Generative AI is widely believed to help people finish work faster. But what really matters is not the feeling of speed—it is whether, after AI makes things faster, people still want to keep working. Four randomized experiments answered that question with performance and psychological measures.

Generative AI has become a tool for making work faster. It organizes writing, polishes emails, generates ideas, and sometimes produces sentences that sound more convincing than a human’s. That is why a common line emerges in real workplaces: when you use AI, work flows smoothly. The problem comes next. When you return to a stretch of work where AI is not available, the same tasks can suddenly feel heavy and boring—and that experience repeats. What matters is whether this is a temporary mood swing or a structural side effect. If it is temporary, time will fix it. If it is structural, the adoption approach itself has to change.

This paradox is closer to human psychology than to technical performance. People do not work on results alone. To sustain effort, there has to be a sense that you are in control of the work, that you are growing, and that you are making real choices. But the moment a productivity tool takes over the “fun” parts of the process, the output may improve while the tactile satisfaction of doing the work can disappear. Then work becomes faster, yet the will to work becomes weaker—an odd situation created by the same tool that raised performance.

What makes this topic convincing is not observation but experimentation. A research team at Zhejiang University’s School of Management in China conducted four online experiments with a total sample of 3,562 participants. The key was that they designed not a single task but two consecutive tasks. In the first task, one group collaborated with a generative AI system, while the other group worked alone from start to finish. In the second task, everyone worked under the same condition—alone, without AI.

This structure mirrors reality. In real work, some tasks are handled with AI, while other tasks ultimately must be finished by a person working alone. The crucial question is not simply what happens in the AI-enabled segment, but what happens in the human-only segment that follows. The study pulled that transition into the experiment.

The tasks modeled common text-based office work: writing work emails, drafting documents, brainstorming ideas, and solving problems through writing. The researchers evaluated the quality of task outputs while also measuring psychological variables after the tasks—such as perceived control, intrinsic motivation, and boredom. The design aimed to test how a tool that increases speed affects the “speed of the mind.”

The first finding aligns with intuition. In the condition where people collaborated with AI, performance on the first task improved. Within the same time, participants tended to produce more complete writing, organize content more neatly, and use a more persuasive tone. In some cases, AI-assisted documents were observed to be longer, more analytical, and more considerate in style.

But a critical question emerges in the second task: does the experience of working with AI remain as human skill, automatically carrying over and lifting performance on the next task? The study suggests lowering expectations. Overall, the evidence was not strong that performance gains from AI collaboration carried over directly into performance on the next human-only task. In other words, it was not easy to establish a mechanism where AI-produced high-quality output naturally becomes internalized by the person and continues to generate advantages in subsequent tasks.

There were, however, some interesting exceptions. For instance, in a creativity task—such as generating as many new uses as possible for a soda can—AI collaboration showed some positive signals in the later qualitative evaluation of ideas. But these outcomes did not harden into a stable, broad-based upward trend. The conclusion is clear: AI can boost immediate performance, but it is difficult to claim it automatically functions as a growth engine that carries you forward into the next task.

The second result is the core of the study. When people collaborated with AI and then immediately transitioned into a human-only task, psychological costs appeared. Intrinsic motivation decreased and boredom increased, and this pattern was observed repeatedly across multiple experiments. Intrinsic motivation here refers to the drive that comes not from rewards or pressure, but from finding the process itself interesting and meaningful. Boredom is not simply “this is not fun”—it is a signal that the energy needed to keep going is draining away.

This shift may occur because the mind operates by relative comparison. If the AI-assisted segment is extremely smooth, the next segment—done alone—can feel rougher. Even for the same type of work, it may feel as if the difficulty suddenly jumped. And if AI handles the structuring and expression, a person may experience less of the sense of accomplishment and control that normally comes from the process. The output is better, yet the feeling of “I did this” can weaken. Then the work that must be done alone may feel flatter and heavier by comparison.

The study also measured perceived control. Interestingly, perceived control can vary depending on the direction and sequence of the transition. In some conditions, moving from AI collaboration back to working alone showed partial recovery of perceived control. Conversely, moving from working alone into AI collaboration was reported to reduce perceived control. In other words, perceived control is sensitive not only to whether AI is present, but to task order and role allocation. Even so, shifts in intrinsic motivation and boredom remained more stubborn. Speed can rise, but the sustainability of effort requires separate design.

The study’s most uncomfortable message is here. Generative AI can increase productivity while slowly drying out the experience of work. If you look only at immediate output, it appears to be success. But if you look at the full rhythm of repeated work, motivation may erode over time.

This is not a message to stop adopting AI. It is a message to adopt it more precisely. The core is how roles are divided. If AI-enabled segments and non-AI segments are mixed indiscriminately, transition shock can grow. But if segments are designed intentionally, speed and motivation can be protected together.

First, it helps to define and limit what AI is responsible for. Assign AI to segments where speed matters—drafting, organizing, classifying, polishing—while fixing standard-setting and final judgment as human responsibilities. If AI produces a draft, the person should re-check the problem definition, set the key standards, verify the evidence, and take responsibility for the final sentences. The clearer the human role becomes as judgment and accountability, the more likely perceived control is to be preserved.

Second, after using AI, it helps to attach a task where a human “handprint” remains. For example, after AI produces a draft, the person can look for counterexamples, catch errors, rearrange logic, and then move to a stage where they must explain the final output to someone else. Without this segment, AI may take all the interesting parts and leave only closing labor to the person. Then speed rises but motivation falls.

Third, performance management standards must change alongside adoption. If evaluation is based only on the speed AI creates, people may swing more easily between apathy and burnout. The system should push people to use the time AI saves for learning, improvement, and verification. The more metrics focus only on speed, the more people may feel that work “gets done” only in the AI-assisted segment, while the segments that must be done without AI feel increasingly powerless and draining.

Fourth, reducing the number of transitions can also help. If a single day forces frequent switching between AI collaboration and solo work, psychological friction can increase. But if the rhythm is grouped—such as using AI for organizing work in the morning, and doing judgment, persuasion, and verification alone in the afternoon—transition shock may be softened. Transition is not a technical problem; it is an operating problem.

This study is based on online experiments and a relatively short structure of two consecutive tasks. In real organizations, variables such as teamwork, interdependence, reward systems, deadline pressure, and managerial feedback are far more complex. So the result should not be copied as a conclusion for every workplace, but treated as a warning signal that demands careful design.

Even so, the lesson is clear. Generative AI is a technology that increases productivity and, at the same time, reshapes the experience of work. If success is judged only by adoption rates, the motivation gap left behind speed can be missed. The order of work, the boundary of roles, the standards of evaluation, and the rhythm of transitions must all be redesigned together. The next step is to restructure work so that, as things get faster, human motivation does not leave an empty space behind.

Wu, Suqing; Liu, Yukun; Ruan, Mengqi; Chen, Siyu; Xie, Xiao-Yun. (2025). Human-generative AI collaboration enhances task performance but undermines human’s intrinsic motivation. Scientific Reports.