AI’s advance is steady, not sudden, study finds

Artificial intelligence is spreading through the workplace in a steady climb rather than a sudden leap, according to new research from MIT Computer Science and Artificial Intelligence Laboratory. Researchers say this is a shift that could make changes more visible to workers and businesses even as the technology improves rapidly.Thestudy, based on thousands of real-world job tasks, finds that AI capabilities are improving across a wide range of text-based work simultaneously, rather than arriving in abrupt breakthroughs that transform entire categories overnight. That pattern, the researchers argue, means progress is more trackable over time, even if the long-term effects on jobs remain significant.“The crashing waves narrative was always more intuitive, but less accurate in enterprise settings,”Ayhan Sebin, Head of Product Incubation at IBM’s Software Innovation Lab, who was not involved in the research, toldIBM Thinkin an interview. “What we see instead is simultaneous, broad-based improvement that quietly raises the floor across everything at once.”

The MIT researchers analyzed more than 17,000 evaluations of AI systems performing tasks from across the US labor market, focusing on work that involves generating or processing text, such as writing, analysis and communication. They found that performance gains tended to occur across many tasks at once, rather than concentrating in narrow areas where systems suddenly improve.“We find little evidence of crashing waves and substantial evidence that rising tides are the primary form of AI automation,” the authors write.The distinction shows up in how AI performance changes as tasks become more complex. In a “crashing wave” scenario, systems would suddenly master certain tasks after long periods of failure. In a “rising tide” scenario, improvements spread more evenly across tasks of different length and difficulty. The researchers found a relatively flat relationship between task duration and success rates, suggesting that gains occurred broadly rather than in concentrated bursts.At the same time, the underlying pace of improvement remains fast. The researchers find that AI can already complete roughly half to three-quarters of these text-based, real-world job tasks at a minimally sufficient level without human edits.“Models can do a minimally sufficient job without human edits on roughly half to three-quarters of potential tasks presented to them,” the authors write.If current trends continue, the study estimates that most text-based tasks could reach success rates of roughly 80 percent to 95 percent by 2029 at that level of quality, though higher levels of performance may take longer.

Inside organizations, that steady improvement often appears less as disruption and more as diffusion, expanding what employees across roles can do.“Developers and non-technical domain experts alike are becoming 10x builders,” Sebin said. “They’re starting to automate processes for both personal and organizational productivity across a remarkably wide range of use cases, at a much faster rate.”The shift reflects a broader change in how work gets done. Tasks that once required specialized programming skills can now be handled by people with domain expertise using AI tools, lowering the barrier to building software and automation.“That democratization of building is itself a rising tide dynamic,” Sebin said. “Capability improvements aren’t landing in one profession or one task type; they’re enabling a fundamentally new class of builder across organizations.”Rather than disappear, Sebin said, the role of human expertise appears to be changing. As AI systems take on more execution, people increasingly focus on directing and evaluating their output.“Experts are becoming agent managers, steering agents, reviewing outputs, confirming quality and compliance,” Sebin said.That shift may concentrate responsibility in a smaller group of highly skilled workers who can manage multiple AI systems effectively. “Only the best of the best AI native experts will have the capacity and judgment to steer effectively, review meaningfully and improve agents over time while managing multiple agents simultaneously,” Sebin said.At the same time, the researchers caution that a gradual trajectory does not mean a limited impact. Improvements are accumulating quickly, and their effects could spread widely as adoption increases.“Our findings do not imply that LLM capabilities are growing slowly, nor that workers will be insulated from AI automation effects,” the authors write.