A machine can complete its assignment while making someone else’s job harder. An AI tool might produce code faster than a developer can meaningfully review it. A robot might follow its route while leaving nearby workers uncertain about where it will move next. In both cases, apparent efficiency can conceal a poorly designed working relationship.
As businesses consider how much work to delegate to AI and robots, I believe they should ask a broader question than how quickly a task gets done. Does the technology help people remain capable, informed, and in control? That should be a condition of successful adoption.
Research offers reason to examine that question. A 2025 study published in the proceedings of the Conference on Human Factors in Computing Systems (CHI), an international conference on human-computer interaction, surveyed 319 knowledge workers and found that greater confidence in generative AI was associated with less reported critical thinking. This does not establish that AI causes lasting skill loss. It does suggest that confidence in a tool deserves scrutiny when people remain responsible for its results
...
My concern is practical. When a system makes accepting an answer easier than understanding it, maintaining meaningful oversight can become harder. Productivity measures should account for that burden.
When AI Makes Work Harder
I encountered this problem after spending three-and-a-half years working manually on teleoperation and mission control systems for robotic quadrupeds; our software organization received an opportunity to try AI.
As we rolled out AI and our team's mandate diverged from the work I was interested in doing, I made the mistake of over-relying on AI tools. My work required more correction, and the apparent shortcut sometimes made delivery slower. I left shortly after to focus on my game studio that I was running on the side.
Around the same time, while running the game studio, we began accepting AI-generated changes faster than we could thoroughly understand them. We later responded by separating the game's rendering and user interface for a handwritten rebuild while retaining AI-supported backend systems, but the damage was done.
We were functionally reliant on AI while controlling less of the codebase than a fully human team would've been able to handle. It was an imperfect compromise, shaped by limited resources, but it reinforced my view that human understanding needs a protected place in the workflow.
I found myself considering the long-term behavioral implications associated with AI and autonomous machine usage.
Designing Robots Around People
Robotics brings this concern into shared physical space. Consider a hypothetical inspection robot approaching a worker moving equipment. Following its assigned route may be technically correct, yet still leave the worker guessing about whether it will stop, turn, or continue.
This becomes a greater design problem when you take into account that workers may be facing away from the robot as it passes by. Let us consider the long-term effects on worker productivity when people are constantly checking behind them to see if there is a robot nearby.
I think of this as behavioral ergonomics: designing machine behavior around human perception, expectations, and working conditions. A robot's interface should serve everyone who encounters it, not only the person holding its controller. People need understandable signals about its intended movement, whether it registered their request, and when it needs assistance.
This is also a measurable engineering concern. The National Institute of Standards and Technology identifies intuitive information sharing, operator experience, and communication effectiveness among its human-robot interaction research priorities. These offer a broader basis for evaluating a deployment than task completion alone.
Animal behavior may offer useful design ideas. A dog-shaped robot could use a visible change in orientation or a distinct sound to indicate movement or proximity. Those cues would need testing with actual workers. Familiar appearance should never imply that a machine understands people as an animal might, or replace established safeguards.
Predictable behavior also deserves more attention in the systems controlling robots. I would like engineers to evaluate smaller, specialized models combined with explicit decision rules where appropriate. The industry term for this would be somewhere between expert systems and probabilistically swapped decision trees.
Such approaches are not automatically safer than language models, but they are considerably more transparent to programmers. Their value would depend on whether teams can test their limits and make their responses understandable in the intended environment.
Keeping Humans in Control
The potential benefit is a workplace where machines reduce burdens while people retain judgment and confidence. The alternative is a gradual transfer of responsibility without corresponding understanding. Workers could become accountable for outputs they struggle to evaluate, or spend their day adapting to machines whose behavior remains unclear.
Before expanding a deployment, employers should run trials with both operators and nearby workers. Test whether people can anticipate movement, recognize misunderstandings, and use appropriate intervention procedures. For AI-assisted work, assess whether employees can explain and correct the output. Track interruptions and rework alongside speed, and preserve opportunities to practice essential skills independently.
These checks would make human capability part of the purchasing decision. A machine's performance should include how well people can work alongside it.
Robert Brownstein is founder of the game studio Gnarled Helix and a former robotics software developer whose work informs his views on AI-assisted coding and human-machine interaction.
All opinions are the author’s own.