditions sink deployments more often than missing capability. The five conditions are real value, personalization, trust, social acceptance, and standards Why do capable AI agents still fail in real deployments?.

Taken together, the most serious threats aren't dramatic model failures. They're slow, structural ones: hidden checking work, gaps between what was asked and what was built, errors nobody can see, and the gradual loss of the human skill that oversight depends on. The corpus is stronger on diagnosing these than on proven fixes. Measuring and designing against them is still an open problem.